[
{
"path": "KDSR-GAN/.gitignore",
"content": "# ignored folders\ndatasets/*\nexperiments/*\nresults/*\ntb_logger/*\nwandb/*\ntmp/*\nrealesrgan/weights/*\n\nversion.py\n\n# Byte-compiled / optimized / DLL files\n__pycache__/\n*.py[cod]\n*$py.class\n\n# C extensions\n*.so\n\n# Distribution / packaging\n.Python\nbuild/\ndevelop-eggs/\ndist/\ndownloads/\neggs/\n.eggs/\nlib/\nlib64/\nparts/\nsdist/\nvar/\nwheels/\npip-wheel-metadata/\nshare/python-wheels/\n*.egg-info/\n.installed.cfg\n*.egg\nMANIFEST\n\n# PyInstaller\n# Usually these files are written by a python script from a template\n# before PyInstaller builds the exe, so as to inject date/other infos into it.\n*.manifest\n*.spec\n\n# Installer logs\npip-log.txt\npip-delete-this-directory.txt\n\n# Unit test / coverage reports\nhtmlcov/\n.tox/\n.nox/\n.coverage\n.coverage.*\n.cache\nnosetests.xml\ncoverage.xml\n*.cover\n*.py,cover\n.hypothesis/\n.pytest_cache/\n\n# Translations\n*.mo\n*.pot\n\n# Django stuff:\n*.log\nlocal_settings.py\ndb.sqlite3\ndb.sqlite3-journal\n\n# Flask stuff:\ninstance/\n.webassets-cache\n\n# Scrapy stuff:\n.scrapy\n\n# Sphinx documentation\ndocs/_build/\n\n# PyBuilder\ntarget/\n\n# Jupyter Notebook\n.ipynb_checkpoints\n\n# IPython\nprofile_default/\nipython_config.py\n\n# pyenv\n.python-version\n\n# pipenv\n# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.\n# However, in case of collaboration, if having platform-specific dependencies or dependencies\n# having no cross-platform support, pipenv may install dependencies that don't work, or not\n# install all needed dependencies.\n#Pipfile.lock\n\n# PEP 582; used by e.g. github.com/David-OConnor/pyflow\n__pypackages__/\n\n# Celery stuff\ncelerybeat-schedule\ncelerybeat.pid\n\n# SageMath parsed files\n*.sage.py\n\n# Environments\n.env\n.venv\nenv/\nvenv/\nENV/\nenv.bak/\nvenv.bak/\n\n# Spyder project settings\n.spyderproject\n.spyproject\n\n# Rope project settings\n.ropeproject\n\n# mkdocs documentation\n/site\n\n# mypy\n.mypy_cache/\n.dmypy.json\ndmypy.json\n\n# Pyre type checker\n.pyre/\n"
},
{
"path": "KDSR-GAN/.pre-commit-config.yaml",
"content": "repos:\n # flake8\n - repo: https://github.com/PyCQA/flake8\n rev: 3.8.3\n hooks:\n - id: flake8\n args: [\"--config=setup.cfg\", \"--ignore=W504, W503\"]\n\n # modify known_third_party\n - repo: https://github.com/asottile/seed-isort-config\n rev: v2.2.0\n hooks:\n - id: seed-isort-config\n\n # isort\n - repo: https://github.com/timothycrosley/isort\n rev: 5.2.2\n hooks:\n - id: isort\n\n # yapf\n - repo: https://github.com/pre-commit/mirrors-yapf\n rev: v0.30.0\n hooks:\n - id: yapf\n\n # codespell\n - repo: https://github.com/codespell-project/codespell\n rev: v2.1.0\n hooks:\n - id: codespell\n\n # pre-commit-hooks\n - repo: https://github.com/pre-commit/pre-commit-hooks\n rev: v3.2.0\n hooks:\n - id: trailing-whitespace # Trim trailing whitespace\n - id: check-yaml # Attempt to load all yaml files to verify syntax\n - id: check-merge-conflict # Check for files that contain merge conflict strings\n - id: double-quote-string-fixer # Replace double quoted strings with single quoted strings\n - id: end-of-file-fixer # Make sure files end in a newline and only a newline\n - id: requirements-txt-fixer # Sort entries in requirements.txt and remove incorrect entry for pkg-resources==0.0.0\n - id: fix-encoding-pragma # Remove the coding pragma: # -*- coding: utf-8 -*-\n args: [\"--remove\"]\n - id: mixed-line-ending # Replace or check mixed line ending\n args: [\"--fix=lf\"]\n"
},
{
"path": "KDSR-GAN/CODE_OF_CONDUCT.md",
"content": "# Contributor Covenant Code of Conduct\n\n## Our Pledge\n\nWe as members, contributors, and leaders pledge to make participation in our\ncommunity a harassment-free experience for everyone, regardless of age, body\nsize, visible or invisible disability, ethnicity, sex characteristics, gender\nidentity and expression, level of experience, education, socio-economic status,\nnationality, personal appearance, race, religion, or sexual identity\nand orientation.\n\nWe pledge to act and interact in ways that contribute to an open, welcoming,\ndiverse, inclusive, and healthy community.\n\n## Our Standards\n\nExamples of behavior that contributes to a positive environment for our\ncommunity include:\n\n* Demonstrating empathy and kindness toward other people\n* Being respectful of differing opinions, viewpoints, and experiences\n* Giving and gracefully accepting constructive feedback\n* Accepting responsibility and apologizing to those affected by our mistakes,\n and learning from the experience\n* Focusing on what is best not just for us as individuals, but for the\n overall community\n\nExamples of unacceptable behavior include:\n\n* The use of sexualized language or imagery, and sexual attention or\n advances of any kind\n* Trolling, insulting or derogatory comments, and personal or political attacks\n* Public or private harassment\n* Publishing others' private information, such as a physical or email\n address, without their explicit permission\n* Other conduct which could reasonably be considered inappropriate in a\n professional setting\n\n## Enforcement Responsibilities\n\nCommunity leaders are responsible for clarifying and enforcing our standards of\nacceptable behavior and will take appropriate and fair corrective action in\nresponse to any behavior that they deem inappropriate, threatening, offensive,\nor harmful.\n\nCommunity leaders have the right and responsibility to remove, edit, or reject\ncomments, commits, code, wiki edits, issues, and other contributions that are\nnot aligned to this Code of Conduct, and will communicate reasons for moderation\ndecisions when appropriate.\n\n## Scope\n\nThis Code of Conduct applies within all community spaces, and also applies when\nan individual is officially representing the community in public spaces.\nExamples of representing our community include using an official e-mail address,\nposting via an official social media account, or acting as an appointed\nrepresentative at an online or offline event.\n\n## Enforcement\n\nInstances of abusive, harassing, or otherwise unacceptable behavior may be\nreported to the community leaders responsible for enforcement at\nxintao.wang@outlook.com or xintaowang@tencent.com.\nAll complaints will be reviewed and investigated promptly and fairly.\n\nAll community leaders are obligated to respect the privacy and security of the\nreporter of any incident.\n\n## Enforcement Guidelines\n\nCommunity leaders will follow these Community Impact Guidelines in determining\nthe consequences for any action they deem in violation of this Code of Conduct:\n\n### 1. Correction\n\n**Community Impact**: Use of inappropriate language or other behavior deemed\nunprofessional or unwelcome in the community.\n\n**Consequence**: A private, written warning from community leaders, providing\nclarity around the nature of the violation and an explanation of why the\nbehavior was inappropriate. A public apology may be requested.\n\n### 2. Warning\n\n**Community Impact**: A violation through a single incident or series\nof actions.\n\n**Consequence**: A warning with consequences for continued behavior. No\ninteraction with the people involved, including unsolicited interaction with\nthose enforcing the Code of Conduct, for a specified period of time. This\nincludes avoiding interactions in community spaces as well as external channels\nlike social media. Violating these terms may lead to a temporary or\npermanent ban.\n\n### 3. Temporary Ban\n\n**Community Impact**: A serious violation of community standards, including\nsustained inappropriate behavior.\n\n**Consequence**: A temporary ban from any sort of interaction or public\ncommunication with the community for a specified period of time. No public or\nprivate interaction with the people involved, including unsolicited interaction\nwith those enforcing the Code of Conduct, is allowed during this period.\nViolating these terms may lead to a permanent ban.\n\n### 4. Permanent Ban\n\n**Community Impact**: Demonstrating a pattern of violation of community\nstandards, including sustained inappropriate behavior, harassment of an\nindividual, or aggression toward or disparagement of classes of individuals.\n\n**Consequence**: A permanent ban from any sort of public interaction within\nthe community.\n\n## Attribution\n\nThis Code of Conduct is adapted from the [Contributor Covenant][homepage],\nversion 2.0, available at\nhttps://www.contributor-covenant.org/version/2/0/code_of_conduct.html.\n\nCommunity Impact Guidelines were inspired by [Mozilla's code of conduct\nenforcement ladder](https://github.com/mozilla/diversity).\n\n[homepage]: https://www.contributor-covenant.org\n\nFor answers to common questions about this code of conduct, see the FAQ at\nhttps://www.contributor-covenant.org/faq. Translations are available at\nhttps://www.contributor-covenant.org/translations.\n"
},
{
"path": "KDSR-GAN/Metric/DISTS/DISTS_pytorch/DISTS_pt.py",
"content": "# This is a pytoch implementation of DISTS metric.\n# Requirements: python >= 3.6, pytorch >= 1.0\n\nimport numpy as np\nimport os,sys\nimport torch\nfrom torchvision import models,transforms\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nclass L2pooling(nn.Module):\n def __init__(self, filter_size=5, stride=2, channels=None, pad_off=0):\n super(L2pooling, self).__init__()\n self.padding = (filter_size - 2 )//2\n self.stride = stride\n self.channels = channels\n a = np.hanning(filter_size)[1:-1]\n g = torch.Tensor(a[:,None]*a[None,:])\n g = g/torch.sum(g)\n self.register_buffer('filter', g[None,None,:,:].repeat((self.channels,1,1,1)))\n\n def forward(self, input):\n input = input**2\n out = F.conv2d(input, self.filter, stride=self.stride, padding=self.padding, groups=input.shape[1])\n return (out+1e-12).sqrt()\n\nclass DISTS(torch.nn.Module):\n def __init__(self, load_weights=True):\n super(DISTS, self).__init__()\n vgg_pretrained_features = models.vgg16(pretrained=True).features\n self.stage1 = torch.nn.Sequential()\n self.stage2 = torch.nn.Sequential()\n self.stage3 = torch.nn.Sequential()\n self.stage4 = torch.nn.Sequential()\n self.stage5 = torch.nn.Sequential()\n for x in range(0,4):\n self.stage1.add_module(str(x), vgg_pretrained_features[x])\n self.stage2.add_module(str(4), L2pooling(channels=64))\n for x in range(5, 9):\n self.stage2.add_module(str(x), vgg_pretrained_features[x])\n self.stage3.add_module(str(9), L2pooling(channels=128))\n for x in range(10, 16):\n self.stage3.add_module(str(x), vgg_pretrained_features[x])\n self.stage4.add_module(str(16), L2pooling(channels=256))\n for x in range(17, 23):\n self.stage4.add_module(str(x), vgg_pretrained_features[x])\n self.stage5.add_module(str(23), L2pooling(channels=512))\n for x in range(24, 30):\n self.stage5.add_module(str(x), vgg_pretrained_features[x])\n \n for param in self.parameters():\n param.requires_grad = False\n\n self.register_buffer(\"mean\", torch.tensor([0.485, 0.456, 0.406]).view(1,-1,1,1))\n self.register_buffer(\"std\", torch.tensor([0.229, 0.224, 0.225]).view(1,-1,1,1))\n\n self.chns = [3,64,128,256,512,512]\n self.register_parameter(\"alpha\", nn.Parameter(torch.randn(1, sum(self.chns),1,1)))\n self.register_parameter(\"beta\", nn.Parameter(torch.randn(1, sum(self.chns),1,1)))\n self.alpha.data.normal_(0.1,0.01)\n self.beta.data.normal_(0.1,0.01)\n if load_weights:\n # weights = torch.load(os.path.join(sys.prefix, 'weights.pt'))\n weights = torch.load('scripts/metrics/DISTS/DISTS_pytorch/weights.pt')\n self.alpha.data = weights['alpha']\n self.beta.data = weights['beta']\n \n def forward_once(self, x):\n h = (x-self.mean)/self.std\n h = self.stage1(h)\n h_relu1_2 = h\n h = self.stage2(h)\n h_relu2_2 = h\n h = self.stage3(h)\n h_relu3_3 = h\n h = self.stage4(h)\n h_relu4_3 = h\n h = self.stage5(h)\n h_relu5_3 = h\n return [x,h_relu1_2, h_relu2_2, h_relu3_3, h_relu4_3, h_relu5_3]\n\n def forward(self, x, y, require_grad=False, batch_average=False):\n if require_grad:\n feats0 = self.forward_once(x)\n feats1 = self.forward_once(y) \n else:\n with torch.no_grad():\n feats0 = self.forward_once(x)\n feats1 = self.forward_once(y) \n dist1 = 0 \n dist2 = 0 \n c1 = 1e-6\n c2 = 1e-6\n w_sum = self.alpha.sum() + self.beta.sum()\n alpha = torch.split(self.alpha/w_sum, self.chns, dim=1)\n beta = torch.split(self.beta/w_sum, self.chns, dim=1)\n for k in range(len(self.chns)):\n x_mean = feats0[k].mean([2,3], keepdim=True)\n y_mean = feats1[k].mean([2,3], keepdim=True)\n S1 = (2*x_mean*y_mean+c1)/(x_mean**2+y_mean**2+c1)\n dist1 = dist1+(alpha[k]*S1).sum(1,keepdim=True)\n\n x_var = ((feats0[k]-x_mean)**2).mean([2,3], keepdim=True)\n y_var = ((feats1[k]-y_mean)**2).mean([2,3], keepdim=True)\n xy_cov = (feats0[k]*feats1[k]).mean([2,3],keepdim=True) - x_mean*y_mean\n S2 = (2*xy_cov+c2)/(x_var+y_var+c2)\n dist2 = dist2+(beta[k]*S2).sum(1,keepdim=True)\n\n score = 1 - (dist1+dist2).squeeze()\n if batch_average:\n return score.mean()\n else:\n return score\n\ndef prepare_image(image, resize=True):\n if resize and min(image.size) > 256:\n image = transforms.functional.resize(image, 256)\n image = transforms.ToTensor()(image)\n return image.unsqueeze(0)\n\nif __name__ == '__main__':\n\n from PIL import Image\n import glob\n\n os.environ['CUDA_VISIBLE_DEVICES'] = '0'\n\n # others\n # data_root = '/data1/liangjie/BasicSR_ALL/results/'\n # ref_root = '/data1/liangjie/BasicSR_ALL/datasets/'\n # ref_dirs = ['SISR_Test_matlab/Set5mod12', 'SISR_Test_matlab/Set14mod12', 'SISR_Test_matlab/Manga109mod12', 'SISR_Test_matlab/BSDS100mod12', 'SISR_Test_matlab/General100mod12', 'SISR_Test/Urban100', 'DIV2K/DIV2K_valid_HR/']\n # datasets = ['Set5', 'Set14', 'Manga109', 'BSDS100', 'General100', 'Urban100', 'DIV2K100']\n # img_dirs = ['SRGAN_official', 'ESRGAN_official', 'NatSR_official', 'USRGAN_official', 'SPSR_official', 'SPSR_DF2K', 'ESRGAN_ours_DIV2K', 'ESRGAN_ours_DIV2K_ema', 'ESRGAN_ours_DF2K', 'ESRGAN_ours_DF2K_ema']\n\n # SFTGAN\n # data_root = '/data1/liangjie/BasicSR_ALL/results'\n # ref_root = '/data1/liangjie/BasicSR_ALL/results/SFTGAN_official'\n # ref_dirs = ['GT'] * 7\n # datasets = ['Set5', 'Set14', 'Manga109', 'BSDS100', 'General100', 'Urban100', 'DIV2K100']\n # img_dirs = ['SFTGAN_official']\n\n # new\n data_root = 'results/'\n ref_root = 'datasets/'\n ref_dirs = ['DIV2K/DIV2K_valid_HR/']\n datasets = ['DIV2K100']\n img_dirs = ['ESRGAN_ours_DISTS_300k/visualization/']\n\n logoverall_path = 'results/table_logs/' + 'DISTS_orisize_DISTStrain225k.txt'\n\n for index in range(len(ref_dirs)):\n ref_dir = os.path.join(ref_root, ref_dirs[index])\n for method in img_dirs:\n img_dir = os.path.join(data_root, method, datasets[index])\n\n img_list = sorted(glob.glob(os.path.join(img_dir, '*')))\n\n log_path = 'results/table_logs/' + img_dir.replace('/', '_') + '_DISTS_orisize.txt'\n\n DISTS_all = []\n\n for i, img_path in enumerate(img_list):\n file_name = img_path.split('/')[-1]\n if 'DIV2K100' in img_dir and 'SFTGAN' not in img_dir:\n gt_path = os.path.join(ref_dir, file_name[:4] + '.png')\n elif 'Urban100' in img_dir and 'SFTGAN' not in img_dir:\n gt_path = os.path.join(ref_dir, file_name[:7] + '.png')\n elif 'SFTGAN' in img_dir:\n gt_path = os.path.join(ref_dir, file_name.split('_')[0] + '_gt.png')\n if 'Urban100' in img_dir:\n gt_path = os.path.join(ref_dir, file_name.split('_')[0] + '_' + file_name.split('_')[1] + '_gt.png')\n else:\n if '_' in file_name:\n gt_path = os.path.join(ref_dir, file_name.split('_')[0] + '.png')\n else:\n gt_path = os.path.join(ref_dir, file_name)\n\n ref = prepare_image(Image.open(gt_path).convert(\"RGB\"), resize=False)\n dist = prepare_image(Image.open(img_path).convert(\"RGB\"), resize=False)\n assert ref.shape == dist.shape\n device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n model = DISTS().to(device)\n ref = ref.to(device)\n dist = dist.to(device)\n score = model(ref, dist)\n DISTS_all.append(score.item())\n log = f'{i + 1:3d}: {file_name:25}. \\tDISTS: {score.item():.6f}.'\n with open(log_path, 'a') as f:\n f.write(log + '\\n')\n # print(log)\n\n log = f'Average: DISTS: {sum(DISTS_all) / len(DISTS_all):.6f}'\n with open(log_path, 'a') as f:\n f.write(log + '\\n')\n log_overall = method + '__' + datasets[index] + '__' + log\n with open(logoverall_path, 'a') as f:\n f.write(log_overall + '\\n')\n print(log_overall)\n\n"
},
{
"path": "KDSR-GAN/Metric/DISTS/DISTS_tensorflow/DISTS_tf.py",
"content": "# This is a tensorflow implementation of DISTS metric.\n# Requirements: python >= 3.6, tensorflow-gpu >= 1.15\n\nimport tensorflow.compat.v1 as tf\nimport numpy as np\nimport time\nimport scipy.io as scio\nfrom PIL import Image\nimport argparse\n# tf.enable_eager_execution()\ntf.disable_eager_execution()\n\nclass DISTS():\n def __init__(self):\n self.parameters = scio.loadmat('../weights/net_param.mat')\n self.chns = [3,64,128,256,512,512]\n self.mean = tf.constant(self.parameters['vgg_mean'], dtype=tf.float32, shape=(1,1,1,3),name=\"img_mean\")\n self.std = tf.constant(self.parameters['vgg_std'], dtype=tf.float32, shape=(1,1,1,3),name=\"img_std\")\n # self.alpha = tf.Variable(tf.random_normal(shape=(1,1,1,sum(self.chns)), mean=0.1, stddev=0.01),name=\"alpha\")\n # self.beta = tf.Variable(tf.random_normal(shape=(1,1,1,sum(self.chns)), mean=0.1, stddev=0.01),name=\"beta\")\n self.weights = scio.loadmat('../weights/alpha_beta.mat')\n self.alpha = tf.constant(np.reshape(self.weights['alpha'],(1,1,1,sum(self.chns))),name=\"alpha\")\n self.beta = tf.constant(np.reshape(self.weights['beta'],(1,1,1,sum(self.chns))),name=\"beta\")\n\n def get_features(self, img):\n\n x = (img - self.mean)/self.std\n\n self.conv1_1 = self.conv_layer(x, \"conv1_1\")\n self.conv1_2 = self.conv_layer(self.conv1_1, \"conv1_2\")\n self.pool1 = self.pool_layer(self.conv1_2, name=\"pool_1\")\n\n self.conv2_1 = self.conv_layer(self.pool1, \"conv2_1\")\n self.conv2_2 = self.conv_layer(self.conv2_1, \"conv2_2\")\n self.pool2 = self.pool_layer(self.conv2_2, name=\"pool_2\")\n\n self.conv3_1 = self.conv_layer(self.pool2, \"conv3_1\")\n self.conv3_2 = self.conv_layer(self.conv3_1, \"conv3_2\")\n self.conv3_3 = self.conv_layer(self.conv3_2, \"conv3_3\")\n self.pool3 = self.pool_layer(self.conv3_3, name=\"pool_3\")\n\n self.conv4_1 = self.conv_layer(self.pool3, \"conv4_1\")\n self.conv4_2 = self.conv_layer(self.conv4_1, \"conv4_2\")\n self.conv4_3 = self.conv_layer(self.conv4_2, \"conv4_3\")\n self.pool4 = self.pool_layer(self.conv4_3, name=\"pool_4\")\n\n self.conv5_1 = self.conv_layer(self.pool4, \"conv5_1\")\n self.conv5_2 = self.conv_layer(self.conv5_1, \"conv5_2\")\n self.conv5_3 = self.conv_layer(self.conv5_2, \"conv5_3\")\n\n return [img, self.conv1_2,self.conv2_2,self.conv3_3,self.conv4_3,self.conv5_3]\n\n def conv_layer(self, input, name):\n with tf.variable_scope(name) as _:\n filter = self.get_conv_filter(name)\n conv = tf.nn.conv2d(input, filter, strides=1, padding=\"SAME\")\n bias = self.get_bias(name)\n conv = tf.nn.relu(tf.nn.bias_add(conv, bias))\n return conv\n\n def pool_layer(self, input, name):\n # return tf.nn.max_pool(input, ksize=[1,2,2,1], strides=[1,2,2,1], padding=\"SAME\")\n with tf.variable_scope(name) as _:\n filter = tf.squeeze(tf.constant(self.parameters['L2'+name], name = \"filter\"),3)\n conv = tf.nn.conv2d(input**2, filter, strides=2, padding=[[0, 0], [1, 0], [1, 0], [0, 0]])\n return tf.sqrt(tf.maximum(conv, 1e-12)) \n\n def get_conv_filter(self, name):\n return tf.constant(self.parameters[name+'_weight'], name = \"filter\")\n\n def get_bias(self, name):\n return tf.constant(np.squeeze(self.parameters[name+'_bias']), name = \"bias\")\n\n def get_score(self, img1, img2):\n feats0 = self.get_features(img1)\n feats1 = self.get_features(img2)\n dist1 = 0 \n dist2 = 0 \n c1 = 1e-6\n c2 = 1e-6\n w_sum = tf.reduce_sum(self.alpha) + tf.reduce_sum(self.beta)\n alpha = tf.split(self.alpha/w_sum, self.chns, axis=3)\n beta = tf.split(self.beta/w_sum, self.chns, axis=3)\n for k in range(len(self.chns)):\n x_mean = tf.reduce_mean(feats0[k],[1,2], keepdims=True)\n y_mean = tf.reduce_mean(feats1[k],[1,2], keepdims=True)\n S1 = (2*x_mean*y_mean+c1)/(x_mean**2+y_mean**2+c1)\n dist1 = dist1+tf.reduce_sum(alpha[k]*S1, 3, keepdims=True)\n x_var = tf.reduce_mean((feats0[k]-x_mean)**2,[1,2], keepdims=True)\n y_var = tf.reduce_mean((feats1[k]-y_mean)**2,[1,2], keepdims=True)\n xy_cov = tf.reduce_mean(feats0[k]*feats1[k],[1,2], keepdims=True) - x_mean*y_mean\n S2 = (2*xy_cov+c2)/(x_var+y_var+c2)\n dist2 = dist2+tf.reduce_sum(beta[k]*S2, 3, keepdims=True)\n\n dist = 1-tf.squeeze(dist1+dist2)\n return dist\n\n\nif __name__ == '__main__':\n parser = argparse.ArgumentParser()\n parser.add_argument('--ref', type=str, default='../images/r0.png')\n parser.add_argument('--dist', type=str, default='../images/r1.png')\n args = parser.parse_args()\n model = DISTS()\n\n ref = np.array(Image.open(args.ref).convert(\"RGB\"))\n ref = np.expand_dims(ref,axis=0)/255.\n dist = np.array(Image.open(args.dist).convert(\"RGB\"))\n dist = np.expand_dims(dist,axis=0)/255.\n\n x = tf.placeholder(dtype=tf.float32, shape=ref.shape, name= \"ref\")\n y = tf.placeholder(dtype=tf.float32, shape=dist.shape, name= \"dist\")\n score = model.get_score(x,y)\n with tf.Session() as sess:\n sess.run(tf.global_variables_initializer())\n score = sess.run(score, feed_dict={x: ref, y: dist})\n print(score)\n\n\n"
},
{
"path": "KDSR-GAN/Metric/DISTS/LICENSE",
"content": "MIT License\n\nCopyright (c) 2020 Keyan Ding\n\nPermission is hereby granted, free of charge, to any person obtaining a copy\nof this software and associated documentation files (the \"Software\"), to deal\nin the Software without restriction, including without limitation the rights\nto use, copy, modify, merge, publish, distribute, sublicense, and/or sell\ncopies of the Software, and to permit persons to whom the Software is\nfurnished to do so, subject to the following conditions:\n\nThe above copyright notice and this permission notice shall be included in all\ncopies or substantial portions of the Software.\n\nTHE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR\nIMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,\nFITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE\nAUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER\nLIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,\nOUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE\nSOFTWARE.\n"
},
{
"path": "KDSR-GAN/Metric/DISTS/requirements.txt",
"content": "torch>=1.0"
},
{
"path": "KDSR-GAN/Metric/LPIPS.py",
"content": "import cv2\nimport glob\nimport numpy as np\nimport os.path as osp\nfrom torchvision.transforms.functional import normalize\nfrom basicsr.utils import img2tensor\nimport lpips\nimport argparse\n\n\ndef main():\n # Configurations\n parser = argparse.ArgumentParser()\n parser.add_argument('--folder_gt', type=str, default='/root/datasets/NTIRE2020-Track1/track1-valid-gt')\n parser.add_argument('--folder_restored', type=str, default='/root/results/NTIRE2020-Track1')\n args = parser.parse_args()\n loss_fn_vgg = lpips.LPIPS(net='vgg').cuda(0)\n lpips_all = []\n img_list = sorted(glob.glob(osp.join(args.folder_gt, '*.png')))\n lr_list = sorted(glob.glob(osp.join(args.folder_restored, '*.png')))\n mean = [0.5, 0.5, 0.5]\n std = [0.5, 0.5, 0.5]\n for i, (img_path, lr_path) in enumerate(zip(img_list,lr_list)):\n basename, ext = osp.splitext(osp.basename(img_path))\n img_gt = cv2.imread(img_path, cv2.IMREAD_UNCHANGED).astype(np.float32) / 255.\n img_restored = cv2.imread(osp.join(lr_path), cv2.IMREAD_UNCHANGED).astype(\n np.float32) / 255.\n\n img_gt, img_restored = img2tensor([img_gt, img_restored], bgr2rgb=True, float32=True)\n # norm to [-1, 1]\n normalize(img_gt, mean, std, inplace=True)\n normalize(img_restored, mean, std, inplace=True)\n\n # calculate lpips\n lpips_val = loss_fn_vgg(img_restored.unsqueeze(0).cuda(0), img_gt.unsqueeze(0).cuda(0)).cpu().data.numpy()[0,0,0,0]\n # print(lpips_val)\n lpips_all.append(lpips_val)\n\n print(f'Average: LPIPS: {sum(lpips_all) / len(lpips_all):.6f}')\n\n\nif __name__ == '__main__':\n main()\n"
},
{
"path": "KDSR-GAN/Metric/PSNR.py",
"content": "import cv2\nimport glob\nimport numpy as np\nimport os.path as osp\nfrom torchvision.transforms.functional import normalize\nfrom basicsr.utils import img2tensor\nimport lpips\nimport argparse\nfrom basicsr.metrics import calculate_psnr, calculate_ssim\n\n\ndef main():\n # Configurations\n parser = argparse.ArgumentParser()\n parser.add_argument('--folder_gt', type=str, default='/root/datasets/NTIRE2020-Track1/track1-valid-gt')\n parser.add_argument('--folder_restored', type=str, default='/root/results/NTIRE2020-Track1')\n args = parser.parse_args()\n psnr_all = []\n ssim_all = []\n img_list = sorted(glob.glob(osp.join(args.folder_gt, '*.png')))\n lr_list = sorted(glob.glob(osp.join(args.folder_restored, '*.png')))\n for i, (img_path, lr_path) in enumerate(zip(img_list,lr_list)):\n basename, ext = osp.splitext(osp.basename(img_path))\n img_gt = cv2.imread(img_path, cv2.IMREAD_UNCHANGED)\n img_restored = cv2.imread(osp.join(lr_path), cv2.IMREAD_UNCHANGED)\n psnr=calculate_psnr(img_restored, img_gt, crop_border=4, test_y_channel=True)\n ssim=calculate_ssim(img_restored, img_gt, crop_border=4, test_y_channel=True)\n psnr_all.append(psnr)\n ssim_all.append(ssim)\n\n print(f'Average: PSNR: {sum(psnr_all) / len(psnr_all):.6f}')\n print(f'Average: SSIM: {sum(ssim_all) / len(ssim_all):.6f}')\n\n\nif __name__ == '__main__':\n main()\n"
},
{
"path": "KDSR-GAN/Metric/dists.py",
"content": "# This is a pytoch implementation of DISTS metric.\n# Requirements: python >= 3.6, pytorch >= 1.0\n\nimport numpy as np\nimport os,sys\nimport torch\nfrom torchvision import models,transforms\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport argparse\nimport os.path as osp\n\nclass L2pooling(nn.Module):\n def __init__(self, filter_size=5, stride=2, channels=None, pad_off=0):\n super(L2pooling, self).__init__()\n self.padding = (filter_size - 2 )//2\n self.stride = stride\n self.channels = channels\n a = np.hanning(filter_size)[1:-1]\n g = torch.Tensor(a[:,None]*a[None,:])\n g = g/torch.sum(g)\n self.register_buffer('filter', g[None,None,:,:].repeat((self.channels,1,1,1)))\n\n def forward(self, input):\n input = input**2\n out = F.conv2d(input, self.filter, stride=self.stride, padding=self.padding, groups=input.shape[1])\n return (out+1e-12).sqrt()\n\nclass DISTS(torch.nn.Module):\n def __init__(self, load_weights=True):\n super(DISTS, self).__init__()\n vgg_pretrained_features = models.vgg16(pretrained=True).features\n self.stage1 = torch.nn.Sequential()\n self.stage2 = torch.nn.Sequential()\n self.stage3 = torch.nn.Sequential()\n self.stage4 = torch.nn.Sequential()\n self.stage5 = torch.nn.Sequential()\n for x in range(0,4):\n self.stage1.add_module(str(x), vgg_pretrained_features[x])\n self.stage2.add_module(str(4), L2pooling(channels=64))\n for x in range(5, 9):\n self.stage2.add_module(str(x), vgg_pretrained_features[x])\n self.stage3.add_module(str(9), L2pooling(channels=128))\n for x in range(10, 16):\n self.stage3.add_module(str(x), vgg_pretrained_features[x])\n self.stage4.add_module(str(16), L2pooling(channels=256))\n for x in range(17, 23):\n self.stage4.add_module(str(x), vgg_pretrained_features[x])\n self.stage5.add_module(str(23), L2pooling(channels=512))\n for x in range(24, 30):\n self.stage5.add_module(str(x), vgg_pretrained_features[x])\n \n for param in self.parameters():\n param.requires_grad = False\n\n self.register_buffer(\"mean\", torch.tensor([0.485, 0.456, 0.406]).view(1,-1,1,1))\n self.register_buffer(\"std\", torch.tensor([0.229, 0.224, 0.225]).view(1,-1,1,1))\n\n self.chns = [3,64,128,256,512,512]\n self.register_parameter(\"alpha\", nn.Parameter(torch.randn(1, sum(self.chns),1,1)))\n self.register_parameter(\"beta\", nn.Parameter(torch.randn(1, sum(self.chns),1,1)))\n self.alpha.data.normal_(0.1,0.01)\n self.beta.data.normal_(0.1,0.01)\n if load_weights:\n # weights = torch.load(os.path.join(sys.prefix, 'weights.pt'))\n weights = torch.load('DISTS/DISTS_pytorch/weights.pt')\n self.alpha.data = weights['alpha']\n self.beta.data = weights['beta']\n \n def forward_once(self, x):\n h = (x-self.mean)/self.std\n h = self.stage1(h)\n h_relu1_2 = h\n h = self.stage2(h)\n h_relu2_2 = h\n h = self.stage3(h)\n h_relu3_3 = h\n h = self.stage4(h)\n h_relu4_3 = h\n h = self.stage5(h)\n h_relu5_3 = h\n return [x,h_relu1_2, h_relu2_2, h_relu3_3, h_relu4_3, h_relu5_3]\n\n def forward(self, x, y, require_grad=False, batch_average=False):\n if require_grad:\n feats0 = self.forward_once(x)\n feats1 = self.forward_once(y) \n else:\n with torch.no_grad():\n feats0 = self.forward_once(x)\n feats1 = self.forward_once(y) \n dist1 = 0 \n dist2 = 0 \n c1 = 1e-6\n c2 = 1e-6\n w_sum = self.alpha.sum() + self.beta.sum()\n alpha = torch.split(self.alpha/w_sum, self.chns, dim=1)\n beta = torch.split(self.beta/w_sum, self.chns, dim=1)\n for k in range(len(self.chns)):\n x_mean = feats0[k].mean([2,3], keepdim=True)\n y_mean = feats1[k].mean([2,3], keepdim=True)\n S1 = (2*x_mean*y_mean+c1)/(x_mean**2+y_mean**2+c1)\n dist1 = dist1+(alpha[k]*S1).sum(1,keepdim=True)\n\n x_var = ((feats0[k]-x_mean)**2).mean([2,3], keepdim=True)\n y_var = ((feats1[k]-y_mean)**2).mean([2,3], keepdim=True)\n xy_cov = (feats0[k]*feats1[k]).mean([2,3],keepdim=True) - x_mean*y_mean\n S2 = (2*xy_cov+c2)/(x_var+y_var+c2)\n dist2 = dist2+(beta[k]*S2).sum(1,keepdim=True)\n\n score = 1 - (dist1+dist2).squeeze()\n if batch_average:\n return score.mean()\n else:\n return score\n\ndef prepare_image(image, resize=True):\n if resize and min(image.size) > 256:\n image = transforms.functional.resize(image, 256)\n image = transforms.ToTensor()(image)\n return image.unsqueeze(0)\n\nif __name__ == '__main__':\n\n from PIL import Image\n import glob\n\n os.environ['CUDA_VISIBLE_DEVICES'] = '0'\n parser = argparse.ArgumentParser()\n parser.add_argument('--folder_gt', type=str, default='/root/datasets/NTIRE2020-Track1/track1-valid-gt')\n parser.add_argument('--folder_restored', type=str, default='/root/results/NTIRE2020-Track1')\n args = parser.parse_args()\n img_list = sorted(glob.glob(osp.join(args.folder_gt, '*.png')))\n lr_list = sorted(glob.glob(osp.join(args.folder_restored, '*.png')))\n\n\n DISTS_all = []\n\n for i, (gt_path, lr_path) in enumerate(zip(img_list,lr_list)):\n ref = prepare_image(Image.open(gt_path).convert(\"RGB\"), resize=False)\n dist = prepare_image(Image.open(lr_path).convert(\"RGB\"), resize=False)\n assert ref.shape == dist.shape\n device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n model = DISTS().to(device)\n ref = ref.to(device)\n dist = dist.to(device)\n score = model(ref, dist)\n DISTS_all.append(score.item())\n log = f'{i + 1:3d}:\\tDISTS: {score.item():.6f}.'\n print(log)\n # with open(log_path, 'a') as f:\n # f.write(log + '\\n')\n # # print(log)\n\n print(f'Average: DISTS: {sum(DISTS_all) / len(DISTS_all):.6f}')\n\n\n"
},
{
"path": "KDSR-GAN/Metric/pip.sh",
"content": "pip install lpips --index-url http://pypi.douban.com/simple --trusted-host pypi.douban.com\npip install basicsr --index-url http://pypi.douban.com/simple --trusted-host pypi.douban.com\n"
},
{
"path": "KDSR-GAN/README.md",
"content": "# KDSR-GAN\n\nThis project is the official implementation of 'Knowledge Distillation based Degradation Estimation for Blind Super-Resolution', ICLR2023\n> **Knowledge Distillation based Degradation Estimation for Blind Super-Resolution [[Paper](https://arxiv.org/pdf/2211.16928.pdf)] [[Project](https://github.com/Zj-BinXia/KDSR)]**\n\nThis is code for KDSR-GAN (for Real-world SR)\n\n\n ![](\"images/method.jpg\")
\n
\n\n---\n\n## Dependencies and Installation\n\n- Python >= 3.8 (Recommend to use [Anaconda](https://www.anaconda.com/download/#linux) or [Miniconda](https://docs.conda.io/en/latest/miniconda.html))\n- [PyTorch >= 1.10](https://pytorch.org/)\n\n### Installation\n\nInstall dependent packages\n\n\n pip install basicsr \n pip install -r requirements.txt\n pip install pandas \n sudo python3 setup.py develop\n\n\n---\n\n## Dataset Preparation\n\nWe use the same training datasets as [Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN) (DF2K+OST).\n\n---\n\n## Training (8 V100 GPUs)\n\n1. We train KDSRNet_T (only using L1 loss)\n\n```bash\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m torch.distributed.launch --nproc_per_node=8 --master_port=3309 kdsrgan/train.py -opt options/train_kdsrnet_x4TA.yml --launcher pytorch \n```\n\n2. we train KDSRNet_S (only using L1 loss and KD loss). **It is notable that modify the ''pretrain_network_TA'' and ''pretrain_network_g'' of options/train_kdsrnet_x4ST.yml to the path of trained KDSRNet_T checkpoint.** Then, we run\n\n```bash\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \\\npython3 -m torch.distributed.launch --nproc_per_node=8 --master_port=4349 kdsrgan/train.py -opt options/train_kdsrnet_x4ST.yml --launcher pytorch \n```\n\n3. we train KDSRGAN_S ( using L1 loss, perceptual loss, adversial loss and KD loss). **It is notable that modify the ''pretrain_network_TA'' and ''pretrain_network_g'' of options/train_kdsrnet_x4ST.yml or options/train_kdsrnet_x4STV2.yml to the path of trained KDSRNet_T and KDSRNet_S checkpoint, respectively.** Then, we run\n\n```bash\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \\\npython3 -m torch.distributed.launch --nproc_per_node=8 --master_port=4397 kdsrgan/train.py -opt options/train_kdsrgan_x4ST.yml --launcher pytorch\n\n```\nor\n```bash\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \\\npython3 -m torch.distributed.launch --nproc_per_node=8 --master_port=4397 kdsrgan/train.py -opt options/train_kdsrgan_x4STV2.yml --launcher pytorch\n\n```\n---\n\n## :european_castle: Model Zoo\n\nPlease download checkpoints from [Google Drive](https://drive.google.com/drive/folders/1QlOz4F9Mtp9DFXoaHYbnMnRSonR9YFJA).\n\n---\n## Testing\nYou can download NTIRE2020 Track1 and NTIRE2020 Track2 and AIM2019 Track2 from [official link](https://competitions.codalab.org/competitions/22220) or [Google Drive](https://drive.google.com/drive/folders/1P5x8wuty3IVj7aKoDe6uZy9pA19S9rrQ?usp=sharing).\n\n```bash\npython3 kdsrgan/test.py -opt options/test_kdsrgan_x4ST.yml \n```\n\ncalculate metric\n```bash\npython3 Metric/PSNR.py --folder_gt PathtoGT --folder_restored PathtoSR\n```\n\n```bash\npython3 Metric/LPIPS.py --folder_gt PathtoGT --folder_restored PathtoSR\n```\n\n---\n## Results\n\n ![](\"images/quan.jpg\")
\n
\n\n\n---\n\n## BibTeX\n\n @InProceedings{xia2022knowledge,\n title={Knowledge Distillation based Degradation Estimation for Blind Super-Resolution},\n author={Xia, Bin and Zhang, Yulun and Wang, Yitong and Tian, Yapeng and Yang, Wenming and Timofte, Radu and Van Gool, Luc},\n journal={ICLR},\n year={2023}\n }\n\n## 📧 Contact\n\nIf you have any question, please email `zjbinxia@gmail.com`.\n\n\n\n"
},
{
"path": "KDSR-GAN/VERSION",
"content": "0.2.5.0\n"
},
{
"path": "KDSR-GAN/docs/CONTRIBUTING.md",
"content": "# Contributing to Real-ESRGAN\n\n:art: Real-ESRGAN needs your contributions. Any contributions are welcome, such as new features/models/typo fixes/suggestions/maintenance, *etc*. See [CONTRIBUTING.md](docs/CONTRIBUTING.md). All contributors are list [here](README.md#hugs-acknowledgement).\n\nWe like open-source and want to develop practical algorithms for general image restoration. However, individual strength is limited. So, any kinds of contributions are welcome, such as:\n\n- New features\n- New models (your fine-tuned models)\n- Bug fixes\n- Typo fixes\n- Suggestions\n- Maintenance\n- Documents\n- *etc*\n\n## Workflow\n\n1. Fork and pull the latest Real-ESRGAN repository\n1. Checkout a new branch (do not use master branch for PRs)\n1. Commit your changes\n1. Create a PR\n\n**Note**:\n\n1. Please check the code style and linting\n 1. The style configuration is specified in [setup.cfg](setup.cfg)\n 1. If you use VSCode, the settings are configured in [.vscode/settings.json](.vscode/settings.json)\n1. Strongly recommend using `pre-commit hook`. It will check your code style and linting before your commit.\n 1. In the root path of project folder, run `pre-commit install`\n 1. The pre-commit configuration is listed in [.pre-commit-config.yaml](.pre-commit-config.yaml)\n1. Better to [open a discussion](https://github.com/xinntao/Real-ESRGAN/discussions) before large changes.\n 1. Welcome to discuss :sunglasses:. I will try my best to join the discussion.\n\n## TODO List\n\n:zero: The most straightforward way of improving model performance is to fine-tune on some specific datasets.\n\nHere are some TODOs:\n\n- [ ] optimize for human faces\n- [ ] optimize for texts\n- [ ] support controllable restoration strength\n\n:one: There are also [several issues](https://github.com/xinntao/Real-ESRGAN/issues) that require helpers to improve. If you can help, please let me know :smile:\n"
},
{
"path": "KDSR-GAN/docs/FAQ.md",
"content": "# FAQ\n\n1. **Q: How to select models?**
\nA: Please refer to [docs/model_zoo.md](docs/model_zoo.md)\n\n1. **Q: Can `face_enhance` be used for anime images/animation videos?**
\nA: No, it can only be used for real faces. It is recommended not to use this option for anime images/animation videos to save GPU memory.\n\n1. **Q: Error \"slow_conv2d_cpu\" not implemented for 'Half'**
\nA: In order to save GPU memory consumption and speed up inference, Real-ESRGAN uses half precision (fp16) during inference by default. However, some operators for half inference are not implemented in CPU mode. You need to add **`--fp32` option** for the commands. For example, `python inference_realesrgan.py -n RealESRGAN_x4plus.pth -i inputs --fp32`.\n"
},
{
"path": "KDSR-GAN/docs/Training.md",
"content": "# :computer: How to Train/Finetune Real-ESRGAN\n\n- [Train Real-ESRGAN](#train-real-esrgan)\n - [Overview](#overview)\n - [Dataset Preparation](#dataset-preparation)\n - [Train Real-ESRNet](#Train-Real-ESRNet)\n - [Train Real-ESRGAN](#Train-Real-ESRGAN)\n- [Finetune Real-ESRGAN on your own dataset](#Finetune-Real-ESRGAN-on-your-own-dataset)\n - [Generate degraded images on the fly](#Generate-degraded-images-on-the-fly)\n - [Use paired training data](#use-your-own-paired-data)\n\n[English](Training.md) **|** [简体中文](Training_CN.md)\n\n## Train Real-ESRGAN\n\n### Overview\n\nThe training has been divided into two stages. These two stages have the same data synthesis process and training pipeline, except for the loss functions. Specifically,\n\n1. We first train Real-ESRNet with L1 loss from the pre-trained model ESRGAN.\n1. We then use the trained Real-ESRNet model as an initialization of the generator, and train the Real-ESRGAN with a combination of L1 loss, perceptual loss and GAN loss.\n\n### Dataset Preparation\n\nWe use DF2K (DIV2K and Flickr2K) + OST datasets for our training. Only HR images are required.
\nYou can download from :\n\n1. DIV2K: http://data.vision.ee.ethz.ch/cvl/DIV2K/DIV2K_train_HR.zip\n2. Flickr2K: https://cv.snu.ac.kr/research/EDSR/Flickr2K.tar\n3. OST: https://openmmlab.oss-cn-hangzhou.aliyuncs.com/datasets/OST_dataset.zip\n\nHere are steps for data preparation.\n\n#### Step 1: [Optional] Generate multi-scale images\n\nFor the DF2K dataset, we use a multi-scale strategy, *i.e.*, we downsample HR images to obtain several Ground-Truth images with different scales.
\nYou can use the [scripts/generate_multiscale_DF2K.py](scripts/generate_multiscale_DF2K.py) script to generate multi-scale images.
\nNote that this step can be omitted if you just want to have a fast try.\n\n```bash\npython scripts/generate_multiscale_DF2K.py --input datasets/DF2K/DF2K_HR --output datasets/DF2K/DF2K_multiscale\n```\n\n#### Step 2: [Optional] Crop to sub-images\n\nWe then crop DF2K images into sub-images for faster IO and processing.
\nThis step is optional if your IO is enough or your disk space is limited.\n\nYou can use the [scripts/extract_subimages.py](scripts/extract_subimages.py) script. Here is the example:\n\n```bash\n python scripts/extract_subimages.py --input datasets/DF2K/DF2K_multiscale --output datasets/DF2K/DF2K_multiscale_sub --crop_size 400 --step 200\n```\n\n#### Step 3: Prepare a txt for meta information\n\nYou need to prepare a txt file containing the image paths. The following are some examples in `meta_info_DF2Kmultiscale+OST_sub.txt` (As different users may have different sub-images partitions, this file is not suitable for your purpose and you need to prepare your own txt file):\n\n```txt\nDF2K_HR_sub/000001_s001.png\nDF2K_HR_sub/000001_s002.png\nDF2K_HR_sub/000001_s003.png\n...\n```\n\nYou can use the [scripts/generate_meta_info.py](scripts/generate_meta_info.py) script to generate the txt file.
\nYou can merge several folders into one meta_info txt. Here is the example:\n\n```bash\n python scripts/generate_meta_info.py --input datasets/DF2K/DF2K_HR datasets/DF2K/DF2K_multiscale --root datasets/DF2K datasets/DF2K --meta_info datasets/DF2K/meta_info/meta_info_DF2Kmultiscale.txt\n```\n\n### Train Real-ESRNet\n\n1. Download pre-trained model [ESRGAN](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.1/ESRGAN_SRx4_DF2KOST_official-ff704c30.pth) into `experiments/pretrained_models`.\n ```bash\n wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.1/ESRGAN_SRx4_DF2KOST_official-ff704c30.pth -P experiments/pretrained_models\n ```\n1. Modify the content in the option file `options/train_realesrnet_x4plus.yml` accordingly:\n ```yml\n train:\n name: DF2K+OST\n type: RealESRGANDataset\n dataroot_gt: datasets/DF2K # modify to the root path of your folder\n meta_info: realesrgan/meta_info/meta_info_DF2Kmultiscale+OST_sub.txt # modify to your own generate meta info txt\n io_backend:\n type: disk\n ```\n1. If you want to perform validation during training, uncomment those lines and modify accordingly:\n ```yml\n # Uncomment these for validation\n # val:\n # name: validation\n # type: PairedImageDataset\n # dataroot_gt: path_to_gt\n # dataroot_lq: path_to_lq\n # io_backend:\n # type: disk\n\n ...\n\n # Uncomment these for validation\n # validation settings\n # val:\n # val_freq: !!float 5e3\n # save_img: True\n\n # metrics:\n # psnr: # metric name, can be arbitrary\n # type: calculate_psnr\n # crop_border: 4\n # test_y_channel: false\n ```\n1. Before the formal training, you may run in the `--debug` mode to see whether everything is OK. We use four GPUs for training:\n ```bash\n CUDA_VISIBLE_DEVICES=0,1,2,3 \\\n python -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 realesrgan/train.py -opt options/train_realesrnet_x4plus.yml --launcher pytorch --debug\n ```\n\n Train with **a single GPU** in the *debug* mode:\n ```bash\n python realesrgan/train.py -opt options/train_realesrnet_x4plus.yml --debug\n ```\n1. The formal training. We use four GPUs for training. We use the `--auto_resume` argument to automatically resume the training if necessary.\n ```bash\n CUDA_VISIBLE_DEVICES=0,1,2,3 \\\n python -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 realesrgan/train.py -opt options/train_realesrnet_x4plus.yml --launcher pytorch --auto_resume\n ```\n\n Train with **a single GPU**:\n ```bash\n python realesrgan/train.py -opt options/train_realesrnet_x4plus.yml --auto_resume\n ```\n\n### Train Real-ESRGAN\n\n1. After the training of Real-ESRNet, you now have the file `experiments/train_RealESRNetx4plus_1000k_B12G4_fromESRGAN/model/net_g_1000000.pth`. If you need to specify the pre-trained path to other files, modify the `pretrain_network_g` value in the option file `train_realesrgan_x4plus.yml`.\n1. Modify the option file `train_realesrgan_x4plus.yml` accordingly. Most modifications are similar to those listed above.\n1. Before the formal training, you may run in the `--debug` mode to see whether everything is OK. We use four GPUs for training:\n ```bash\n CUDA_VISIBLE_DEVICES=0,1,2,3 \\\n python -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 realesrgan/train.py -opt options/train_realesrgan_x4plus.yml --launcher pytorch --debug\n ```\n\n Train with **a single GPU** in the *debug* mode:\n ```bash\n python realesrgan/train.py -opt options/train_realesrgan_x4plus.yml --debug\n ```\n1. The formal training. We use four GPUs for training. We use the `--auto_resume` argument to automatically resume the training if necessary.\n ```bash\n CUDA_VISIBLE_DEVICES=0,1,2,3 \\\n python -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 realesrgan/train.py -opt options/train_realesrgan_x4plus.yml --launcher pytorch --auto_resume\n ```\n\n Train with **a single GPU**:\n ```bash\n python realesrgan/train.py -opt options/train_realesrgan_x4plus.yml --auto_resume\n ```\n\n## Finetune Real-ESRGAN on your own dataset\n\nYou can finetune Real-ESRGAN on your own dataset. Typically, the fine-tuning process can be divided into two cases:\n\n1. [Generate degraded images on the fly](#Generate-degraded-images-on-the-fly)\n1. [Use your own **paired** data](#Use-paired-training-data)\n\n### Generate degraded images on the fly\n\nOnly high-resolution images are required. The low-quality images are generated with the degradation process described in Real-ESRGAN during trainig.\n\n**1. Prepare dataset**\n\nSee [this section](#dataset-preparation) for more details.\n\n**2. Download pre-trained models**\n\nDownload pre-trained models into `experiments/pretrained_models`.\n\n- *RealESRGAN_x4plus.pth*:\n ```bash\n wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth -P experiments/pretrained_models\n ```\n\n- *RealESRGAN_x4plus_netD.pth*:\n ```bash\n wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.3/RealESRGAN_x4plus_netD.pth -P experiments/pretrained_models\n ```\n\n**3. Finetune**\n\nModify [options/finetune_realesrgan_x4plus.yml](options/finetune_realesrgan_x4plus.yml) accordingly, especially the `datasets` part:\n\n```yml\ntrain:\n name: DF2K+OST\n type: RealESRGANDataset\n dataroot_gt: datasets/DF2K # modify to the root path of your folder\n meta_info: realesrgan/meta_info/meta_info_DF2Kmultiscale+OST_sub.txt # modify to your own generate meta info txt\n io_backend:\n type: disk\n```\n\nWe use four GPUs for training. We use the `--auto_resume` argument to automatically resume the training if necessary.\n\n```bash\nCUDA_VISIBLE_DEVICES=0,1,2,3 \\\npython -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 realesrgan/train.py -opt options/finetune_realesrgan_x4plus.yml --launcher pytorch --auto_resume\n```\n\nFinetune with **a single GPU**:\n```bash\npython realesrgan/train.py -opt options/finetune_realesrgan_x4plus.yml --auto_resume\n```\n\n### Use your own paired data\n\nYou can also finetune RealESRGAN with your own paired data. It is more similar to fine-tuning ESRGAN.\n\n**1. Prepare dataset**\n\nAssume that you already have two folders:\n\n- **gt folder** (Ground-truth, high-resolution images): *datasets/DF2K/DIV2K_train_HR_sub*\n- **lq folder** (Low quality, low-resolution images): *datasets/DF2K/DIV2K_train_LR_bicubic_X4_sub*\n\nThen, you can prepare the meta_info txt file using the script [scripts/generate_meta_info_pairdata.py](scripts/generate_meta_info_pairdata.py):\n\n```bash\npython scripts/generate_meta_info_pairdata.py --input datasets/DF2K/DIV2K_train_HR_sub datasets/DF2K/DIV2K_train_LR_bicubic_X4_sub --meta_info datasets/DF2K/meta_info/meta_info_DIV2K_sub_pair.txt\n```\n\n**2. Download pre-trained models**\n\nDownload pre-trained models into `experiments/pretrained_models`.\n\n- *RealESRGAN_x4plus.pth*\n ```bash\n wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth -P experiments/pretrained_models\n ```\n\n- *RealESRGAN_x4plus_netD.pth*\n ```bash\n wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.3/RealESRGAN_x4plus_netD.pth -P experiments/pretrained_models\n ```\n\n**3. Finetune**\n\nModify [options/finetune_realesrgan_x4plus_pairdata.yml](options/finetune_realesrgan_x4plus_pairdata.yml) accordingly, especially the `datasets` part:\n\n```yml\ntrain:\n name: DIV2K\n type: RealESRGANPairedDataset\n dataroot_gt: datasets/DF2K # modify to the root path of your folder\n dataroot_lq: datasets/DF2K # modify to the root path of your folder\n meta_info: datasets/DF2K/meta_info/meta_info_DIV2K_sub_pair.txt # modify to your own generate meta info txt\n io_backend:\n type: disk\n```\n\nWe use four GPUs for training. We use the `--auto_resume` argument to automatically resume the training if necessary.\n\n```bash\nCUDA_VISIBLE_DEVICES=0,1,2,3 \\\npython -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 realesrgan/train.py -opt options/finetune_realesrgan_x4plus_pairdata.yml --launcher pytorch --auto_resume\n```\n\nFinetune with **a single GPU**:\n```bash\npython realesrgan/train.py -opt options/finetune_realesrgan_x4plus_pairdata.yml --auto_resume\n```\n"
},
{
"path": "KDSR-GAN/docs/Training_CN.md",
"content": "# :computer: 如何训练/微调 Real-ESRGAN\n\n- [训练 Real-ESRGAN](#训练-real-esrgan)\n - [概述](#概述)\n - [准备数据集](#准备数据集)\n - [训练 Real-ESRNet 模型](#训练-real-esrnet-模型)\n - [训练 Real-ESRGAN 模型](#训练-real-esrgan-模型)\n- [用自己的数据集微调 Real-ESRGAN](#用自己的数据集微调-real-esrgan)\n - [动态生成降级图像](#动态生成降级图像)\n - [使用已配对的数据](#使用已配对的数据)\n\n[English](Training.md) **|** [简体中文](Training_CN.md)\n\n## 训练 Real-ESRGAN\n\n### 概述\n\n训练分为两个步骤。除了 loss 函数外，这两个步骤拥有相同数据合成以及训练的一条龙流程。具体点说：\n\n1. 首先使用 L1 loss 训练 Real-ESRNet 模型，其中 L1 loss 来自预先训练的 ESRGAN 模型。\n\n2. 然后我们将 Real-ESRNet 模型作为生成器初始化，结合L1 loss、感知 loss、GAN loss 三者的参数对 Real-ESRGAN 进行训练。\n\n### 准备数据集\n\n我们使用 DF2K ( DIV2K 和 Flickr2K ) + OST 数据集进行训练。只需要HR图像！
\n下面是网站链接:\n1. DIV2K: http://data.vision.ee.ethz.ch/cvl/DIV2K/DIV2K_train_HR.zip\n2. Flickr2K: https://cv.snu.ac.kr/research/EDSR/Flickr2K.tar\n3. OST: https://openmmlab.oss-cn-hangzhou.aliyuncs.com/datasets/OST_dataset.zip\n\n以下是数据的准备步骤。\n\n#### 第1步：【可选】生成多尺寸图片\n\n针对 DF2K 数据集，我们使用多尺寸缩放策略，*换言之*，我们对 HR 图像进行下采样，就能获得多尺寸的标准参考（Ground-Truth）图像。
\n您可以使用这个 [scripts/generate_multiscale_DF2K.py](scripts/generate_multiscale_DF2K.py) 脚本快速生成多尺寸的图像。
\n注意：如果您只想简单试试，那么可以跳过此步骤。\n\n```bash\npython scripts/generate_multiscale_DF2K.py --input datasets/DF2K/DF2K_HR --output datasets/DF2K/DF2K_multiscale\n```\n\n#### 第2步：【可选】裁切为子图像\n\n我们可以将 DF2K 图像裁切为子图像，以加快 IO 和处理速度。
\n如果你的 IO 够好或储存空间有限，那么此步骤是可选的。
\n\n您可以使用脚本 [scripts/extract_subimages.py](scripts/extract_subimages.py)。这是使用示例:\n\n```bash\n python scripts/extract_subimages.py --input datasets/DF2K/DF2K_multiscale --output datasets/DF2K/DF2K_multiscale_sub --crop_size 400 --step 200\n```\n\n#### 第3步：准备元信息 txt\n\n您需要准备一个包含图像路径的 txt 文件。下面是 `meta_info_DF2Kmultiscale+OST_sub.txt` 中的部分展示（由于各个用户可能有截然不同的子图像划分，这个文件不适合你的需求，你得准备自己的 txt 文件)：\n\n```txt\nDF2K_HR_sub/000001_s001.png\nDF2K_HR_sub/000001_s002.png\nDF2K_HR_sub/000001_s003.png\n...\n```\n\n你可以使用该脚本 [scripts/generate_meta_info.py](scripts/generate_meta_info.py) 生成包含图像路径的 txt 文件。
\n你还可以合并多个文件夹的图像路径到一个元信息（meta_info）txt。这是使用示例:\n\n```bash\n python scripts/generate_meta_info.py --input datasets/DF2K/DF2K_HR, datasets/DF2K/DF2K_multiscale --root datasets/DF2K, datasets/DF2K --meta_info datasets/DF2K/meta_info/meta_info_DF2Kmultiscale.txt\n```\n\n### 训练 Real-ESRNet 模型\n\n1. 下载预先训练的模型 [ESRGAN](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.1/ESRGAN_SRx4_DF2KOST_official-ff704c30.pth)，放到 `experiments/pretrained_models`目录下。\n ```bash\n wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.1/ESRGAN_SRx4_DF2KOST_official-ff704c30.pth -P experiments/pretrained_models\n ```\n2. 相应地修改选项文件 `options/train_realesrnet_x4plus.yml` 中的内容：\n ```yml\n train:\n name: DF2K+OST\n type: RealESRGANDataset\n dataroot_gt: datasets/DF2K # 修改为你的数据集文件夹根目录\n meta_info: realesrgan/meta_info/meta_info_DF2Kmultiscale+OST_sub.txt # 修改为你自己生成的元信息txt\n io_backend:\n type: disk\n ```\n3. 如果你想在训练过程中执行验证，就取消注释这些内容并进行相应的修改：\n ```yml\n # 取消注释这些以进行验证\n # val:\n # name: validation\n # type: PairedImageDataset\n # dataroot_gt: path_to_gt\n # dataroot_lq: path_to_lq\n # io_backend:\n # type: disk\n\n ...\n\n # 取消注释这些以进行验证\n # 验证设置\n # val:\n # val_freq: !!float 5e3\n # save_img: True\n\n # metrics:\n # psnr: # 指标名称，可以是任意的\n # type: calculate_psnr\n # crop_border: 4\n # test_y_channel: false\n ```\n4. 正式训练之前，你可以用 `--debug` 模式检查是否正常运行。我们用了4个GPU进行训练：\n ```bash\n CUDA_VISIBLE_DEVICES=0,1,2,3 \\\n python -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 realesrgan/train.py -opt options/train_realesrnet_x4plus.yml --launcher pytorch --debug\n ```\n\n 用 **1个GPU** 训练的 debug 模式示例:\n ```bash\n python realesrgan/train.py -opt options/train_realesrnet_x4plus.yml --debug\n ```\n5. 正式训练开始。我们用了4个GPU进行训练。还可以使用参数 `--auto_resume` 在必要时自动恢复训练。\n ```bash\n CUDA_VISIBLE_DEVICES=0,1,2,3 \\\n python -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 realesrgan/train.py -opt options/train_realesrnet_x4plus.yml --launcher pytorch --auto_resume\n ```\n\n 用 **1个GPU** 训练：\n ```bash\n python realesrgan/train.py -opt options/train_realesrnet_x4plus.yml --auto_resume\n ```\n\n### 训练 Real-ESRGAN 模型\n\n1. 训练 Real-ESRNet 模型后，您得到了这个 `experiments/train_RealESRNetx4plus_1000k_B12G4_fromESRGAN/model/net_g_1000000.pth` 文件。如果需要指定预训练路径到其他文件，请修改选项文件 `train_realesrgan_x4plus.yml` 中 `pretrain_network_g` 的值。\n1. 修改选项文件 `train_realesrgan_x4plus.yml` 的内容。大多数修改与上节提到的类似。\n1. 正式训练之前，你可以以 `--debug` 模式检查是否正常运行。我们使用了4个GPU进行训练：\n ```bash\n CUDA_VISIBLE_DEVICES=0,1,2,3 \\\n python -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 realesrgan/train.py -opt options/train_realesrgan_x4plus.yml --launcher pytorch --debug\n ```\n\n 用 **1个GPU** 训练的 debug 模式示例:\n ```bash\n python realesrgan/train.py -opt options/train_realesrgan_x4plus.yml --debug\n ```\n1. 正式训练开始。我们使用4个GPU进行训练。还可以使用参数 `--auto_resume` 在必要时自动恢复训练。\n ```bash\n CUDA_VISIBLE_DEVICES=0,1,2,3 \\\n python -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 realesrgan/train.py -opt options/train_realesrgan_x4plus.yml --launcher pytorch --auto_resume\n ```\n\n 用 **1个GPU** 训练：\n ```bash\n python realesrgan/train.py -opt options/train_realesrgan_x4plus.yml --auto_resume\n ```\n\n## 用自己的数据集微调 Real-ESRGAN\n\n你可以用自己的数据集微调 Real-ESRGAN。一般地，微调（Fine-Tune）程序可以分为两种类型:\n\n1. [动态生成降级图像](#动态生成降级图像)\n2. [使用**已配对**的数据](#使用已配对的数据)\n\n### 动态生成降级图像\n\n只需要高分辨率图像。在训练过程中，使用 Real-ESRGAN 描述的降级模型生成低质量图像。\n\n**1. 准备数据集**\n\n完整信息请参见[本节](#准备数据集)。\n\n**2. 下载预训练模型**\n\n下载预先训练的模型到 `experiments/pretrained_models` 目录下。\n\n- *RealESRGAN_x4plus.pth*:\n ```bash\n wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth -P experiments/pretrained_models\n ```\n\n- *RealESRGAN_x4plus_netD.pth*:\n ```bash\n wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.3/RealESRGAN_x4plus_netD.pth -P experiments/pretrained_models\n ```\n\n**3. 微调**\n\n修改选项文件 [options/finetune_realesrgan_x4plus.yml](options/finetune_realesrgan_x4plus.yml) ，特别是 `datasets` 部分：\n\n```yml\ntrain:\n name: DF2K+OST\n type: RealESRGANDataset\n dataroot_gt: datasets/DF2K # 修改为你的数据集文件夹根目录\n meta_info: realesrgan/meta_info/meta_info_DF2Kmultiscale+OST_sub.txt # 修改为你自己生成的元信息txt\n io_backend:\n type: disk\n```\n\n我们使用4个GPU进行训练。还可以使用参数 `--auto_resume` 在必要时自动恢复训练。\n\n```bash\nCUDA_VISIBLE_DEVICES=0,1,2,3 \\\npython -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 realesrgan/train.py -opt options/finetune_realesrgan_x4plus.yml --launcher pytorch --auto_resume\n```\n\n用 **1个GPU** 训练：\n```bash\npython realesrgan/train.py -opt options/finetune_realesrgan_x4plus.yml --auto_resume\n```\n\n### 使用已配对的数据\n\n你还可以用自己已经配对的数据微调 RealESRGAN。这个过程更类似于微调 ESRGAN。\n\n**1. 准备数据集**\n\n假设你已经有两个文件夹（folder）:\n\n- **gt folder**（标准参考，高分辨率图像）：*datasets/DF2K/DIV2K_train_HR_sub*\n- **lq folder**（低质量，低分辨率图像）：*datasets/DF2K/DIV2K_train_LR_bicubic_X4_sub*\n\n然后，您可以使用脚本 [scripts/generate_meta_info_pairdata.py](scripts/generate_meta_info_pairdata.py) 生成元信息（meta_info）txt 文件。\n\n```bash\npython scripts/generate_meta_info_pairdata.py --input datasets/DF2K/DIV2K_train_HR_sub datasets/DF2K/DIV2K_train_LR_bicubic_X4_sub --meta_info datasets/DF2K/meta_info/meta_info_DIV2K_sub_pair.txt\n```\n\n**2. 下载预训练模型**\n\n下载预先训练的模型到 `experiments/pretrained_models` 目录下。\n\n- *RealESRGAN_x4plus.pth*:\n ```bash\n wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth -P experiments/pretrained_models\n ```\n\n- *RealESRGAN_x4plus_netD.pth*:\n ```bash\n wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.3/RealESRGAN_x4plus_netD.pth -P experiments/pretrained_models\n ```\n\n**3. 微调**\n\n修改选项文件 [options/finetune_realesrgan_x4plus_pairdata.yml](options/finetune_realesrgan_x4plus_pairdata.yml) ，特别是 `datasets` 部分：\n\n```yml\ntrain:\n name: DIV2K\n type: RealESRGANPairedDataset\n dataroot_gt: datasets/DF2K # 修改为你的 gt folder 文件夹根目录\n dataroot_lq: datasets/DF2K # 修改为你的 lq folder 文件夹根目录\n meta_info: datasets/DF2K/meta_info/meta_info_DIV2K_sub_pair.txt # 修改为你自己生成的元信息txt\n io_backend:\n type: disk\n```\n\n我们使用4个GPU进行训练。还可以使用参数 `--auto_resume` 在必要时自动恢复训练。\n\n```bash\nCUDA_VISIBLE_DEVICES=0,1,2,3 \\\npython -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 realesrgan/train.py -opt options/finetune_realesrgan_x4plus_pairdata.yml --launcher pytorch --auto_resume\n```\n\n用 **1个GPU** 训练：\n```bash\npython realesrgan/train.py -opt options/finetune_realesrgan_x4plus_pairdata.yml --auto_resume\n```\n"
},
{
"path": "KDSR-GAN/docs/anime_comparisons.md",
"content": "# Comparisons among different anime models\n\n[English](anime_comparisons.md) **|** [简体中文](anime_comparisons_CN.md)\n\n## Update News\n\n- 2022/04/24: Release **AnimeVideo-v3**. We have made the following improvements:\n - **better naturalness**\n - **Fewer artifacts**\n - **more faithful to the original colors**\n - **better texture restoration**\n - **better background restoration**\n\n## Comparisons\n\nWe have compared our RealESRGAN-AnimeVideo-v3 with the following methods.\nOur RealESRGAN-AnimeVideo-v3 can achieve better results with faster inference speed.\n\n- [waifu2x](https://github.com/nihui/waifu2x-ncnn-vulkan) with the hyperparameters: `tile=0`, `noiselevel=2`\n- [Real-CUGAN](https://github.com/bilibili/ailab/tree/main/Real-CUGAN): we use the [20220227](https://github.com/bilibili/ailab/releases/tag/Real-CUGAN-add-faster-low-memory-mode) version, the hyperparameters are: `cache_mode=0`, `tile=0`, `alpha=1`.\n- our RealESRGAN-AnimeVideo-v3\n\n## Results\n\nYou may need to **zoom in** for comparing details, or **click the image** to see in the full size. Please note that the images\nin the table below are the resized and cropped patches from the original images, you can download the original inputs and outputs from [Google Drive](https://drive.google.com/drive/folders/1bc_Hje1Nqop9NDkUvci2VACSjL7HZMRp?usp=sharing) .\n\n**More natural results, better background restoration**\n| Input | waifu2x | Real-CUGAN | RealESRGAN
AnimeVideo-v3 |\n| :---: | :---: | :---: | :---: |\n| |  |  |  |\n| |  |  |  |\n| |  |  |  |\n\n**Fewer artifacts, better detailed textures**\n| Input | waifu2x | Real-CUGAN | RealESRGAN
AnimeVideo-v3 |\n| :---: | :---: | :---: | :---: |\n| |  |  |  |\n| |  |  |  |\n| |  |  |  |\n| |  |  |  |\n\n**Other better results**\n| Input | waifu2x | Real-CUGAN | RealESRGAN
AnimeVideo-v3 |\n| :---: | :---: | :---: | :---: |\n| |  |  |  |\n| |  |  |  |\n|  |   |   |   |\n| |  |  |  |\n| |  |  |  |\n\n## Inference Speed\n\n### PyTorch\n\nNote that we only report the **model** time, and ignore the IO time.\n\n| GPU | Input Resolution | waifu2x | Real-CUGAN | RealESRGAN-AnimeVideo-v3\n| :---: | :---: | :---: | :---: | :---: |\n| V100 | 1921 x 1080 | - | 3.4 fps | **10.0** fps |\n| V100 | 1280 x 720 | - | 7.2 fps | **22.6** fps |\n| V100 | 640 x 480 | - | 24.4 fps | **65.9** fps |\n\n### ncnn\n\n- [ ] TODO\n"
},
{
"path": "KDSR-GAN/docs/anime_comparisons_CN.md",
"content": "# 动漫视频模型比较\n\n[English](anime_comparisons.md) **|** [简体中文](anime_comparisons_CN.md)\n\n## 更新\n\n- 2022/04/24: 发布 **AnimeVideo-v3**. 主要做了以下更新：\n - **更自然**\n - **更少瑕疵**\n - **颜色保持得更好**\n - **更好的纹理恢复**\n - **虚化背景处理**\n\n## 比较\n\n我们将 RealESRGAN-AnimeVideo-v3 与以下方法进行了比较。我们的 RealESRGAN-AnimeVideo-v3 可以以更快的推理速度获得更好的结果。\n\n- [waifu2x](https://github.com/nihui/waifu2x-ncnn-vulkan). 超参数: `tile=0`, `noiselevel=2`\n- [Real-CUGAN](https://github.com/bilibili/ailab/tree/main/Real-CUGAN): 我们使用了[20220227](https://github.com/bilibili/ailab/releases/tag/Real-CUGAN-add-faster-low-memory-mode)版本, 超参: `cache_mode=0`, `tile=0`, `alpha=1`.\n- 我们的 RealESRGAN-AnimeVideo-v3\n\n## 结果\n\n您可能需要**放大**以比较详细信息, 或者**单击图像**以查看完整尺寸。 请注意下面表格的图片是从原图里裁剪patch并且resize后的结果，您可以从\n[Google Drive](https://drive.google.com/drive/folders/1bc_Hje1Nqop9NDkUvci2VACSjL7HZMRp?usp=sharing) 里下载原始的输入和输出。\n\n**更自然的结果，更好的虚化背景恢复**\n\n| 输入 | waifu2x | Real-CUGAN | RealESRGAN
AnimeVideo-v3 |\n| :---: | :---: | :---: | :---: |\n| |  |  |  |\n| |  |  |  |\n| |  |  |  |\n\n**更少瑕疵，更好的细节纹理**\n\n| 输入 | waifu2x | Real-CUGAN | RealESRGAN
AnimeVideo-v3 |\n| :---: | :---: | :---: | :---: |\n| |  |  |  |\n| |  |  |  |\n| |  |  |  |\n| |  |  |  |\n\n**其他更好的结果**\n\n| 输入 | waifu2x | Real-CUGAN | RealESRGAN
AnimeVideo-v3 |\n| :---: | :---: | :---: | :---: |\n| |  |  |  |\n| |  |  |  |\n|  |   |   |   |\n| |  |  |  |\n| |  |  |  |\n\n## 推理速度比较\n\n### PyTorch\n\n请注意，我们只报告了**模型推理**的时间, 而忽略了读写硬盘的时间.\n\n| GPU | 输入尺寸 | waifu2x | Real-CUGAN | RealESRGAN-AnimeVideo-v3\n| :---: | :---: | :---: | :---: | :---: |\n| V100 | 1921 x 1080 | - | 3.4 fps | **10.0** fps |\n| V100 | 1280 x 720 | - | 7.2 fps | **22.6** fps |\n| V100 | 640 x 480 | - | 24.4 fps | **65.9** fps |\n\n### ncnn\n\n- [ ] TODO\n"
},
{
"path": "KDSR-GAN/docs/anime_model.md",
"content": "# Anime Model\n\n:white_check_mark: We add [*RealESRGAN_x4plus_anime_6B.pth*](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth), which is optimized for **anime** images with much smaller model size.\n\n- [How to Use](#how-to-use)\n - [PyTorch Inference](#pytorch-inference)\n - [ncnn Executable File](#ncnn-executable-file)\n- [Comparisons with waifu2x](#comparisons-with-waifu2x)\n- [Comparisons with Sliding Bars](#comparisons-with-sliding-bars)\n\n\n ![](\"https://raw.githubusercontent.com/xinntao/public-figures/master/Real-ESRGAN/cmp_realesrgan_anime_1.png\")
\n
\n\nThe following is a video comparison with sliding bar. You may need to use the full-screen mode for better visual quality, as the original image is large; otherwise, you may encounter aliasing issue.\n\n\n\n## How to Use\n\n### PyTorch Inference\n\nPre-trained models: [RealESRGAN_x4plus_anime_6B](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth)\n\n```bash\n# download model\nwget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth -P experiments/pretrained_models\n# inference\npython inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i inputs\n```\n\n### ncnn Executable File\n\nDownload the latest portable [Windows](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesrgan-ncnn-vulkan-20220424-windows.zip) / [Linux](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesrgan-ncnn-vulkan-20220424-ubuntu.zip) / [MacOS](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesrgan-ncnn-vulkan-20220424-macos.zip) **executable files for Intel/AMD/Nvidia GPU**.\n\nTaking the Windows as example, run:\n\n```bash\n./realesrgan-ncnn-vulkan.exe -i input.jpg -o output.png -n realesrgan-x4plus-anime\n```\n\n## Comparisons with waifu2x\n\nWe compare Real-ESRGAN-anime with [waifu2x](https://github.com/nihui/waifu2x-ncnn-vulkan). We use the `-n 2 -s 4` for waifu2x.\n\n\n \n
\n\n ![](\"https://raw.githubusercontent.com/xinntao/public-figures/master/Real-ESRGAN/cmp_realesrgan_anime_2.png\")
\n
\n\n ![](\"https://raw.githubusercontent.com/xinntao/public-figures/master/Real-ESRGAN/cmp_realesrgan_anime_3.png\")
\n
\n\n ![](\"https://raw.githubusercontent.com/xinntao/public-figures/master/Real-ESRGAN/cmp_realesrgan_anime_4.png\")
\n
\n\n ![](\"https://raw.githubusercontent.com/xinntao/public-figures/master/Real-ESRGAN/cmp_realesrgan_anime_5.png\")
\n
\n\n## Comparisons with Sliding Bars\n\nThe following are video comparisons with sliding bar. You may need to use the full-screen mode for better visual quality, as the original image is large; otherwise, you may encounter aliasing issue.\n\n\n\n\n"
},
{
"path": "KDSR-GAN/docs/anime_video_model.md",
"content": "# Anime Video Models\n\n:white_check_mark: We add small models that are optimized for anime videos :-)
\nMore comparisons can be found in [anime_comparisons.md](anime_comparisons.md)\n\n- [How to Use](#how-to-use)\n- [PyTorch Inference](#pytorch-inference)\n- [ncnn Executable File](#ncnn-executable-file)\n - [Step 1: Use ffmpeg to extract frames from video](#step-1-use-ffmpeg-to-extract-frames-from-video)\n - [Step 2: Inference with Real-ESRGAN executable file](#step-2-inference-with-real-esrgan-executable-file)\n - [Step 3: Merge the enhanced frames back into a video](#step-3-merge-the-enhanced-frames-back-into-a-video)\n- [More Demos](#more-demos)\n\n| Models | Scale | Description |\n| ---------------------------------------------------------------------------------------------------------------------------------- | :---- | :----------------------------- |\n| [realesr-animevideov3](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-animevideov3.pth) | X4 1 | Anime video model with XS size |\n\nNote:
\n1 This model can also be used for X1, X2, X3.\n\n---\n\nThe following are some demos (best view in the full screen mode).\n\n\n\n\n\n\n\n## How to Use\n\n### PyTorch Inference\n\n```bash\n# download model\nwget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-animevideov3.pth -P realesrgan/weights\n# single gpu and single process inference\nCUDA_VISIBLE_DEVICES=0 python inference_realesrgan_video.py -i inputs/video/onepiece_demo.mp4 -n realesr-animevideov3 -s 2 --suffix outx2\n# single gpu and multi process inference (you can use multi-processing to improve GPU utilization)\nCUDA_VISIBLE_DEVICES=0 python inference_realesrgan_video.py -i inputs/video/onepiece_demo.mp4 -n realesr-animevideov3 -s 2 --suffix outx2 --num_process_per_gpu 2\n# multi gpu and multi process inference\nCUDA_VISIBLE_DEVICES=0,1,2,3 python inference_realesrgan_video.py -i inputs/video/onepiece_demo.mp4 -n realesr-animevideov3 -s 2 --suffix outx2 --num_process_per_gpu 2\n```\n\n```console\nUsage:\n--num_process_per_gpu The total number of process is num_gpu * num_process_per_gpu. The bottleneck of\n the program lies on the IO, so the GPUs are usually not fully utilized. To alleviate\n this issue, you can use multi-processing by setting this parameter. As long as it\n does not exceed the CUDA memory\n--extract_frame_first If you encounter ffmpeg error when using multi-processing, you can turn this option on.\n```\n\n### NCNN Executable File\n\n#### Step 1: Use ffmpeg to extract frames from video\n\n```bash\nffmpeg -i onepiece_demo.mp4 -qscale:v 1 -qmin 1 -qmax 1 -vsync 0 tmp_frames/frame%08d.png\n```\n\n- Remember to create the folder `tmp_frames` ahead\n\n#### Step 2: Inference with Real-ESRGAN executable file\n\n1. Download the latest portable [Windows](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesrgan-ncnn-vulkan-20220424-windows.zip) / [Linux](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesrgan-ncnn-vulkan-20220424-ubuntu.zip) / [MacOS](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesrgan-ncnn-vulkan-20220424-macos.zip) **executable files for Intel/AMD/Nvidia GPU**\n\n1. Taking the Windows as example, run:\n\n ```bash\n ./realesrgan-ncnn-vulkan.exe -i tmp_frames -o out_frames -n realesr-animevideov3 -s 2 -f jpg\n ```\n\n - Remember to create the folder `out_frames` ahead\n\n#### Step 3: Merge the enhanced frames back into a video\n\n1. First obtain fps from input videos by\n\n ```bash\n ffmpeg -i onepiece_demo.mp4\n ```\n\n ```console\n Usage:\n -i input video path\n ```\n\n You will get the output similar to the following screenshot.\n\n \n ![](\"https://user-images.githubusercontent.com/17445847/145710145-c4f3accf-b82f-4307-9f20-3803a2c73f57.png\")
\n
\n\n2. Merge frames\n\n ```bash\n ffmpeg -r 23.98 -i out_frames/frame%08d.jpg -c:v libx264 -r 23.98 -pix_fmt yuv420p output.mp4\n ```\n\n ```console\n Usage:\n -i input video path\n -c:v video encoder (usually we use libx264)\n -r fps, remember to modify it to meet your needs\n -pix_fmt pixel format in video\n ```\n\n If you also want to copy audio from the input videos, run:\n\n ```bash\n ffmpeg -r 23.98 -i out_frames/frame%08d.jpg -i onepiece_demo.mp4 -map 0:v:0 -map 1:a:0 -c:a copy -c:v libx264 -r 23.98 -pix_fmt yuv420p output_w_audio.mp4\n ```\n\n ```console\n Usage:\n -i input video path, here we use two input streams\n -c:v video encoder (usually we use libx264)\n -r fps, remember to modify it to meet your needs\n -pix_fmt pixel format in video\n ```\n\n## More Demos\n\n- Input video for One Piece:\n\n \n\n- Out video for One Piece\n\n \n\n**More comparisons**\n\n\n"
},
{
"path": "KDSR-GAN/docs/feedback.md",
"content": "# Feedback 反馈\n\n## 动漫插画模型\n\n1. 视频处理不了: 目前的模型，不是针对视频的，所以视频效果很很不好。我们在探究针对视频的模型了\n1. 景深虚化有问题: 现在的模型把一些景深 和 特意的虚化 都复原了，感觉不好。这个后面我们会考虑把这个信息结合进入。一个简单的做法是识别景深和虚化，然后作为条件告诉神经网络，哪些地方复原强一些，哪些地方复原要弱一些\n1. 不可以调节: 像 Waifu2X 可以调节。可以根据自己的喜好，做调整，但是 Real-ESRGAN-anime 并不可以。导致有些恢复效果过了\n1. 把原来的风格改变了: 不同的动漫插画都有自己的风格，现在的 Real-ESRGAN-anime 倾向于恢复成一种风格（这是受到训练数据集影响的）。风格是动漫很重要的一个要素，所以要尽可能保持\n1. 模型太大: 目前的模型处理太慢，能够更快。这个我们有相关的工作在探究，希望能够尽快有结果，并应用到 Real-ESRGAN 这一系列的模型上\n\nThanks for the [detailed and valuable feedbacks/suggestions](https://github.com/xinntao/Real-ESRGAN/issues/131) by [2ji3150](https://github.com/2ji3150).\n"
},
{
"path": "KDSR-GAN/docs/model_zoo.md",
"content": "# :european_castle: Model Zoo\n\n- [For General Images](#for-general-images)\n- [For Anime Images](#for-anime-images)\n- [For Anime Videos](#for-anime-videos)\n\n---\n\n## For General Images\n\n| Models | Scale | Description |\n| ------------------------------------------------------------------------------------------------------------------------------- | :---- | :------------------------------------------- |\n| [RealESRGAN_x4plus](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth) | X4 | X4 model for general images |\n| [RealESRGAN_x2plus](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth) | X2 | X2 model for general images |\n| [RealESRNet_x4plus](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.1/RealESRNet_x4plus.pth) | X4 | X4 model with MSE loss (over-smooth effects) |\n| [official ESRGAN_x4](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.1/ESRGAN_SRx4_DF2KOST_official-ff704c30.pth) | X4 | official ESRGAN model |\n\nThe following models are **discriminators**, which are usually used for fine-tuning.\n\n| Models | Corresponding model |\n| ---------------------------------------------------------------------------------------------------------------------- | :------------------ |\n| [RealESRGAN_x4plus_netD](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.3/RealESRGAN_x4plus_netD.pth) | RealESRGAN_x4plus |\n| [RealESRGAN_x2plus_netD](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.3/RealESRGAN_x2plus_netD.pth) | RealESRGAN_x2plus |\n\n## For Anime Images / Illustrations\n\n| Models | Scale | Description |\n| ------------------------------------------------------------------------------------------------------------------------------ | :---- | :---------------------------------------------------------- |\n| [RealESRGAN_x4plus_anime_6B](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth) | X4 | Optimized for anime images; 6 RRDB blocks (smaller network) |\n\nThe following models are **discriminators**, which are usually used for fine-tuning.\n\n| Models | Corresponding model |\n| ---------------------------------------------------------------------------------------------------------------------------------------- | :------------------------- |\n| [RealESRGAN_x4plus_anime_6B_netD](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B_netD.pth) | RealESRGAN_x4plus_anime_6B |\n\n## For Animation Videos\n\n| Models | Scale | Description |\n| ---------------------------------------------------------------------------------------------------------------------------------- | :---- | :----------------------------- |\n| [realesr-animevideov3](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-animevideov3.pth) | X41 | Anime video model with XS size |\n\nNote:
\n1 This model can also be used for X1, X2, X3.\n\nThe following models are **discriminators**, which are usually used for fine-tuning.\n\nTODO\n"
},
{
"path": "KDSR-GAN/docs/ncnn_conversion.md",
"content": "# Instructions on converting to NCNN models\n\n1. Convert to onnx model with `scripts/pytorch2onnx.py`. Remember to modify codes accordingly\n1. Convert onnx model to ncnn model\n 1. `cd ncnn-master\\ncnn\\build\\tools\\onnx`\n 1. `onnx2ncnn.exe realesrgan-x4.onnx realesrgan-x4-raw.param realesrgan-x4-raw.bin`\n1. Optimize ncnn model\n 1. fp16 mode\n 1. `cd ncnn-master\\ncnn\\build\\tools`\n 1. `ncnnoptimize.exe realesrgan-x4-raw.param realesrgan-x4-raw.bin realesrgan-x4.param realesrgan-x4.bin 1`\n1. Modify the blob name in `realesrgan-x4.param`: `data` and `output`\n"
},
{
"path": "KDSR-GAN/kdsrgan/__init__.py",
"content": "# flake8: noqa\nfrom .losses import *\nfrom .archs import *\nfrom .data import *\nfrom .models import *\nfrom .utils import *\nfrom .version import *\n"
},
{
"path": "KDSR-GAN/kdsrgan/archs/ST_arch.py",
"content": "from basicsr.utils.registry import ARCH_REGISTRY\nfrom torch.nn import functional as F\nimport torch\nfrom torch import nn\nimport kdsrgan.archs.common as common\n\n\nclass IDR_DDC(nn.Module):\n def __init__(self, channels_in, channels_out, kernel_size, reduction):\n super(IDR_DDC, self).__init__()\n self.channels_out = channels_out\n self.channels_in = channels_in\n self.kernel_size = kernel_size\n\n self.kernel = nn.Sequential(\n nn.Linear(channels_in, channels_in, bias=False),\n nn.LeakyReLU(0.1, True),\n nn.Linear(channels_in, channels_in * self.kernel_size * self.kernel_size, bias=False)\n )\n\n def forward(self, x):\n '''\n :param x[0]: feature map: B * C * H * W\n :param x[1]: degradation representation: B * C\n '''\n b, c, h, w = x[0].size()\n\n # branch 1\n kernel = self.kernel(x[1]).view(-1, 1, self.kernel_size, self.kernel_size)\n out = F.conv2d(x[0].view(1, -1, h, w), kernel, groups=b*c, padding=(self.kernel_size-1)//2)\n out = out.view(b, -1, h, w)\n\n\n return out\n\n\n\nclass IDR_DCRB(nn.Module):\n def __init__(self, conv, n_feat, kernel_size, reduction):\n super(IDR_DCRB, self).__init__()\n\n self.da_conv1 = IDR_DDC(n_feat, n_feat, kernel_size, reduction)\n self.conv1 = conv(n_feat, n_feat, kernel_size)\n self.relu = nn.LeakyReLU(0.1, True)\n\n def forward(self, x):\n '''\n :param x[0]: feature map: B * C * H * W\n :param x[1]: degradation representation: B * C\n '''\n\n out = self.relu(self.da_conv1(x))\n out = self.conv1(out)\n out = out + x[0]\n\n return out\n\n\nclass DAG(nn.Module):\n def __init__(self, conv, n_feat, kernel_size, reduction, n_blocks):\n super(DAG, self).__init__()\n self.n_blocks = n_blocks\n modules_body = [\n IDR_DCRB(conv, n_feat, kernel_size, reduction) \\\n for _ in range(n_blocks)\n ]\n # modules_body.append(conv(n_feat, n_feat, kernel_size))\n\n self.body = nn.Sequential(*modules_body)\n\n def forward(self, x):\n '''\n :param x[0]: feature map: B * C * H * W\n :param x[1]: degradation representation: B * C\n '''\n res = x[0]\n for i in range(self.n_blocks):\n res = self.body[i]([res, x[1]])\n\n return res\n\n\nclass KDSR(nn.Module):\n def __init__(self, n_feats=128, scale=4,n_sr_blocks = 42, rgb_range=1,reduction = 8, conv=common.default_conv):\n super(KDSR, self).__init__()\n kernel_size = 3\n\n # RGB mean for DIV2K\n rgb_mean = (0.4488, 0.4371, 0.4040)\n rgb_std = (1.0, 1.0, 1.0)\n self.sub_mean = common.MeanShift(rgb_range, rgb_mean, rgb_std)\n self.add_mean = common.MeanShift(rgb_range, rgb_mean, rgb_std, 1)\n\n # head module\n modules_head = [conv(3, n_feats, kernel_size)]\n self.head = nn.Sequential(*modules_head)\n\n\n # body\n modules_body = [\n DAG(common.default_conv, n_feats, kernel_size, reduction, n_sr_blocks)\n ]\n modules_body.append(conv(n_feats, n_feats, kernel_size))\n self.body = nn.Sequential(*modules_body)\n\n # tail\n modules_tail = [common.Upsampler(conv, scale, n_feats, act=False),\n conv(n_feats, 3, kernel_size)]\n self.tail = nn.Sequential(*modules_tail)\n\n def forward(self, x, k_v):\n\n # sub mean\n x = self.sub_mean(x)\n\n # head\n x = self.head(x)\n\n # body\n res = x\n res = self.body[0]([res, k_v])\n res = self.body[-1](res)\n res = res + x\n\n # tail\n x = self.tail(res)\n\n # add mean\n x = self.add_mean(x)\n\n return x\n\n\n\nclass KD_IDE(nn.Module):\n def __init__(self,n_feats = 128, n_encoder_res = 6):\n super(KD_IDE, self).__init__()\n E1=[nn.Conv2d(3, n_feats, kernel_size=3, padding=1),\n nn.LeakyReLU(0.1, True)]\n E2=[\n common.ResBlock(\n common.default_conv, n_feats, kernel_size=3\n ) for _ in range(n_encoder_res)\n ]\n E3=[\n nn.Conv2d(n_feats, n_feats * 2, kernel_size=3, padding=1),\n nn.LeakyReLU(0.1, True),\n nn.Conv2d(n_feats * 2, n_feats * 2, kernel_size=3, padding=1),\n nn.LeakyReLU(0.1, True),\n nn.Conv2d(n_feats * 2, n_feats * 4, kernel_size=3, padding=1),\n nn.LeakyReLU(0.1, True),\n nn.AdaptiveAvgPool2d(1),\n ]\n E=E1+E2+E3\n self.E = nn.Sequential(\n *E\n )\n self.mlp = nn.Sequential(\n nn.Linear(n_feats * 4, n_feats * 4),\n nn.LeakyReLU(0.1, True),\n nn.Linear(n_feats * 4, n_feats * 4),\n nn.LeakyReLU(0.1, True)\n )\n self.compress = nn.Sequential(\n nn.Linear(n_feats*4, n_feats),\n nn.LeakyReLU(0.1, True)\n )\n\n def forward(self, x):\n fea = self.E(x).squeeze(-1).squeeze(-1)\n S_fea = []\n fea1 = self.mlp(fea)\n fea = self.compress(fea1)\n S_fea.append(fea1)\n return fea,S_fea\n\n@ARCH_REGISTRY.register()\nclass BlindSR_ST(nn.Module):\n def __init__(self, n_feats=128, n_encoder_res=6, scale=4,n_sr_blocks=42 ):\n super(BlindSR_ST, self).__init__()\n\n # Generator\n self.G = KDSR(n_feats=n_feats, scale=scale,n_sr_blocks=n_sr_blocks)\n\n self.E_st = KD_IDE(n_feats=n_feats, n_encoder_res=n_encoder_res)\n\n self.pixel_unshuffle = nn.PixelUnshuffle(scale)\n\n\n def forward(self, x):\n if self.training:\n # degradation-aware represenetion learning\n deg_repre, S_fea = self.E_st(x)\n\n # degradation-aware SR\n sr = self.G(x, deg_repre)\n\n return sr, S_fea\n else:\n # degradation-aware represenetion learning\n deg_repre, _ = self.E_st(x)\n\n # degradation-aware SR\n sr = self.G(x, deg_repre)\n\n return sr\n"
},
{
"path": "KDSR-GAN/kdsrgan/archs/TA_arch.py",
"content": "from basicsr.utils.registry import ARCH_REGISTRY\nfrom torch.nn import functional as F\nimport torch\nfrom torch import nn\nimport kdsrgan.archs.common as common\n\nclass IDR_DDC(nn.Module):\n def __init__(self, channels_in, channels_out, kernel_size, reduction):\n super(IDR_DDC, self).__init__()\n self.channels_out = channels_out\n self.channels_in = channels_in\n self.kernel_size = kernel_size\n\n self.kernel = nn.Sequential(\n nn.Linear(channels_in, channels_in, bias=False),\n nn.LeakyReLU(0.1, True),\n nn.Linear(channels_in, channels_in * self.kernel_size * self.kernel_size, bias=False)\n )\n\n def forward(self, x):\n '''\n :param x[0]: feature map: B * C * H * W\n :param x[1]: degradation representation: B * C\n '''\n b, c, h, w = x[0].size()\n\n # branch 1\n kernel = self.kernel(x[1]).view(-1, 1, self.kernel_size, self.kernel_size)\n out = F.conv2d(x[0].view(1, -1, h, w), kernel, groups=b*c, padding=(self.kernel_size-1)//2)\n out = out.view(b, -1, h, w)\n\n\n return out\n\n\n\nclass IDR_DCRB(nn.Module):\n def __init__(self, conv, n_feat, kernel_size, reduction):\n super(IDR_DCRB, self).__init__()\n\n self.da_conv1 = IDR_DDC(n_feat, n_feat, kernel_size, reduction)\n self.conv1 = conv(n_feat, n_feat, kernel_size)\n self.relu = nn.LeakyReLU(0.1, True)\n\n def forward(self, x):\n '''\n :param x[0]: feature map: B * C * H * W\n :param x[1]: degradation representation: B * C\n '''\n\n out = self.relu(self.da_conv1(x))\n out = self.conv1(out)\n out = out + x[0]\n\n return out\n\n\nclass DAG(nn.Module):\n def __init__(self, conv, n_feat, kernel_size, reduction, n_blocks):\n super(DAG, self).__init__()\n self.n_blocks = n_blocks\n modules_body = [\n IDR_DCRB(conv, n_feat, kernel_size, reduction) \\\n for _ in range(n_blocks)\n ]\n # modules_body.append(conv(n_feat, n_feat, kernel_size))\n\n self.body = nn.Sequential(*modules_body)\n\n def forward(self, x):\n '''\n :param x[0]: feature map: B * C * H * W\n :param x[1]: degradation representation: B * C\n '''\n res = x[0]\n for i in range(self.n_blocks):\n res = self.body[i]([res, x[1]])\n # res = self.body[-1](res)\n #res = res + x[0]\n\n return res\n\n\nclass KDSR(nn.Module):\n def __init__(self, n_feats=128, scale=4,n_sr_blocks = 42, rgb_range=1,reduction = 8, conv=common.default_conv):\n super(KDSR, self).__init__()\n kernel_size = 3\n\n # RGB mean for DIV2K\n rgb_mean = (0.4488, 0.4371, 0.4040)\n rgb_std = (1.0, 1.0, 1.0)\n self.sub_mean = common.MeanShift(rgb_range, rgb_mean, rgb_std)\n self.add_mean = common.MeanShift(rgb_range, rgb_mean, rgb_std, 1)\n\n # head module\n modules_head = [conv(3, n_feats, kernel_size)]\n self.head = nn.Sequential(*modules_head)\n\n\n # body\n modules_body = [\n DAG(common.default_conv, n_feats, kernel_size, reduction, n_sr_blocks)\n ]\n modules_body.append(conv(n_feats, n_feats, kernel_size))\n self.body = nn.Sequential(*modules_body)\n\n # tail\n modules_tail = [common.Upsampler(conv, scale, n_feats, act=False),\n conv(n_feats, 3, kernel_size)]\n self.tail = nn.Sequential(*modules_tail)\n\n def forward(self, x, k_v):\n\n # sub mean\n x = self.sub_mean(x)\n\n # head\n x = self.head(x)\n\n # body\n res = x\n res = self.body[0]([res, k_v])\n res = self.body[-1](res)\n res = res + x\n\n # tail\n x = self.tail(res)\n\n # add mean\n x = self.add_mean(x)\n\n return x\n\n\n\nclass KD_IDE(nn.Module):\n def __init__(self,n_feats = 128, n_encoder_res = 6):\n super(KD_IDE, self).__init__()\n E1=[nn.Conv2d(51, n_feats, kernel_size=3, padding=1),\n nn.LeakyReLU(0.1, True)]\n E2=[\n common.ResBlock(\n common.default_conv, n_feats, kernel_size=3\n ) for _ in range(n_encoder_res)\n ]\n E3=[\n nn.Conv2d(n_feats, n_feats * 2, kernel_size=3, padding=1),\n nn.LeakyReLU(0.1, True),\n nn.Conv2d(n_feats * 2, n_feats * 2, kernel_size=3, padding=1),\n nn.LeakyReLU(0.1, True),\n nn.Conv2d(n_feats * 2, n_feats * 4, kernel_size=3, padding=1),\n nn.LeakyReLU(0.1, True),\n nn.AdaptiveAvgPool2d(1),\n ]\n E=E1+E2+E3\n self.E = nn.Sequential(\n *E\n )\n self.mlp = nn.Sequential(\n nn.Linear(n_feats * 4, n_feats * 4),\n nn.LeakyReLU(0.1, True),\n nn.Linear(n_feats * 4, n_feats * 4),\n nn.LeakyReLU(0.1, True)\n )\n self.compress = nn.Sequential(\n nn.Linear(n_feats*4, n_feats),\n nn.LeakyReLU(0.1, True)\n )\n\n def forward(self, x):\n fea = self.E(x).squeeze(-1).squeeze(-1)\n T_fea = []\n fea1 = self.mlp(fea)\n fea = self.compress(fea1)\n T_fea.append(fea1)\n return fea,T_fea\n\n@ARCH_REGISTRY.register()\nclass BlindSR_TA(nn.Module):\n def __init__(self, n_feats=128, n_encoder_res=6, scale=4,n_sr_blocks=42 ):\n super(BlindSR_TA, self).__init__()\n\n # Generator\n self.G = KDSR(n_feats=n_feats, scale=scale,n_sr_blocks=n_sr_blocks)\n\n self.E = KD_IDE(n_feats=n_feats, n_encoder_res=n_encoder_res)\n\n self.pixel_unshuffle = nn.PixelUnshuffle(scale)\n\n\n def forward(self, x, gt):\n if self.training:\n hr = self.pixel_unshuffle(gt)\n deg_repre = torch.cat([x, hr], dim=1)\n # degradation-aware represenetion learning\n deg_repre, T_fea = self.E(deg_repre)\n\n # degradation-aware SR\n sr = self.G(x, deg_repre)\n\n return sr, T_fea\n else:\n # degradation-aware represenetion learning\n hr = self.pixel_unshuffle(gt)\n deg_repre = torch.cat([x, hr], dim=1)\n deg_repre, _ = self.E(deg_repre)\n\n # degradation-aware SR\n sr = self.G(x, deg_repre)\n\n return sr"
},
{
"path": "KDSR-GAN/kdsrgan/archs/__init__.py",
"content": "import importlib\nfrom basicsr.utils import scandir\nfrom os import path as osp\n\n# automatically scan and import arch modules for registry\n# scan all the files that end with '_arch.py' under the archs folder\narch_folder = osp.dirname(osp.abspath(__file__))\narch_filenames = [osp.splitext(osp.basename(v))[0] for v in scandir(arch_folder) if v.endswith('_arch.py')]\n# import all the arch modules\n_arch_modules = [importlib.import_module(f'kdsrgan.archs.{file_name}') for file_name in arch_filenames]\n"
},
{
"path": "KDSR-GAN/kdsrgan/archs/common.py",
"content": "import math\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\n\ndef default_conv(in_channels, out_channels, kernel_size, bias=True):\n return nn.Conv2d(in_channels, out_channels, kernel_size, padding=(kernel_size//2), bias=bias)\n\nclass ResBlock(nn.Module):\n def __init__(\n self, conv, n_feats, kernel_size,\n bias=True, bn=False, act=nn.LeakyReLU(0.1, inplace=True), res_scale=1):\n\n super(ResBlock, self).__init__()\n m = []\n for i in range(2):\n m.append(conv(n_feats, n_feats, kernel_size, bias=bias))\n if bn:\n m.append(nn.BatchNorm2d(n_feats))\n if i == 0:\n m.append(act)\n\n self.body = nn.Sequential(*m)\n # self.res_scale = res_scale\n\n def forward(self, x):\n res = self.body(x)\n res += x\n\n return res\n\nclass MeanShift(nn.Conv2d):\n def __init__(self, rgb_range, rgb_mean, rgb_std, sign=-1):\n super(MeanShift, self).__init__(3, 3, kernel_size=1)\n std = torch.Tensor(rgb_std)\n self.weight.data = torch.eye(3).view(3, 3, 1, 1)\n self.weight.data.div_(std.view(3, 1, 1, 1))\n self.bias.data = sign * rgb_range * torch.Tensor(rgb_mean)\n self.bias.data.div_(std)\n self.weight.requires_grad = False\n self.bias.requires_grad = False\n\n\nclass Upsampler(nn.Sequential):\n def __init__(self, conv, scale, n_feat, act=False, bias=True):\n m = []\n if (int(scale) & (int(scale) - 1)) == 0: # Is scale = 2^n?\n for _ in range(int(math.log(scale, 2))):\n m.append(conv(n_feat, 4 * n_feat, 3, bias))\n m.append(nn.PixelShuffle(2))\n if act: m.append(act())\n elif scale == 3:\n m.append(conv(n_feat, 9 * n_feat, 3, bias))\n m.append(nn.PixelShuffle(3))\n if act: m.append(act())\n else:\n raise NotImplementedError\n\n super(Upsampler, self).__init__(*m)"
},
{
"path": "KDSR-GAN/kdsrgan/archs/discriminator_arch.py",
"content": "from basicsr.utils.registry import ARCH_REGISTRY\nfrom torch import nn as nn\nfrom torch.nn import functional as F\nfrom torch.nn.utils import spectral_norm\n\n\n@ARCH_REGISTRY.register()\nclass UNetDiscriminatorSN(nn.Module):\n \"\"\"Defines a U-Net discriminator with spectral normalization (SN)\n\n It is used in Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data.\n\n Arg:\n num_in_ch (int): Channel number of inputs. Default: 3.\n num_feat (int): Channel number of base intermediate features. Default: 64.\n skip_connection (bool): Whether to use skip connections between U-Net. Default: True.\n \"\"\"\n\n def __init__(self, num_in_ch, num_feat=64, skip_connection=True):\n super(UNetDiscriminatorSN, self).__init__()\n self.skip_connection = skip_connection\n norm = spectral_norm\n # the first convolution\n self.conv0 = nn.Conv2d(num_in_ch, num_feat, kernel_size=3, stride=1, padding=1)\n # downsample\n self.conv1 = norm(nn.Conv2d(num_feat, num_feat * 2, 4, 2, 1, bias=False))\n self.conv2 = norm(nn.Conv2d(num_feat * 2, num_feat * 4, 4, 2, 1, bias=False))\n self.conv3 = norm(nn.Conv2d(num_feat * 4, num_feat * 8, 4, 2, 1, bias=False))\n # upsample\n self.conv4 = norm(nn.Conv2d(num_feat * 8, num_feat * 4, 3, 1, 1, bias=False))\n self.conv5 = norm(nn.Conv2d(num_feat * 4, num_feat * 2, 3, 1, 1, bias=False))\n self.conv6 = norm(nn.Conv2d(num_feat * 2, num_feat, 3, 1, 1, bias=False))\n # extra convolutions\n self.conv7 = norm(nn.Conv2d(num_feat, num_feat, 3, 1, 1, bias=False))\n self.conv8 = norm(nn.Conv2d(num_feat, num_feat, 3, 1, 1, bias=False))\n self.conv9 = nn.Conv2d(num_feat, 1, 3, 1, 1)\n\n def forward(self, x):\n # downsample\n x0 = F.leaky_relu(self.conv0(x), negative_slope=0.2, inplace=True)\n x1 = F.leaky_relu(self.conv1(x0), negative_slope=0.2, inplace=True)\n x2 = F.leaky_relu(self.conv2(x1), negative_slope=0.2, inplace=True)\n x3 = F.leaky_relu(self.conv3(x2), negative_slope=0.2, inplace=True)\n\n # upsample\n x3 = F.interpolate(x3, scale_factor=2, mode='bilinear', align_corners=False)\n x4 = F.leaky_relu(self.conv4(x3), negative_slope=0.2, inplace=True)\n\n if self.skip_connection:\n x4 = x4 + x2\n x4 = F.interpolate(x4, scale_factor=2, mode='bilinear', align_corners=False)\n x5 = F.leaky_relu(self.conv5(x4), negative_slope=0.2, inplace=True)\n\n if self.skip_connection:\n x5 = x5 + x1\n x5 = F.interpolate(x5, scale_factor=2, mode='bilinear', align_corners=False)\n x6 = F.leaky_relu(self.conv6(x5), negative_slope=0.2, inplace=True)\n\n if self.skip_connection:\n x6 = x6 + x0\n\n # extra convolutions\n out = F.leaky_relu(self.conv7(x6), negative_slope=0.2, inplace=True)\n out = F.leaky_relu(self.conv8(out), negative_slope=0.2, inplace=True)\n out = self.conv9(out)\n\n return out\n"
},
{
"path": "KDSR-GAN/kdsrgan/archs/srvgg_arch.py",
"content": "from basicsr.utils.registry import ARCH_REGISTRY\nfrom torch import nn as nn\nfrom torch.nn import functional as F\n\n\n@ARCH_REGISTRY.register()\nclass SRVGGNetCompact(nn.Module):\n \"\"\"A compact VGG-style network structure for super-resolution.\n\n It is a compact network structure, which performs upsampling in the last layer and no convolution is\n conducted on the HR feature space.\n\n Args:\n num_in_ch (int): Channel number of inputs. Default: 3.\n num_out_ch (int): Channel number of outputs. Default: 3.\n num_feat (int): Channel number of intermediate features. Default: 64.\n num_conv (int): Number of convolution layers in the body network. Default: 16.\n upscale (int): Upsampling factor. Default: 4.\n act_type (str): Activation type, options: 'relu', 'prelu', 'leakyrelu'. Default: prelu.\n \"\"\"\n\n def __init__(self, num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=16, upscale=4, act_type='prelu'):\n super(SRVGGNetCompact, self).__init__()\n self.num_in_ch = num_in_ch\n self.num_out_ch = num_out_ch\n self.num_feat = num_feat\n self.num_conv = num_conv\n self.upscale = upscale\n self.act_type = act_type\n\n self.body = nn.ModuleList()\n # the first conv\n self.body.append(nn.Conv2d(num_in_ch, num_feat, 3, 1, 1))\n # the first activation\n if act_type == 'relu':\n activation = nn.ReLU(inplace=True)\n elif act_type == 'prelu':\n activation = nn.PReLU(num_parameters=num_feat)\n elif act_type == 'leakyrelu':\n activation = nn.LeakyReLU(negative_slope=0.1, inplace=True)\n self.body.append(activation)\n\n # the body structure\n for _ in range(num_conv):\n self.body.append(nn.Conv2d(num_feat, num_feat, 3, 1, 1))\n # activation\n if act_type == 'relu':\n activation = nn.ReLU(inplace=True)\n elif act_type == 'prelu':\n activation = nn.PReLU(num_parameters=num_feat)\n elif act_type == 'leakyrelu':\n activation = nn.LeakyReLU(negative_slope=0.1, inplace=True)\n self.body.append(activation)\n\n # the last conv\n self.body.append(nn.Conv2d(num_feat, num_out_ch * upscale * upscale, 3, 1, 1))\n # upsample\n self.upsampler = nn.PixelShuffle(upscale)\n\n def forward(self, x):\n out = x\n for i in range(0, len(self.body)):\n out = self.body[i](out)\n\n out = self.upsampler(out)\n # add the nearest upsampled image, so that the network learns the residual\n base = F.interpolate(x, scale_factor=self.upscale, mode='nearest')\n out += base\n return out\n"
},
{
"path": "KDSR-GAN/kdsrgan/data/__init__.py",
"content": "import importlib\nfrom basicsr.utils import scandir\nfrom os import path as osp\n\n# automatically scan and import dataset modules for registry\n# scan all the files that end with '_dataset.py' under the data folder\ndata_folder = osp.dirname(osp.abspath(__file__))\ndataset_filenames = [osp.splitext(osp.basename(v))[0] for v in scandir(data_folder) if v.endswith('_dataset.py')]\n# import all the dataset modules\n_dataset_modules = [importlib.import_module(f'kdsrgan.data.{file_name}') for file_name in dataset_filenames]\n"
},
{
"path": "KDSR-GAN/kdsrgan/data/kdsrgan_dataset.py",
"content": "import cv2\nimport math\nimport numpy as np\nimport os\nimport os.path as osp\nimport random\nimport time\nimport torch\nfrom basicsr.data.degradations import circular_lowpass_kernel, random_mixed_kernels\nfrom basicsr.data.transforms import augment\nfrom basicsr.utils import FileClient, get_root_logger, imfrombytes, img2tensor\nfrom basicsr.utils.registry import DATASET_REGISTRY\nfrom torch.utils import data as data\n\n\n@DATASET_REGISTRY.register()\nclass KDSRGANDataset(data.Dataset):\n \"\"\"Dataset used for KDSRGAN model:\n KDSRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data.\n\n It loads gt (Ground-Truth) images, and augments them.\n It also generates blur kernels and sinc kernels for generating low-quality images.\n Note that the low-quality images are processed in tensors on GPUS for faster processing.\n\n Args:\n opt (dict): Config for train datasets. It contains the following keys:\n dataroot_gt (str): Data root path for gt.\n meta_info (str): Path for meta information file.\n io_backend (dict): IO backend type and other kwarg.\n use_hflip (bool): Use horizontal flips.\n use_rot (bool): Use rotation (use vertical flip and transposing h and w for implementation).\n Please see more options in the codes.\n \"\"\"\n\n def __init__(self, opt):\n super(KDSRGANDataset, self).__init__()\n self.opt = opt\n self.file_client = None\n self.io_backend_opt = opt['io_backend']\n self.gt_folder = opt['dataroot_gt']\n\n # file client (lmdb io backend)\n if self.io_backend_opt['type'] == 'lmdb':\n self.io_backend_opt['db_paths'] = [self.gt_folder]\n self.io_backend_opt['client_keys'] = ['gt']\n if not self.gt_folder.endswith('.lmdb'):\n raise ValueError(f\"'dataroot_gt' should end with '.lmdb', but received {self.gt_folder}\")\n with open(osp.join(self.gt_folder, 'meta_info.txt')) as fin:\n self.paths = [line.split('.')[0] for line in fin]\n else:\n # disk backend with meta_info\n # Each line in the meta_info describes the relative path to an image\n with open(self.opt['meta_info']) as fin:\n paths = [line.strip().split(' ')[0] for line in fin]\n self.paths = [os.path.join(self.gt_folder, v) for v in paths]\n\n # blur settings for the first degradation\n self.blur_kernel_size = opt['blur_kernel_size']\n self.kernel_list = opt['kernel_list']\n self.kernel_prob = opt['kernel_prob'] # a list for each kernel probability\n self.blur_sigma = opt['blur_sigma']\n self.betag_range = opt['betag_range'] # betag used in generalized Gaussian blur kernels\n self.betap_range = opt['betap_range'] # betap used in plateau blur kernels\n self.sinc_prob = opt['sinc_prob'] # the probability for sinc filters\n\n # blur settings for the second degradation\n self.blur_kernel_size2 = opt['blur_kernel_size2']\n self.kernel_list2 = opt['kernel_list2']\n self.kernel_prob2 = opt['kernel_prob2']\n self.blur_sigma2 = opt['blur_sigma2']\n self.betag_range2 = opt['betag_range2']\n self.betap_range2 = opt['betap_range2']\n self.sinc_prob2 = opt['sinc_prob2']\n\n # a final sinc filter\n self.final_sinc_prob = opt['final_sinc_prob']\n\n self.kernel_range = [2 * v + 1 for v in range(3, 11)] # kernel size ranges from 7 to 21\n # TODO: kernel range is now hard-coded, should be in the configure file\n self.pulse_tensor = torch.zeros(21, 21).float() # convolving with pulse tensor brings no blurry effect\n self.pulse_tensor[10, 10] = 1\n\n def __getitem__(self, index):\n if self.file_client is None:\n self.file_client = FileClient(self.io_backend_opt.pop('type'), **self.io_backend_opt)\n\n # -------------------------------- Load gt images -------------------------------- #\n # Shape: (h, w, c); channel order: BGR; image range: [0, 1], float32.\n gt_path = self.paths[index]\n # avoid errors caused by high latency in reading files\n retry = 3\n while retry > 0:\n try:\n img_bytes = self.file_client.get(gt_path, 'gt')\n except (IOError, OSError) as e:\n logger = get_root_logger()\n logger.warn(f'File client error: {e}, remaining retry times: {retry - 1}')\n # change another file to read\n index = random.randint(0, self.__len__())\n gt_path = self.paths[index]\n time.sleep(1) # sleep 1s for occasional server congestion\n else:\n break\n finally:\n retry -= 1\n img_gt = imfrombytes(img_bytes, float32=True)\n\n # -------------------- Do augmentation for training: flip, rotation -------------------- #\n img_gt = augment(img_gt, self.opt['use_hflip'], self.opt['use_rot'])\n\n # crop or pad to 400\n # TODO: 400 is hard-coded. You may change it accordingly\n h, w = img_gt.shape[0:2]\n crop_pad_size = 400\n # pad\n if h < crop_pad_size or w < crop_pad_size:\n pad_h = max(0, crop_pad_size - h)\n pad_w = max(0, crop_pad_size - w)\n img_gt = cv2.copyMakeBorder(img_gt, 0, pad_h, 0, pad_w, cv2.BORDER_REFLECT_101)\n # crop\n if img_gt.shape[0] > crop_pad_size or img_gt.shape[1] > crop_pad_size:\n h, w = img_gt.shape[0:2]\n # randomly choose top and left coordinates\n top = random.randint(0, h - crop_pad_size)\n left = random.randint(0, w - crop_pad_size)\n img_gt = img_gt[top:top + crop_pad_size, left:left + crop_pad_size, ...]\n\n # ------------------------ Generate kernels (used in the first degradation) ------------------------ #\n kernel_size = random.choice(self.kernel_range)\n if np.random.uniform() < self.opt['sinc_prob']:\n # this sinc filter setting is for kernels ranging from [7, 21]\n if kernel_size < 13:\n omega_c = np.random.uniform(np.pi / 3, np.pi)\n else:\n omega_c = np.random.uniform(np.pi / 5, np.pi)\n kernel = circular_lowpass_kernel(omega_c, kernel_size, pad_to=False)\n else:\n kernel = random_mixed_kernels(\n self.kernel_list,\n self.kernel_prob,\n kernel_size,\n self.blur_sigma,\n self.blur_sigma, [-math.pi, math.pi],\n self.betag_range,\n self.betap_range,\n noise_range=None)\n # pad kernel\n pad_size = (21 - kernel_size) // 2\n kernel = np.pad(kernel, ((pad_size, pad_size), (pad_size, pad_size)))\n\n # ------------------------ Generate kernels (used in the second degradation) ------------------------ #\n kernel_size = random.choice(self.kernel_range)\n if np.random.uniform() < self.opt['sinc_prob2']:\n if kernel_size < 13:\n omega_c = np.random.uniform(np.pi / 3, np.pi)\n else:\n omega_c = np.random.uniform(np.pi / 5, np.pi)\n kernel2 = circular_lowpass_kernel(omega_c, kernel_size, pad_to=False)\n else:\n kernel2 = random_mixed_kernels(\n self.kernel_list2,\n self.kernel_prob2,\n kernel_size,\n self.blur_sigma2,\n self.blur_sigma2, [-math.pi, math.pi],\n self.betag_range2,\n self.betap_range2,\n noise_range=None)\n\n # pad kernel\n pad_size = (21 - kernel_size) // 2\n kernel2 = np.pad(kernel2, ((pad_size, pad_size), (pad_size, pad_size)))\n\n # ------------------------------------- the final sinc kernel ------------------------------------- #\n if np.random.uniform() < self.opt['final_sinc_prob']:\n kernel_size = random.choice(self.kernel_range)\n omega_c = np.random.uniform(np.pi / 3, np.pi)\n sinc_kernel = circular_lowpass_kernel(omega_c, kernel_size, pad_to=21)\n sinc_kernel = torch.FloatTensor(sinc_kernel)\n else:\n sinc_kernel = self.pulse_tensor\n\n # BGR to RGB, HWC to CHW, numpy to tensor\n img_gt = img2tensor([img_gt], bgr2rgb=True, float32=True)[0]\n kernel = torch.FloatTensor(kernel)\n kernel2 = torch.FloatTensor(kernel2)\n\n return_d = {'gt': img_gt, 'kernel1': kernel, 'kernel2': kernel2, 'sinc_kernel': sinc_kernel, 'gt_path': gt_path}\n return return_d\n\n def __len__(self):\n return len(self.paths)\n"
},
{
"path": "KDSR-GAN/kdsrgan/data/kdsrgan_paired_dataset.py",
"content": "import os\nfrom basicsr.data.data_util import paired_paths_from_folder, paired_paths_from_lmdb\nfrom basicsr.data.transforms import augment, paired_random_crop\nfrom basicsr.utils import FileClient, imfrombytes, img2tensor\nfrom basicsr.utils.registry import DATASET_REGISTRY\nfrom torch.utils import data as data\nfrom torchvision.transforms.functional import normalize\n\n\n@DATASET_REGISTRY.register()\nclass KDSRGANPairedDataset(data.Dataset):\n \"\"\"Paired image dataset for image restoration.\n\n Read LQ (Low Quality, e.g. LR (Low Resolution), blurry, noisy, etc) and GT image pairs.\n\n There are three modes:\n 1. 'lmdb': Use lmdb files.\n If opt['io_backend'] == lmdb.\n 2. 'meta_info': Use meta information file to generate paths.\n If opt['io_backend'] != lmdb and opt['meta_info'] is not None.\n 3. 'folder': Scan folders to generate paths.\n The rest.\n\n Args:\n opt (dict): Config for train datasets. It contains the following keys:\n dataroot_gt (str): Data root path for gt.\n dataroot_lq (str): Data root path for lq.\n meta_info (str): Path for meta information file.\n io_backend (dict): IO backend type and other kwarg.\n filename_tmpl (str): Template for each filename. Note that the template excludes the file extension.\n Default: '{}'.\n gt_size (int): Cropped patched size for gt patches.\n use_hflip (bool): Use horizontal flips.\n use_rot (bool): Use rotation (use vertical flip and transposing h\n and w for implementation).\n\n scale (bool): Scale, which will be added automatically.\n phase (str): 'train' or 'val'.\n \"\"\"\n\n def __init__(self, opt):\n super(KDSRGANPairedDataset, self).__init__()\n self.opt = opt\n self.file_client = None\n self.io_backend_opt = opt['io_backend']\n # mean and std for normalizing the input images\n self.mean = opt['mean'] if 'mean' in opt else None\n self.std = opt['std'] if 'std' in opt else None\n\n self.gt_folder, self.lq_folder = opt['dataroot_gt'], opt['dataroot_lq']\n self.filename_tmpl = opt['filename_tmpl'] if 'filename_tmpl' in opt else '{}'\n\n # file client (lmdb io backend)\n if self.io_backend_opt['type'] == 'lmdb':\n self.io_backend_opt['db_paths'] = [self.lq_folder, self.gt_folder]\n self.io_backend_opt['client_keys'] = ['lq', 'gt']\n self.paths = paired_paths_from_lmdb([self.lq_folder, self.gt_folder], ['lq', 'gt'])\n elif 'meta_info' in self.opt and self.opt['meta_info'] is not None:\n # disk backend with meta_info\n # Each line in the meta_info describes the relative path to an image\n with open(self.opt['meta_info']) as fin:\n paths = [line.strip() for line in fin]\n self.paths = []\n for path in paths:\n gt_path, lq_path = path.split(', ')\n gt_path = os.path.join(self.gt_folder, gt_path)\n lq_path = os.path.join(self.lq_folder, lq_path)\n self.paths.append(dict([('gt_path', gt_path), ('lq_path', lq_path)]))\n else:\n # disk backend\n # it will scan the whole folder to get meta info\n # it will be time-consuming for folders with too many files. It is recommended using an extra meta txt file\n self.paths = paired_paths_from_folder([self.lq_folder, self.gt_folder], ['lq', 'gt'], self.filename_tmpl)\n\n def __getitem__(self, index):\n if self.file_client is None:\n self.file_client = FileClient(self.io_backend_opt.pop('type'), **self.io_backend_opt)\n\n scale = self.opt['scale']\n\n # Load gt and lq images. Dimension order: HWC; channel order: BGR;\n # image range: [0, 1], float32.\n gt_path = self.paths[index]['gt_path']\n img_bytes = self.file_client.get(gt_path, 'gt')\n img_gt = imfrombytes(img_bytes, float32=True)\n lq_path = self.paths[index]['lq_path']\n img_bytes = self.file_client.get(lq_path, 'lq')\n img_lq = imfrombytes(img_bytes, float32=True)\n\n # augmentation for training\n if self.opt['phase'] == 'train':\n gt_size = self.opt['gt_size']\n # random crop\n img_gt, img_lq = paired_random_crop(img_gt, img_lq, gt_size, scale, gt_path)\n # flip, rotation\n img_gt, img_lq = augment([img_gt, img_lq], self.opt['use_hflip'], self.opt['use_rot'])\n\n # BGR to RGB, HWC to CHW, numpy to tensor\n img_gt, img_lq = img2tensor([img_gt, img_lq], bgr2rgb=True, float32=True)\n # normalize\n if self.mean is not None or self.std is not None:\n normalize(img_lq, self.mean, self.std, inplace=True)\n normalize(img_gt, self.mean, self.std, inplace=True)\n\n return {'lq': img_lq, 'gt': img_gt, 'lq_path': lq_path, 'gt_path': gt_path}\n\n def __len__(self):\n return len(self.paths)\n"
},
{
"path": "KDSR-GAN/kdsrgan/losses/__init__.py",
"content": "import importlib\nfrom basicsr.utils import scandir\nfrom os import path as osp\n\n# automatically scan and import arch modules for registry\n# scan all the files that end with '_arch.py' under the archs folder\narch_folder = osp.dirname(osp.abspath(__file__))\narch_filenames = [osp.splitext(osp.basename(v))[0] for v in scandir(arch_folder) if v.endswith('_loss.py')]\n# import all the arch modules\n_arch_modules = [importlib.import_module(f'kdsrgan.losses.{file_name}') for file_name in arch_filenames]\n"
},
{
"path": "KDSR-GAN/kdsrgan/losses/my_loss.py",
"content": "import torch\nfrom torch import nn as nn\nfrom torch.nn import functional as F\nfrom basicsr.utils.registry import LOSS_REGISTRY\n\n\n@LOSS_REGISTRY.register()\nclass KDLoss(nn.Module):\n \"\"\"Knowledge distillation loss.\n Args:\n loss_weight (float): Loss weight for KD loss. Default: 1.0.\n \"\"\"\n\n def __init__(self, loss_weight=1.0, temperature = 0.15):\n super(KDLoss, self).__init__()\n \n self.loss_weight = loss_weight\n self.temperature = temperature\n\n def forward(self, T_fea, S_fea):\n \"\"\"\n Args:\n T_fea (List): contain shape (N, L) vector of BlindSR_TA. \n S_fea (List): contain shape (N, L) vector of BlindSR_ST.\n weight (Tensor, optional): of shape (N, C, H, W). Element-wise weights. Default: None.\n \"\"\"\n loss_distill_dis = 0\n for i in range(len(T_fea)):\n student_distance = F.log_softmax(S_fea[i] / self.temperature, dim=1)\n teacher_distance = F.softmax(T_fea[i].detach()/ self.temperature, dim=1)\n loss_distill_dis += F.kl_div(\n student_distance, teacher_distance, reduction='batchmean')\n #loss_distill_abs += nn.L1Loss()(S_fea[i], T_fea[i].detach())\n return self.loss_weight * loss_distill_dis\n\n "
},
{
"path": "KDSR-GAN/kdsrgan/models/__init__.py",
"content": "import importlib\nfrom basicsr.utils import scandir\nfrom os import path as osp\n\n# automatically scan and import model modules for registry\n# scan all the files that end with '_model.py' under the model folder\nmodel_folder = osp.dirname(osp.abspath(__file__))\nmodel_filenames = [osp.splitext(osp.basename(v))[0] for v in scandir(model_folder) if v.endswith('_model.py')]\n# import all the model modules\n_model_modules = [importlib.import_module(f'kdsrgan.models.{file_name}') for file_name in model_filenames]\n"
},
{
"path": "KDSR-GAN/kdsrgan/models/kdsrgan_ST_model.py",
"content": "import numpy as np\nimport random\nimport torch\nfrom basicsr.data.degradations import random_add_gaussian_noise_pt, random_add_poisson_noise_pt\nfrom basicsr.data.transforms import paired_random_crop\nfrom basicsr.models.srgan_model import SRGANModel\nfrom basicsr.utils import DiffJPEG, USMSharp\nfrom basicsr.utils.img_process_util import filter2D\nfrom basicsr.utils.registry import MODEL_REGISTRY\nfrom collections import OrderedDict\nfrom torch.nn import functional as F\nfrom basicsr.archs import build_network\nfrom basicsr.utils import get_root_logger\nfrom basicsr.losses import build_loss\nfrom torch import nn\n\n@MODEL_REGISTRY.register()\nclass KDSRGANSTModel(SRGANModel):\n \"\"\"\n\n It mainly performs:\n 1. randomly synthesize LQ images in GPU tensors\n 2. optimize the networks with GAN training.\n \"\"\"\n\n def __init__(self, opt):\n super(KDSRGANSTModel, self).__init__(opt)\n self.jpeger = DiffJPEG(differentiable=False).cuda() # simulate JPEG compression artifacts\n self.usm_sharpener = USMSharp().cuda() # do usm sharpening\n self.queue_size = opt.get('queue_size', 180)\n\n self.net_g_TA = build_network(opt['network_TA'])\n self.net_g_TA = self.model_to_device(self.net_g_TA)\n load_path = self.opt['path'].get('pretrain_network_TA', None)\n\n if load_path is not None:\n param_key = self.opt['path'].get('param_key_g', 'params')\n self.load_network(self.net_g_TA, load_path, True, param_key)\n\n self.net_g_TA.eval()\n if self.opt['dist']:\n self.model_Est = self.net_g.module.E_st\n self.model_Eta = self.net_g_TA.module.E\n else:\n self.model_Est = self.net_g.E_st\n self.model_Eta = self.net_g_TA.E\n\n self.pixel_unshuffle = nn.PixelUnshuffle(opt[\"scale\"])\n if self.is_train:\n self.encoder_iter = opt[\"train\"][\"encoder_iter\"]\n self.lr_encoder = opt[\"train\"][\"lr_encoder\"]\n self.lr_sr = opt[\"train\"][\"lr_sr\"]\n self.gamma_encoder = opt[\"train\"][\"gamma_encoder\"]\n self.gamma_sr = opt[\"train\"][\"gamma_sr\"]\n self.lr_decay_encoder = opt[\"train\"][\"lr_decay_encoder\"]\n self.lr_decay_sr = opt[\"train\"][\"lr_decay_sr\"]\n\n def init_training_settings(self):\n train_opt = self.opt['train']\n\n if train_opt.get('kd_opt'):\n self.cri_kd = build_loss(train_opt['kd_opt']).to(self.device)\n else:\n self.cri_kd = None\n super(KDSRGANSTModel, self).init_training_settings()\n \n\n @torch.no_grad()\n def _dequeue_and_enqueue(self):\n \"\"\"It is the training pair pool for increasing the diversity in a batch.\n\n Batch processing limits the diversity of synthetic degradations in a batch. For example, samples in a\n batch could not have different resize scaling factors. Therefore, we employ this training pair pool\n to increase the degradation diversity in a batch.\n \"\"\"\n # initialize\n b, c, h, w = self.lq.size()\n if not hasattr(self, 'queue_lr'):\n assert self.queue_size % b == 0, f'queue size {self.queue_size} should be divisible by batch size {b}'\n self.queue_lr = torch.zeros(self.queue_size, c, h, w).cuda()\n _, c, h, w = self.gt.size()\n self.queue_gt = torch.zeros(self.queue_size, c, h, w).cuda()\n self.queue_ptr = 0\n if self.queue_ptr == self.queue_size: # the pool is full\n # do dequeue and enqueue\n # shuffle\n idx = torch.randperm(self.queue_size)\n self.queue_lr = self.queue_lr[idx]\n self.queue_gt = self.queue_gt[idx]\n # get first b samples\n lq_dequeue = self.queue_lr[0:b, :, :, :].clone()\n gt_dequeue = self.queue_gt[0:b, :, :, :].clone()\n # update the queue\n self.queue_lr[0:b, :, :, :] = self.lq.clone()\n self.queue_gt[0:b, :, :, :] = self.gt.clone()\n\n self.lq = lq_dequeue\n self.gt = gt_dequeue\n else:\n # only do enqueue\n self.queue_lr[self.queue_ptr:self.queue_ptr + b, :, :, :] = self.lq.clone()\n self.queue_gt[self.queue_ptr:self.queue_ptr + b, :, :, :] = self.gt.clone()\n self.queue_ptr = self.queue_ptr + b\n\n @torch.no_grad()\n def feed_data(self, data):\n \"\"\"Accept data from dataloader, and then add two-order degradations to obtain LQ images.\n \"\"\"\n if self.is_train and self.opt.get('high_order_degradation', True):\n # training data synthesis\n self.gt = data['gt'].to(self.device)\n self.gt_usm = self.usm_sharpener(self.gt)\n\n self.kernel1 = data['kernel1'].to(self.device)\n self.kernel2 = data['kernel2'].to(self.device)\n self.sinc_kernel = data['sinc_kernel'].to(self.device)\n\n ori_h, ori_w = self.gt.size()[2:4]\n\n # ----------------------- The first degradation process ----------------------- #\n # blur\n out = filter2D(self.gt, self.kernel1)\n # random resize\n updown_type = random.choices(['up', 'down', 'keep'], self.opt['resize_prob'])[0]\n if updown_type == 'up':\n scale = np.random.uniform(1, self.opt['resize_range'][1])\n elif updown_type == 'down':\n scale = np.random.uniform(self.opt['resize_range'][0], 1)\n else:\n scale = 1\n mode = random.choice(['area', 'bilinear', 'bicubic'])\n out = F.interpolate(out, scale_factor=scale, mode=mode)\n # add noise\n gray_noise_prob = self.opt['gray_noise_prob']\n if np.random.uniform() < self.opt['gaussian_noise_prob']:\n out = random_add_gaussian_noise_pt(\n out, sigma_range=self.opt['noise_range'], clip=True, rounds=False, gray_prob=gray_noise_prob)\n else:\n out = random_add_poisson_noise_pt(\n out,\n scale_range=self.opt['poisson_scale_range'],\n gray_prob=gray_noise_prob,\n clip=True,\n rounds=False)\n # JPEG compression\n jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range'])\n out = torch.clamp(out, 0, 1) # clamp to [0, 1], otherwise JPEGer will result in unpleasant artifacts\n out = self.jpeger(out, quality=jpeg_p)\n\n # ----------------------- The second degradation process ----------------------- #\n # blur\n if np.random.uniform() < self.opt['second_blur_prob']:\n out = filter2D(out, self.kernel2)\n # random resize\n updown_type = random.choices(['up', 'down', 'keep'], self.opt['resize_prob2'])[0]\n if updown_type == 'up':\n scale = np.random.uniform(1, self.opt['resize_range2'][1])\n elif updown_type == 'down':\n scale = np.random.uniform(self.opt['resize_range2'][0], 1)\n else:\n scale = 1\n mode = random.choice(['area', 'bilinear', 'bicubic'])\n out = F.interpolate(\n out, size=(int(ori_h / self.opt['scale'] * scale), int(ori_w / self.opt['scale'] * scale)), mode=mode)\n # add noise\n gray_noise_prob = self.opt['gray_noise_prob2']\n if np.random.uniform() < self.opt['gaussian_noise_prob2']:\n out = random_add_gaussian_noise_pt(\n out, sigma_range=self.opt['noise_range2'], clip=True, rounds=False, gray_prob=gray_noise_prob)\n else:\n out = random_add_poisson_noise_pt(\n out,\n scale_range=self.opt['poisson_scale_range2'],\n gray_prob=gray_noise_prob,\n clip=True,\n rounds=False)\n\n # JPEG compression + the final sinc filter\n # We also need to resize images to desired sizes. We group [resize back + sinc filter] together\n # as one operation.\n # We consider two orders:\n # 1. [resize back + sinc filter] + JPEG compression\n # 2. JPEG compression + [resize back + sinc filter]\n # Empirically, we find other combinations (sinc + JPEG + Resize) will introduce twisted lines.\n if np.random.uniform() < 0.5:\n # resize back + the final sinc filter\n mode = random.choice(['area', 'bilinear', 'bicubic'])\n out = F.interpolate(out, size=(ori_h // self.opt['scale'], ori_w // self.opt['scale']), mode=mode)\n out = filter2D(out, self.sinc_kernel)\n # JPEG compression\n jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range2'])\n out = torch.clamp(out, 0, 1)\n out = self.jpeger(out, quality=jpeg_p)\n else:\n # JPEG compression\n jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range2'])\n out = torch.clamp(out, 0, 1)\n out = self.jpeger(out, quality=jpeg_p)\n # resize back + the final sinc filter\n mode = random.choice(['area', 'bilinear', 'bicubic'])\n out = F.interpolate(out, size=(ori_h // self.opt['scale'], ori_w // self.opt['scale']), mode=mode)\n out = filter2D(out, self.sinc_kernel)\n\n # clamp and round\n self.lq = torch.clamp((out * 255.0).round(), 0, 255) / 255.\n\n # random crop\n gt_size = self.opt['gt_size']\n (self.gt, self.gt_usm), self.lq = paired_random_crop([self.gt, self.gt_usm], self.lq, gt_size,\n self.opt['scale'])\n\n # training pair pool\n self._dequeue_and_enqueue()\n # sharpen self.gt again, as we have changed the self.gt with self._dequeue_and_enqueue\n self.gt_usm = self.usm_sharpener(self.gt)\n self.lq = self.lq.contiguous() # for the warning: grad and param do not obey the gradient layout contract\n else:\n # for paired training or validation\n self.lq = data['lq'].to(self.device)\n if 'gt' in data:\n self.gt = data['gt'].to(self.device)\n self.gt_usm = self.usm_sharpener(self.gt)\n\n def nondist_validation(self, dataloader, current_iter, tb_logger, save_img):\n # do not use the synthetic process during validation\n self.is_train = False\n super(KDSRGANSTModel, self).nondist_validation(dataloader, current_iter, tb_logger, save_img)\n self.is_train = True\n \n def test(self):\n if hasattr(self, 'net_g_ema'):\n self.net_g_ema.eval()\n with torch.no_grad():\n self.output = self.net_g_ema(self.lq)\n else:\n self.net_g.eval()\n with torch.no_grad():\n self.output = self.net_g(self.lq)\n self.net_g.train()\n\n def optimize_parameters(self, current_iter):\n lr = self.lr_sr * (self.gamma_sr ** ((current_iter ) // self.lr_decay_sr))\n for param_group in self.optimizer_g.param_groups:\n param_group['lr'] = lr \n\n l1_gt = self.gt_usm\n percep_gt = self.gt_usm\n gan_gt = self.gt_usm\n if self.opt['l1_gt_usm'] is False:\n l1_gt = self.gt\n if self.opt['percep_gt_usm'] is False:\n percep_gt = self.gt\n if self.opt['gan_gt_usm'] is False:\n gan_gt = self.gt\n \n hr2 = self.pixel_unshuffle(l1_gt)\n _, T_fea = self.model_Eta(torch.cat([self.lq,hr2],dim=1))\n\n # optimize net_g\n for p in self.net_d.parameters():\n p.requires_grad = False\n\n self.optimizer_g.zero_grad()\n self.output,S_fea = self.net_g(self.lq)\n\n l_g_total = 0\n loss_dict = OrderedDict()\n if (current_iter % self.net_d_iters == 0 and current_iter > self.net_d_init_iters):\n # pixel loss\n if self.cri_pix:\n l_g_pix = self.cri_pix(self.output, self.gt)\n l_g_total += l_g_pix\n loss_dict['l_g_pix'] = l_g_pix\n \n if self.cri_kd:\n l_kd = self.cri_kd(T_fea, S_fea)\n l_g_total += l_kd\n loss_dict['l_kd'] = l_kd\n\n # perceptual loss\n if self.cri_perceptual:\n l_g_percep, l_g_style = self.cri_perceptual(self.output, self.gt)\n if l_g_percep is not None:\n l_g_total += l_g_percep\n loss_dict['l_g_percep'] = l_g_percep\n if l_g_style is not None:\n l_g_total += l_g_style\n loss_dict['l_g_style'] = l_g_style\n # gan loss\n fake_g_pred = self.net_d(self.output)\n l_g_gan = self.cri_gan(fake_g_pred, True, is_disc=False)\n l_g_total += l_g_gan\n loss_dict['l_g_gan'] = l_g_gan\n\n l_g_total.backward()\n self.optimizer_g.step()\n\n # optimize net_d\n for p in self.net_d.parameters():\n p.requires_grad = True\n\n self.optimizer_d.zero_grad()\n # real\n real_d_pred = self.net_d(self.gt)\n l_d_real = self.cri_gan(real_d_pred, True, is_disc=True)\n loss_dict['l_d_real'] = l_d_real\n loss_dict['out_d_real'] = torch.mean(real_d_pred.detach())\n l_d_real.backward()\n # fake\n fake_d_pred = self.net_d(self.output.detach().clone()) # clone for pt1.9\n l_d_fake = self.cri_gan(fake_d_pred, False, is_disc=True)\n loss_dict['l_d_fake'] = l_d_fake\n loss_dict['out_d_fake'] = torch.mean(fake_d_pred.detach())\n l_d_fake.backward()\n self.optimizer_d.step()\n\n if self.ema_decay > 0:\n self.model_ema(decay=self.ema_decay)\n\n self.log_dict = self.reduce_loss_dict(loss_dict)\n"
},
{
"path": "KDSR-GAN/kdsrgan/models/kdsrgan_TA_model.py",
"content": "import numpy as np\nimport random\nimport torch\nfrom basicsr.data.degradations import random_add_gaussian_noise_pt, random_add_poisson_noise_pt\nfrom basicsr.data.transforms import paired_random_crop\nfrom basicsr.models.srgan_model import SRGANModel\nfrom basicsr.utils import DiffJPEG, USMSharp\nfrom basicsr.utils.img_process_util import filter2D\nfrom basicsr.utils.registry import MODEL_REGISTRY\nfrom collections import OrderedDict\nfrom torch.nn import functional as F\n\n\n@MODEL_REGISTRY.register()\nclass KDSRGANTAModel(SRGANModel):\n \"\"\"\n\n It mainly performs:\n 1. randomly synthesize LQ images in GPU tensors\n 2. optimize the networks with GAN training.\n \"\"\"\n\n def __init__(self, opt):\n super(KDSRGANTAModel, self).__init__(opt)\n self.jpeger = DiffJPEG(differentiable=False).cuda() # simulate JPEG compression artifacts\n self.usm_sharpener = USMSharp().cuda() # do usm sharpening\n self.queue_size = opt.get('queue_size', 180)\n\n @torch.no_grad()\n def _dequeue_and_enqueue(self):\n \"\"\"It is the training pair pool for increasing the diversity in a batch.\n\n Batch processing limits the diversity of synthetic degradations in a batch. For example, samples in a\n batch could not have different resize scaling factors. Therefore, we employ this training pair pool\n to increase the degradation diversity in a batch.\n \"\"\"\n # initialize\n b, c, h, w = self.lq.size()\n if not hasattr(self, 'queue_lr'):\n assert self.queue_size % b == 0, f'queue size {self.queue_size} should be divisible by batch size {b}'\n self.queue_lr = torch.zeros(self.queue_size, c, h, w).cuda()\n _, c, h, w = self.gt.size()\n self.queue_gt = torch.zeros(self.queue_size, c, h, w).cuda()\n self.queue_ptr = 0\n if self.queue_ptr == self.queue_size: # the pool is full\n # do dequeue and enqueue\n # shuffle\n idx = torch.randperm(self.queue_size)\n self.queue_lr = self.queue_lr[idx]\n self.queue_gt = self.queue_gt[idx]\n # get first b samples\n lq_dequeue = self.queue_lr[0:b, :, :, :].clone()\n gt_dequeue = self.queue_gt[0:b, :, :, :].clone()\n # update the queue\n self.queue_lr[0:b, :, :, :] = self.lq.clone()\n self.queue_gt[0:b, :, :, :] = self.gt.clone()\n\n self.lq = lq_dequeue\n self.gt = gt_dequeue\n else:\n # only do enqueue\n self.queue_lr[self.queue_ptr:self.queue_ptr + b, :, :, :] = self.lq.clone()\n self.queue_gt[self.queue_ptr:self.queue_ptr + b, :, :, :] = self.gt.clone()\n self.queue_ptr = self.queue_ptr + b\n\n @torch.no_grad()\n def feed_data(self, data):\n \"\"\"Accept data from dataloader, and then add two-order degradations to obtain LQ images.\n \"\"\"\n if self.is_train and self.opt.get('high_order_degradation', True):\n # training data synthesis\n self.gt = data['gt'].to(self.device)\n self.gt_usm = self.usm_sharpener(self.gt)\n\n self.kernel1 = data['kernel1'].to(self.device)\n self.kernel2 = data['kernel2'].to(self.device)\n self.sinc_kernel = data['sinc_kernel'].to(self.device)\n\n ori_h, ori_w = self.gt.size()[2:4]\n\n # ----------------------- The first degradation process ----------------------- #\n # blur\n out = filter2D(self.gt_usm, self.kernel1)\n # random resize\n updown_type = random.choices(['up', 'down', 'keep'], self.opt['resize_prob'])[0]\n if updown_type == 'up':\n scale = np.random.uniform(1, self.opt['resize_range'][1])\n elif updown_type == 'down':\n scale = np.random.uniform(self.opt['resize_range'][0], 1)\n else:\n scale = 1\n mode = random.choice(['area', 'bilinear', 'bicubic'])\n out = F.interpolate(out, scale_factor=scale, mode=mode)\n # add noise\n gray_noise_prob = self.opt['gray_noise_prob']\n if np.random.uniform() < self.opt['gaussian_noise_prob']:\n out = random_add_gaussian_noise_pt(\n out, sigma_range=self.opt['noise_range'], clip=True, rounds=False, gray_prob=gray_noise_prob)\n else:\n out = random_add_poisson_noise_pt(\n out,\n scale_range=self.opt['poisson_scale_range'],\n gray_prob=gray_noise_prob,\n clip=True,\n rounds=False)\n # JPEG compression\n jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range'])\n out = torch.clamp(out, 0, 1) # clamp to [0, 1], otherwise JPEGer will result in unpleasant artifacts\n out = self.jpeger(out, quality=jpeg_p)\n\n # ----------------------- The second degradation process ----------------------- #\n # blur\n if np.random.uniform() < self.opt['second_blur_prob']:\n out = filter2D(out, self.kernel2)\n # random resize\n updown_type = random.choices(['up', 'down', 'keep'], self.opt['resize_prob2'])[0]\n if updown_type == 'up':\n scale = np.random.uniform(1, self.opt['resize_range2'][1])\n elif updown_type == 'down':\n scale = np.random.uniform(self.opt['resize_range2'][0], 1)\n else:\n scale = 1\n mode = random.choice(['area', 'bilinear', 'bicubic'])\n out = F.interpolate(\n out, size=(int(ori_h / self.opt['scale'] * scale), int(ori_w / self.opt['scale'] * scale)), mode=mode)\n # add noise\n gray_noise_prob = self.opt['gray_noise_prob2']\n if np.random.uniform() < self.opt['gaussian_noise_prob2']:\n out = random_add_gaussian_noise_pt(\n out, sigma_range=self.opt['noise_range2'], clip=True, rounds=False, gray_prob=gray_noise_prob)\n else:\n out = random_add_poisson_noise_pt(\n out,\n scale_range=self.opt['poisson_scale_range2'],\n gray_prob=gray_noise_prob,\n clip=True,\n rounds=False)\n\n # JPEG compression + the final sinc filter\n # We also need to resize images to desired sizes. We group [resize back + sinc filter] together\n # as one operation.\n # We consider two orders:\n # 1. [resize back + sinc filter] + JPEG compression\n # 2. JPEG compression + [resize back + sinc filter]\n # Empirically, we find other combinations (sinc + JPEG + Resize) will introduce twisted lines.\n if np.random.uniform() < 0.5:\n # resize back + the final sinc filter\n mode = random.choice(['area', 'bilinear', 'bicubic'])\n out = F.interpolate(out, size=(ori_h // self.opt['scale'], ori_w // self.opt['scale']), mode=mode)\n out = filter2D(out, self.sinc_kernel)\n # JPEG compression\n jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range2'])\n out = torch.clamp(out, 0, 1)\n out = self.jpeger(out, quality=jpeg_p)\n else:\n # JPEG compression\n jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range2'])\n out = torch.clamp(out, 0, 1)\n out = self.jpeger(out, quality=jpeg_p)\n # resize back + the final sinc filter\n mode = random.choice(['area', 'bilinear', 'bicubic'])\n out = F.interpolate(out, size=(ori_h // self.opt['scale'], ori_w // self.opt['scale']), mode=mode)\n out = filter2D(out, self.sinc_kernel)\n\n # clamp and round\n self.lq = torch.clamp((out * 255.0).round(), 0, 255) / 255.\n\n # random crop\n gt_size = self.opt['gt_size']\n (self.gt, self.gt_usm), self.lq = paired_random_crop([self.gt, self.gt_usm], self.lq, gt_size,\n self.opt['scale'])\n\n # training pair pool\n self._dequeue_and_enqueue()\n # sharpen self.gt again, as we have changed the self.gt with self._dequeue_and_enqueue\n self.gt_usm = self.usm_sharpener(self.gt)\n self.lq = self.lq.contiguous() # for the warning: grad and param do not obey the gradient layout contract\n else:\n # for paired training or validation\n self.lq = data['lq'].to(self.device)\n if 'gt' in data:\n self.gt = data['gt'].to(self.device)\n self.gt_usm = self.usm_sharpener(self.gt)\n\n def nondist_validation(self, dataloader, current_iter, tb_logger, save_img):\n # do not use the synthetic process during validation\n self.is_train = False\n super(KDSRGANTAModel, self).nondist_validation(dataloader, current_iter, tb_logger, save_img)\n self.is_train = True\n\n def optimize_parameters(self, current_iter):\n # optimize net_g\n for p in self.net_d.parameters():\n p.requires_grad = False\n\n self.optimizer_g.zero_grad()\n self.output = self.net_g(self.lq)\n\n l_g_total = 0\n loss_dict = OrderedDict()\n if (current_iter % self.net_d_iters == 0 and current_iter > self.net_d_init_iters):\n # pixel loss\n if self.cri_pix:\n l_g_pix = self.cri_pix(self.output, self.gt)\n l_g_total += l_g_pix\n loss_dict['l_g_pix'] = l_g_pix\n # perceptual loss\n if self.cri_perceptual:\n l_g_percep, l_g_style = self.cri_perceptual(self.output, self.gt)\n if l_g_percep is not None:\n l_g_total += l_g_percep\n loss_dict['l_g_percep'] = l_g_percep\n if l_g_style is not None:\n l_g_total += l_g_style\n loss_dict['l_g_style'] = l_g_style\n # gan loss\n fake_g_pred = self.net_d(self.output)\n l_g_gan = self.cri_gan(fake_g_pred, True, is_disc=False)\n l_g_total += l_g_gan\n loss_dict['l_g_gan'] = l_g_gan\n\n l_g_total.backward()\n self.optimizer_g.step()\n\n # optimize net_d\n for p in self.net_d.parameters():\n p.requires_grad = True\n\n self.optimizer_d.zero_grad()\n # real\n real_d_pred = self.net_d(self.gt)\n l_d_real = self.cri_gan(real_d_pred, True, is_disc=True)\n loss_dict['l_d_real'] = l_d_real\n loss_dict['out_d_real'] = torch.mean(real_d_pred.detach())\n l_d_real.backward()\n # fake\n fake_d_pred = self.net_d(self.output.detach().clone()) # clone for pt1.9\n l_d_fake = self.cri_gan(fake_d_pred, False, is_disc=True)\n loss_dict['l_d_fake'] = l_d_fake\n loss_dict['out_d_fake'] = torch.mean(fake_d_pred.detach())\n l_d_fake.backward()\n self.optimizer_d.step()\n\n if self.ema_decay > 0:\n self.model_ema(decay=self.ema_decay)\n\n self.log_dict = self.reduce_loss_dict(loss_dict)\n"
},
{
"path": "KDSR-GAN/kdsrgan/models/kdsrnet_ST_model.py",
"content": "import numpy as np\nimport random\nimport torch\nfrom basicsr.data.degradations import random_add_gaussian_noise_pt, random_add_poisson_noise_pt\nfrom basicsr.data.transforms import paired_random_crop\nfrom basicsr.models.sr_model import SRModel\nfrom basicsr.utils import DiffJPEG, USMSharp\nfrom basicsr.utils.img_process_util import filter2D\nfrom basicsr.utils.registry import MODEL_REGISTRY\nfrom torch.nn import functional as F\nfrom collections import OrderedDict\nfrom basicsr.archs import build_network\nfrom basicsr.utils import get_root_logger\nfrom basicsr.losses import build_loss\nfrom torch import nn\nfrom basicsr.models import lr_scheduler as lr_scheduler\n\n@MODEL_REGISTRY.register()\nclass KDSRNetSTModel(SRModel):\n \"\"\"\n It is trained without GAN losses.\n It mainly performs:\n 1. randomly synthesize LQ images in GPU tensors\n 2. optimize the networks with GAN training.\n \"\"\"\n\n def __init__(self, opt):\n super(KDSRNetSTModel, self).__init__(opt)\n self.jpeger = DiffJPEG(differentiable=False).cuda() # simulate JPEG compression artifacts\n self.usm_sharpener = USMSharp().cuda() # do usm sharpening\n self.queue_size = opt.get('queue_size', 180)\n\n self.net_g_TA = build_network(opt['network_TA'])\n self.net_g_TA = self.model_to_device(self.net_g_TA)\n\n # load pretrained models\n load_path = self.opt['path'].get('pretrain_network_TA', None)\n if load_path is not None:\n param_key = self.opt['path'].get('param_key_g', 'params')\n self.load_network(self.net_g_TA, load_path, True, param_key)\n \n self.net_g_TA.eval()\n if self.opt['dist']:\n self.model_Est = self.net_g.module.E_st\n self.model_Eta = self.net_g_TA.module.E\n else:\n self.model_Est = self.net_g.E_st\n self.model_Eta = self.net_g_TA.E\n self.pixel_unshuffle = nn.PixelUnshuffle(opt[\"scale\"])\n self.encoder_iter = opt[\"train\"][\"encoder_iter\"]\n self.lr_encoder = opt[\"train\"][\"lr_encoder\"]\n self.lr_sr = opt[\"train\"][\"lr_sr\"]\n self.gamma_encoder = opt[\"train\"][\"gamma_encoder\"]\n self.gamma_sr = opt[\"train\"][\"gamma_sr\"]\n self.lr_decay_encoder = opt[\"train\"][\"lr_decay_encoder\"]\n self.lr_decay_sr = opt[\"train\"][\"lr_decay_sr\"]\n\n\n def setup_optimizers(self):\n train_opt = self.opt['train']\n optim_params = []\n for k, v in self.net_g.named_parameters():\n if v.requires_grad:\n optim_params.append(v)\n else:\n logger = get_root_logger()\n logger.warning(f'Params {k} will not be optimized in the second stage.')\n\n optim_type = train_opt['optim_g'].pop('type')\n self.optimizer_g = self.get_optimizer(optim_type, optim_params, **train_opt['optim_g'])\n self.optimizers.append(self.optimizer_g)\n \n def setup_schedulers(self):\n \"\"\"Set up schedulers.\"\"\"\n train_opt = self.opt['train']\n\n scheduler_type = train_opt['scheduler'].pop('type')\n if scheduler_type in ['MultiStepLR', 'MultiStepRestartLR']:\n self.schedulers.append(lr_scheduler.MultiStepRestartLR(self.optimizer_g, **train_opt['scheduler']))\n elif scheduler_type == 'CosineAnnealingRestartLR':\n self.schedulers.append(lr_scheduler.CosineAnnealingRestartLR(self.optimizer_g, **train_opt['scheduler']))\n else:\n raise NotImplementedError(f'Scheduler {scheduler_type} is not implemented yet.')\n\n\n\n def init_training_settings(self):\n self.net_g.train()\n train_opt = self.opt['train']\n\n self.ema_decay = train_opt.get('ema_decay', 0)\n if self.ema_decay > 0:\n logger = get_root_logger()\n logger.info(f'Use Exponential Moving Average with decay: {self.ema_decay}')\n # define network net_g with Exponential Moving Average (EMA)\n # net_g_ema is used only for testing on one GPU and saving\n # There is no need to wrap with DistributedDataParallel\n self.net_g_ema = build_network(self.opt['network_g']).to(self.device)\n # load pretrained model\n load_path = self.opt['path'].get('pretrain_network_g', None)\n if load_path is not None:\n self.load_network(self.net_g_ema, load_path, self.opt['path'].get('strict_load_g', True), 'params_ema')\n else:\n self.model_ema(0) # copy net_g weight\n self.net_g_ema.eval()\n\n # define losses\n if train_opt.get('pixel_opt'):\n self.cri_pix = build_loss(train_opt['pixel_opt']).to(self.device)\n else:\n self.cri_pix = None\n\n if train_opt.get('perceptual_opt'):\n self.cri_perceptual = build_loss(train_opt['perceptual_opt']).to(self.device)\n else:\n self.cri_perceptual = None\n\n if train_opt.get('kd_opt'):\n self.cri_kd = build_loss(train_opt['kd_opt']).to(self.device)\n else:\n self.cri_kd = None\n\n if self.cri_pix is None and self.cri_perceptual is None:\n raise ValueError('Both pixel and perceptual losses are None.')\n\n # set up optimizers and schedulers\n self.setup_optimizers()\n self.setup_schedulers()\n\n @torch.no_grad()\n def _dequeue_and_enqueue(self):\n \"\"\"It is the training pair pool for increasing the diversity in a batch.\n\n Batch processing limits the diversity of synthetic degradations in a batch. For example, samples in a\n batch could not have different resize scaling factors. Therefore, we employ this training pair pool\n to increase the degradation diversity in a batch.\n \"\"\"\n # initialize\n b, c, h, w = self.lq.size()\n if not hasattr(self, 'queue_lr'):\n assert self.queue_size % b == 0, f'queue size {self.queue_size} should be divisible by batch size {b}'\n self.queue_lr = torch.zeros(self.queue_size, c, h, w).cuda()\n _, c, h, w = self.gt.size()\n self.queue_gt = torch.zeros(self.queue_size, c, h, w).cuda()\n self.queue_ptr = 0\n if self.queue_ptr == self.queue_size: # the pool is full\n # do dequeue and enqueue\n # shuffle\n idx = torch.randperm(self.queue_size)\n self.queue_lr = self.queue_lr[idx]\n self.queue_gt = self.queue_gt[idx]\n # get first b samples\n lq_dequeue = self.queue_lr[0:b, :, :, :].clone()\n gt_dequeue = self.queue_gt[0:b, :, :, :].clone()\n # update the queue\n self.queue_lr[0:b, :, :, :] = self.lq.clone()\n self.queue_gt[0:b, :, :, :] = self.gt.clone()\n\n self.lq = lq_dequeue\n self.gt = gt_dequeue\n else:\n # only do enqueue\n self.queue_lr[self.queue_ptr:self.queue_ptr + b, :, :, :] = self.lq.clone()\n self.queue_gt[self.queue_ptr:self.queue_ptr + b, :, :, :] = self.gt.clone()\n self.queue_ptr = self.queue_ptr + b\n\n @torch.no_grad()\n def feed_data(self, data):\n \"\"\"Accept data from dataloader, and then add two-order degradations to obtain LQ images.\n \"\"\"\n if self.is_train and self.opt.get('high_order_degradation', True):\n # training data synthesis\n self.gt = data['gt'].to(self.device)\n # USM sharpen the GT images\n if self.opt['gt_usm'] is True:\n self.gt = self.usm_sharpener(self.gt)\n\n self.kernel1 = data['kernel1'].to(self.device)\n self.kernel2 = data['kernel2'].to(self.device)\n self.sinc_kernel = data['sinc_kernel'].to(self.device)\n\n ori_h, ori_w = self.gt.size()[2:4]\n\n # ----------------------- The first degradation process ----------------------- #\n # blur\n out = filter2D(self.gt, self.kernel1)\n # random resize\n updown_type = random.choices(['up', 'down', 'keep'], self.opt['resize_prob'])[0]\n if updown_type == 'up':\n scale = np.random.uniform(1, self.opt['resize_range'][1])\n elif updown_type == 'down':\n scale = np.random.uniform(self.opt['resize_range'][0], 1)\n else:\n scale = 1\n mode = random.choice(['area', 'bilinear', 'bicubic'])\n out = F.interpolate(out, scale_factor=scale, mode=mode)\n # add noise\n gray_noise_prob = self.opt['gray_noise_prob']\n if np.random.uniform() < self.opt['gaussian_noise_prob']:\n out = random_add_gaussian_noise_pt(\n out, sigma_range=self.opt['noise_range'], clip=True, rounds=False, gray_prob=gray_noise_prob)\n else:\n out = random_add_poisson_noise_pt(\n out,\n scale_range=self.opt['poisson_scale_range'],\n gray_prob=gray_noise_prob,\n clip=True,\n rounds=False)\n # JPEG compression\n jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range'])\n out = torch.clamp(out, 0, 1) # clamp to [0, 1], otherwise JPEGer will result in unpleasant artifacts\n out = self.jpeger(out, quality=jpeg_p)\n\n # ----------------------- The second degradation process ----------------------- #\n # blur\n if np.random.uniform() < self.opt['second_blur_prob']:\n out = filter2D(out, self.kernel2)\n # random resize\n updown_type = random.choices(['up', 'down', 'keep'], self.opt['resize_prob2'])[0]\n if updown_type == 'up':\n scale = np.random.uniform(1, self.opt['resize_range2'][1])\n elif updown_type == 'down':\n scale = np.random.uniform(self.opt['resize_range2'][0], 1)\n else:\n scale = 1\n mode = random.choice(['area', 'bilinear', 'bicubic'])\n out = F.interpolate(\n out, size=(int(ori_h / self.opt['scale'] * scale), int(ori_w / self.opt['scale'] * scale)), mode=mode)\n # add noise\n gray_noise_prob = self.opt['gray_noise_prob2']\n if np.random.uniform() < self.opt['gaussian_noise_prob2']:\n out = random_add_gaussian_noise_pt(\n out, sigma_range=self.opt['noise_range2'], clip=True, rounds=False, gray_prob=gray_noise_prob)\n else:\n out = random_add_poisson_noise_pt(\n out,\n scale_range=self.opt['poisson_scale_range2'],\n gray_prob=gray_noise_prob,\n clip=True,\n rounds=False)\n\n # JPEG compression + the final sinc filter\n # We also need to resize images to desired sizes. We group [resize back + sinc filter] together\n # as one operation.\n # We consider two orders:\n # 1. [resize back + sinc filter] + JPEG compression\n # 2. JPEG compression + [resize back + sinc filter]\n # Empirically, we find other combinations (sinc + JPEG + Resize) will introduce twisted lines.\n if np.random.uniform() < 0.5:\n # resize back + the final sinc filter\n mode = random.choice(['area', 'bilinear', 'bicubic'])\n out = F.interpolate(out, size=(ori_h // self.opt['scale'], ori_w // self.opt['scale']), mode=mode)\n out = filter2D(out, self.sinc_kernel)\n # JPEG compression\n jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range2'])\n out = torch.clamp(out, 0, 1)\n out = self.jpeger(out, quality=jpeg_p)\n else:\n # JPEG compression\n jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range2'])\n out = torch.clamp(out, 0, 1)\n out = self.jpeger(out, quality=jpeg_p)\n # resize back + the final sinc filter\n mode = random.choice(['area', 'bilinear', 'bicubic'])\n out = F.interpolate(out, size=(ori_h // self.opt['scale'], ori_w // self.opt['scale']), mode=mode)\n out = filter2D(out, self.sinc_kernel)\n\n # clamp and round\n self.lq = torch.clamp((out * 255.0).round(), 0, 255) / 255.\n\n # random crop\n gt_size = self.opt['gt_size']\n self.gt, self.lq = paired_random_crop(self.gt, self.lq, gt_size, self.opt['scale'])\n\n # training pair pool\n self._dequeue_and_enqueue()\n self.lq = self.lq.contiguous() # for the warning: grad and param do not obey the gradient layout contract\n else:\n # for paired training or validation\n self.lq = data['lq'].to(self.device)\n if 'gt' in data:\n self.gt = data['gt'].to(self.device)\n self.gt_usm = self.usm_sharpener(self.gt)\n\n def nondist_validation(self, dataloader, current_iter, tb_logger, save_img):\n # do not use the synthetic process during validation\n self.is_train = False\n super(KDSRNetSTModel, self).nondist_validation(dataloader, current_iter, tb_logger, save_img)\n self.is_train = True\n\n def test(self):\n if hasattr(self, 'net_g_ema'):\n self.net_g_ema.eval()\n with torch.no_grad():\n self.output = self.net_g_ema(self.lq)\n else:\n self.net_g.eval()\n with torch.no_grad():\n\n self.output = self.net_g(self.lq)\n self.net_g.train()\n\n def optimize_parameters(self, current_iter):\n lr = self.lr_sr * (self.gamma_sr ** ((current_iter ) // self.lr_decay_sr))\n for param_group in self.optimizer_g.param_groups:\n param_group['lr'] = lr \n \n l_total = 0\n loss_dict = OrderedDict()\n\n hr2 = self.pixel_unshuffle(self.gt)\n _, T_fea = self.model_Eta(torch.cat([self.lq,hr2],dim=1))\n\n\n self.optimizer_g.zero_grad()\n self.output, S_fea = self.net_g(self.lq)\n l_pix = self.cri_pix(self.output, self.gt)\n l_total += l_pix\n loss_dict['l_pix'] = l_pix\n l_kd = self.cri_kd(T_fea, S_fea)\n l_total += l_kd\n loss_dict['l_kd'] = l_kd\n l_total.backward()\n self.optimizer_g.step()\n\n self.log_dict = self.reduce_loss_dict(loss_dict)\n\n if self.ema_decay > 0:\n self.model_ema(decay=self.ema_decay)\n"
},
{
"path": "KDSR-GAN/kdsrgan/models/kdsrnet_TA_model.py",
"content": "import numpy as np\nimport random\nimport torch\nfrom basicsr.data.degradations import random_add_gaussian_noise_pt, random_add_poisson_noise_pt\nfrom basicsr.data.transforms import paired_random_crop\nfrom basicsr.models.sr_model import SRModel\nfrom basicsr.utils import DiffJPEG, USMSharp\nfrom basicsr.utils.img_process_util import filter2D\nfrom basicsr.utils.registry import MODEL_REGISTRY\nfrom torch.nn import functional as F\nfrom collections import OrderedDict\n\n\n@MODEL_REGISTRY.register()\nclass KDSRNetTAModel(SRModel):\n \"\"\"\n It is trained without GAN losses.\n It mainly performs:\n 1. randomly synthesize LQ images in GPU tensors\n 2. optimize the networks with GAN training.\n \"\"\"\n\n def __init__(self, opt):\n super(KDSRNetTAModel, self).__init__(opt)\n self.jpeger = DiffJPEG(differentiable=False).cuda() # simulate JPEG compression artifacts\n self.usm_sharpener = USMSharp().cuda() # do usm sharpening\n self.queue_size = opt.get('queue_size', 180)\n\n\n\n @torch.no_grad()\n def _dequeue_and_enqueue(self):\n \"\"\"It is the training pair pool for increasing the diversity in a batch.\n\n Batch processing limits the diversity of synthetic degradations in a batch. For example, samples in a\n batch could not have different resize scaling factors. Therefore, we employ this training pair pool\n to increase the degradation diversity in a batch.\n \"\"\"\n # initialize\n b, c, h, w = self.lq.size()\n if not hasattr(self, 'queue_lr'):\n assert self.queue_size % b == 0, f'queue size {self.queue_size} should be divisible by batch size {b}'\n self.queue_lr = torch.zeros(self.queue_size, c, h, w).cuda()\n _, c, h, w = self.gt.size()\n self.queue_gt = torch.zeros(self.queue_size, c, h, w).cuda()\n self.queue_ptr = 0\n if self.queue_ptr == self.queue_size: # the pool is full\n # do dequeue and enqueue\n # shuffle\n idx = torch.randperm(self.queue_size)\n self.queue_lr = self.queue_lr[idx]\n self.queue_gt = self.queue_gt[idx]\n # get first b samples\n lq_dequeue = self.queue_lr[0:b, :, :, :].clone()\n gt_dequeue = self.queue_gt[0:b, :, :, :].clone()\n # update the queue\n self.queue_lr[0:b, :, :, :] = self.lq.clone()\n self.queue_gt[0:b, :, :, :] = self.gt.clone()\n\n self.lq = lq_dequeue\n self.gt = gt_dequeue\n else:\n # only do enqueue\n self.queue_lr[self.queue_ptr:self.queue_ptr + b, :, :, :] = self.lq.clone()\n self.queue_gt[self.queue_ptr:self.queue_ptr + b, :, :, :] = self.gt.clone()\n self.queue_ptr = self.queue_ptr + b\n\n @torch.no_grad()\n def feed_data(self, data):\n \"\"\"Accept data from dataloader, and then add two-order degradations to obtain LQ images.\n \"\"\"\n if self.is_train and self.opt.get('high_order_degradation', True):\n # training data synthesis\n self.gt = data['gt'].to(self.device)\n # USM sharpen the GT images\n if self.opt['gt_usm'] is True:\n self.gt = self.usm_sharpener(self.gt)\n\n self.kernel1 = data['kernel1'].to(self.device)\n self.kernel2 = data['kernel2'].to(self.device)\n self.sinc_kernel = data['sinc_kernel'].to(self.device)\n\n ori_h, ori_w = self.gt.size()[2:4]\n\n # ----------------------- The first degradation process ----------------------- #\n # blur\n out = filter2D(self.gt, self.kernel1)\n # random resize\n updown_type = random.choices(['up', 'down', 'keep'], self.opt['resize_prob'])[0]\n if updown_type == 'up':\n scale = np.random.uniform(1, self.opt['resize_range'][1])\n elif updown_type == 'down':\n scale = np.random.uniform(self.opt['resize_range'][0], 1)\n else:\n scale = 1\n mode = random.choice(['area', 'bilinear', 'bicubic'])\n out = F.interpolate(out, scale_factor=scale, mode=mode)\n # add noise\n gray_noise_prob = self.opt['gray_noise_prob']\n if np.random.uniform() < self.opt['gaussian_noise_prob']:\n out = random_add_gaussian_noise_pt(\n out, sigma_range=self.opt['noise_range'], clip=True, rounds=False, gray_prob=gray_noise_prob)\n else:\n out = random_add_poisson_noise_pt(\n out,\n scale_range=self.opt['poisson_scale_range'],\n gray_prob=gray_noise_prob,\n clip=True,\n rounds=False)\n # JPEG compression\n jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range'])\n out = torch.clamp(out, 0, 1) # clamp to [0, 1], otherwise JPEGer will result in unpleasant artifacts\n out = self.jpeger(out, quality=jpeg_p)\n\n # ----------------------- The second degradation process ----------------------- #\n # blur\n if np.random.uniform() < self.opt['second_blur_prob']:\n out = filter2D(out, self.kernel2)\n # random resize\n updown_type = random.choices(['up', 'down', 'keep'], self.opt['resize_prob2'])[0]\n if updown_type == 'up':\n scale = np.random.uniform(1, self.opt['resize_range2'][1])\n elif updown_type == 'down':\n scale = np.random.uniform(self.opt['resize_range2'][0], 1)\n else:\n scale = 1\n mode = random.choice(['area', 'bilinear', 'bicubic'])\n out = F.interpolate(\n out, size=(int(ori_h / self.opt['scale'] * scale), int(ori_w / self.opt['scale'] * scale)), mode=mode)\n # add noise\n gray_noise_prob = self.opt['gray_noise_prob2']\n if np.random.uniform() < self.opt['gaussian_noise_prob2']:\n out = random_add_gaussian_noise_pt(\n out, sigma_range=self.opt['noise_range2'], clip=True, rounds=False, gray_prob=gray_noise_prob)\n else:\n out = random_add_poisson_noise_pt(\n out,\n scale_range=self.opt['poisson_scale_range2'],\n gray_prob=gray_noise_prob,\n clip=True,\n rounds=False)\n\n # JPEG compression + the final sinc filter\n # We also need to resize images to desired sizes. We group [resize back + sinc filter] together\n # as one operation.\n # We consider two orders:\n # 1. [resize back + sinc filter] + JPEG compression\n # 2. JPEG compression + [resize back + sinc filter]\n # Empirically, we find other combinations (sinc + JPEG + Resize) will introduce twisted lines.\n if np.random.uniform() < 0.5:\n # resize back + the final sinc filter\n mode = random.choice(['area', 'bilinear', 'bicubic'])\n out = F.interpolate(out, size=(ori_h // self.opt['scale'], ori_w // self.opt['scale']), mode=mode)\n out = filter2D(out, self.sinc_kernel)\n # JPEG compression\n jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range2'])\n out = torch.clamp(out, 0, 1)\n out = self.jpeger(out, quality=jpeg_p)\n else:\n # JPEG compression\n jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range2'])\n out = torch.clamp(out, 0, 1)\n out = self.jpeger(out, quality=jpeg_p)\n # resize back + the final sinc filter\n mode = random.choice(['area', 'bilinear', 'bicubic'])\n out = F.interpolate(out, size=(ori_h // self.opt['scale'], ori_w // self.opt['scale']), mode=mode)\n out = filter2D(out, self.sinc_kernel)\n\n # clamp and round\n self.lq = torch.clamp((out * 255.0).round(), 0, 255) / 255.\n\n # random crop\n gt_size = self.opt['gt_size']\n self.gt, self.lq = paired_random_crop(self.gt, self.lq, gt_size, self.opt['scale'])\n\n # training pair pool\n self._dequeue_and_enqueue()\n self.lq = self.lq.contiguous() # for the warning: grad and param do not obey the gradient layout contract\n else:\n # for paired training or validation\n self.lq = data['lq'].to(self.device)\n if 'gt' in data:\n self.gt = data['gt'].to(self.device)\n self.gt_usm = self.usm_sharpener(self.gt)\n # self.gt = self.usm_sharpener(self.gt)\n\n def nondist_validation(self, dataloader, current_iter, tb_logger, save_img):\n # do not use the synthetic process during validation\n self.is_train = False\n super(KDSRNetTAModel, self).nondist_validation(dataloader, current_iter, tb_logger, save_img)\n self.is_train = True\n\n def test(self):\n if hasattr(self, 'net_g_ema'):\n self.net_g_ema.eval()\n with torch.no_grad():\n self.output = self.net_g_ema(self.lq, self.gt)\n else:\n self.net_g.eval()\n with torch.no_grad():\n # self.output = self.net_g(self.lq, self.gt)\n self.output = self.net_g(self.lq, self.gt)\n self.net_g.train()\n\n def optimize_parameters(self, current_iter):\n self.optimizer_g.zero_grad()\n self.output, _ = self.net_g(self.lq, self.gt)\n\n l_total = 0\n loss_dict = OrderedDict()\n # pixel loss\n if self.cri_pix:\n l_pix = self.cri_pix(self.output, self.gt)\n l_total += l_pix\n loss_dict['l_pix'] = l_pix\n # perceptual loss\n if self.cri_perceptual:\n l_percep, l_style = self.cri_perceptual(self.output, self.gt)\n if l_percep is not None:\n l_total += l_percep\n loss_dict['l_percep'] = l_percep\n if l_style is not None:\n l_total += l_style\n loss_dict['l_style'] = l_style\n\n l_total.backward()\n self.optimizer_g.step()\n\n self.log_dict = self.reduce_loss_dict(loss_dict)\n\n if self.ema_decay > 0:\n self.model_ema(decay=self.ema_decay)\n"
},
{
"path": "KDSR-GAN/kdsrgan/test.py",
"content": "# flake8: noqa\nimport os.path as osp\nfrom basicsr.test import test_pipeline\n\nimport kdsrgan.archs\nimport kdsrgan.data\nimport kdsrgan.models\n\nif __name__ == '__main__':\n root_path = osp.abspath(osp.join(__file__, osp.pardir, osp.pardir))\n test_pipeline(root_path)\n"
},
{
"path": "KDSR-GAN/kdsrgan/train.py",
"content": "# flake8: noqa\nimport os.path as osp\nfrom basicsr.train import train_pipeline\n\nimport kdsrgan.archs\nimport kdsrgan.data\nimport kdsrgan.models\nimport kdsrgan.losses\nimport warnings\n\nwarnings.filterwarnings(\"ignore\")\n\nif __name__ == '__main__':\n root_path = osp.abspath(osp.join(__file__, osp.pardir, osp.pardir))\n train_pipeline(root_path)\n"
},
{
"path": "KDSR-GAN/kdsrgan/utils.py",
"content": "import cv2\nimport math\nimport numpy as np\nimport os\nimport queue\nimport threading\nimport torch\nfrom basicsr.utils.download_util import load_file_from_url\nfrom torch.nn import functional as F\n\nROOT_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))\n\n\nclass RealESRGANer():\n \"\"\"A helper class for upsampling images with RealESRGAN.\n\n Args:\n scale (int): Upsampling scale factor used in the networks. It is usually 2 or 4.\n model_path (str): The path to the pretrained model. It can be urls (will first download it automatically).\n model (nn.Module): The defined network. Default: None.\n tile (int): As too large images result in the out of GPU memory issue, so this tile option will first crop\n input images into tiles, and then process each of them. Finally, they will be merged into one image.\n 0 denotes for do not use tile. Default: 0.\n tile_pad (int): The pad size for each tile, to remove border artifacts. Default: 10.\n pre_pad (int): Pad the input images to avoid border artifacts. Default: 10.\n half (float): Whether to use half precision during inference. Default: False.\n \"\"\"\n\n def __init__(self,\n scale,\n model_path,\n model=None,\n tile=0,\n tile_pad=10,\n pre_pad=10,\n half=False,\n device=None,\n gpu_id=None):\n self.scale = scale\n self.tile_size = tile\n self.tile_pad = tile_pad\n self.pre_pad = pre_pad\n self.mod_scale = None\n self.half = half\n\n # initialize model\n if gpu_id:\n self.device = torch.device(\n f'cuda:{gpu_id}' if torch.cuda.is_available() else 'cpu') if device is None else device\n else:\n self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') if device is None else device\n # if the model_path starts with https, it will first download models to the folder: realesrgan/weights\n if model_path.startswith('https://'):\n model_path = load_file_from_url(\n url=model_path, model_dir=os.path.join(ROOT_DIR, 'realesrgan/weights'), progress=True, file_name=None)\n loadnet = torch.load(model_path, map_location=torch.device('cpu'))\n # prefer to use params_ema\n if 'params_ema' in loadnet:\n keyname = 'params_ema'\n else:\n keyname = 'params'\n model.load_state_dict(loadnet[keyname], strict=True)\n model.eval()\n self.model = model.to(self.device)\n if self.half:\n self.model = self.model.half()\n\n def pre_process(self, img):\n \"\"\"Pre-process, such as pre-pad and mod pad, so that the images can be divisible\n \"\"\"\n img = torch.from_numpy(np.transpose(img, (2, 0, 1))).float()\n self.img = img.unsqueeze(0).to(self.device)\n if self.half:\n self.img = self.img.half()\n\n # pre_pad\n if self.pre_pad != 0:\n self.img = F.pad(self.img, (0, self.pre_pad, 0, self.pre_pad), 'reflect')\n # mod pad for divisible borders\n if self.scale == 2:\n self.mod_scale = 2\n elif self.scale == 1:\n self.mod_scale = 4\n if self.mod_scale is not None:\n self.mod_pad_h, self.mod_pad_w = 0, 0\n _, _, h, w = self.img.size()\n if (h % self.mod_scale != 0):\n self.mod_pad_h = (self.mod_scale - h % self.mod_scale)\n if (w % self.mod_scale != 0):\n self.mod_pad_w = (self.mod_scale - w % self.mod_scale)\n self.img = F.pad(self.img, (0, self.mod_pad_w, 0, self.mod_pad_h), 'reflect')\n\n def process(self):\n # model inference\n self.output = self.model(self.img)\n\n def tile_process(self):\n \"\"\"It will first crop input images to tiles, and then process each tile.\n Finally, all the processed tiles are merged into one images.\n\n Modified from: https://github.com/ata4/esrgan-launcher\n \"\"\"\n batch, channel, height, width = self.img.shape\n output_height = height * self.scale\n output_width = width * self.scale\n output_shape = (batch, channel, output_height, output_width)\n\n # start with black image\n self.output = self.img.new_zeros(output_shape)\n tiles_x = math.ceil(width / self.tile_size)\n tiles_y = math.ceil(height / self.tile_size)\n\n # loop over all tiles\n for y in range(tiles_y):\n for x in range(tiles_x):\n # extract tile from input image\n ofs_x = x * self.tile_size\n ofs_y = y * self.tile_size\n # input tile area on total image\n input_start_x = ofs_x\n input_end_x = min(ofs_x + self.tile_size, width)\n input_start_y = ofs_y\n input_end_y = min(ofs_y + self.tile_size, height)\n\n # input tile area on total image with padding\n input_start_x_pad = max(input_start_x - self.tile_pad, 0)\n input_end_x_pad = min(input_end_x + self.tile_pad, width)\n input_start_y_pad = max(input_start_y - self.tile_pad, 0)\n input_end_y_pad = min(input_end_y + self.tile_pad, height)\n\n # input tile dimensions\n input_tile_width = input_end_x - input_start_x\n input_tile_height = input_end_y - input_start_y\n tile_idx = y * tiles_x + x + 1\n input_tile = self.img[:, :, input_start_y_pad:input_end_y_pad, input_start_x_pad:input_end_x_pad]\n\n # upscale tile\n try:\n with torch.no_grad():\n output_tile = self.model(input_tile)\n except RuntimeError as error:\n print('Error', error)\n print(f'\\tTile {tile_idx}/{tiles_x * tiles_y}')\n\n # output tile area on total image\n output_start_x = input_start_x * self.scale\n output_end_x = input_end_x * self.scale\n output_start_y = input_start_y * self.scale\n output_end_y = input_end_y * self.scale\n\n # output tile area without padding\n output_start_x_tile = (input_start_x - input_start_x_pad) * self.scale\n output_end_x_tile = output_start_x_tile + input_tile_width * self.scale\n output_start_y_tile = (input_start_y - input_start_y_pad) * self.scale\n output_end_y_tile = output_start_y_tile + input_tile_height * self.scale\n\n # put tile into output image\n self.output[:, :, output_start_y:output_end_y,\n output_start_x:output_end_x] = output_tile[:, :, output_start_y_tile:output_end_y_tile,\n output_start_x_tile:output_end_x_tile]\n\n def post_process(self):\n # remove extra pad\n if self.mod_scale is not None:\n _, _, h, w = self.output.size()\n self.output = self.output[:, :, 0:h - self.mod_pad_h * self.scale, 0:w - self.mod_pad_w * self.scale]\n # remove prepad\n if self.pre_pad != 0:\n _, _, h, w = self.output.size()\n self.output = self.output[:, :, 0:h - self.pre_pad * self.scale, 0:w - self.pre_pad * self.scale]\n return self.output\n\n @torch.no_grad()\n def enhance(self, img, outscale=None, alpha_upsampler='realesrgan'):\n h_input, w_input = img.shape[0:2]\n # img: numpy\n img = img.astype(np.float32)\n if np.max(img) > 256: # 16-bit image\n max_range = 65535\n print('\\tInput is a 16-bit image')\n else:\n max_range = 255\n img = img / max_range\n if len(img.shape) == 2: # gray image\n img_mode = 'L'\n img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB)\n elif img.shape[2] == 4: # RGBA image with alpha channel\n img_mode = 'RGBA'\n alpha = img[:, :, 3]\n img = img[:, :, 0:3]\n img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n if alpha_upsampler == 'realesrgan':\n alpha = cv2.cvtColor(alpha, cv2.COLOR_GRAY2RGB)\n else:\n img_mode = 'RGB'\n img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n\n # ------------------- process image (without the alpha channel) ------------------- #\n self.pre_process(img)\n if self.tile_size > 0:\n self.tile_process()\n else:\n self.process()\n output_img = self.post_process()\n output_img = output_img.data.squeeze().float().cpu().clamp_(0, 1).numpy()\n output_img = np.transpose(output_img[[2, 1, 0], :, :], (1, 2, 0))\n if img_mode == 'L':\n output_img = cv2.cvtColor(output_img, cv2.COLOR_BGR2GRAY)\n\n # ------------------- process the alpha channel if necessary ------------------- #\n if img_mode == 'RGBA':\n if alpha_upsampler == 'realesrgan':\n self.pre_process(alpha)\n if self.tile_size > 0:\n self.tile_process()\n else:\n self.process()\n output_alpha = self.post_process()\n output_alpha = output_alpha.data.squeeze().float().cpu().clamp_(0, 1).numpy()\n output_alpha = np.transpose(output_alpha[[2, 1, 0], :, :], (1, 2, 0))\n output_alpha = cv2.cvtColor(output_alpha, cv2.COLOR_BGR2GRAY)\n else: # use the cv2 resize for alpha channel\n h, w = alpha.shape[0:2]\n output_alpha = cv2.resize(alpha, (w * self.scale, h * self.scale), interpolation=cv2.INTER_LINEAR)\n\n # merge the alpha channel\n output_img = cv2.cvtColor(output_img, cv2.COLOR_BGR2BGRA)\n output_img[:, :, 3] = output_alpha\n\n # ------------------------------ return ------------------------------ #\n if max_range == 65535: # 16-bit image\n output = (output_img * 65535.0).round().astype(np.uint16)\n else:\n output = (output_img * 255.0).round().astype(np.uint8)\n\n if outscale is not None and outscale != float(self.scale):\n output = cv2.resize(\n output, (\n int(w_input * outscale),\n int(h_input * outscale),\n ), interpolation=cv2.INTER_LANCZOS4)\n\n return output, img_mode\n\n\nclass PrefetchReader(threading.Thread):\n \"\"\"Prefetch images.\n\n Args:\n img_list (list[str]): A image list of image paths to be read.\n num_prefetch_queue (int): Number of prefetch queue.\n \"\"\"\n\n def __init__(self, img_list, num_prefetch_queue):\n super().__init__()\n self.que = queue.Queue(num_prefetch_queue)\n self.img_list = img_list\n\n def run(self):\n for img_path in self.img_list:\n img = cv2.imread(img_path, cv2.IMREAD_UNCHANGED)\n self.que.put(img)\n\n self.que.put(None)\n\n def __next__(self):\n next_item = self.que.get()\n if next_item is None:\n raise StopIteration\n return next_item\n\n def __iter__(self):\n return self\n\n\nclass IOConsumer(threading.Thread):\n\n def __init__(self, opt, que, qid):\n super().__init__()\n self._queue = que\n self.qid = qid\n self.opt = opt\n\n def run(self):\n while True:\n msg = self._queue.get()\n if isinstance(msg, str) and msg == 'quit':\n break\n\n output = msg['output']\n save_path = msg['save_path']\n cv2.imwrite(save_path, output)\n print(f'IO worker {self.qid} is done.')\n"
},
{
"path": "KDSR-GAN/kdsrgan/weights/README.md",
"content": "# Weights\n\nPut the downloaded weights to this folder.\n"
},
{
"path": "KDSR-GAN/options/test_kdsrgan_x4ST.yml",
"content": "# general settings\nname: test_KDSRGANx4STplus_400k_B12G4\nmodel_type: KDSRGANSTModel\nscale: 4\nnum_gpu: auto # auto: can infer from your visible devices automatically. official: 4 GPUs\nmanual_seed: 0\n\n# ----------------- options for synthesizing training data in RealESRGANModel ----------------- #\n\n# dataset and data loader settings\ndatasets:\n # Uncomment these for validation\n val_1:\n name: NTIRE2020-Track1\n type: PairedImageDataset\n dataroot_gt: /root/dataset/NTIRE2020-Track1/track1-valid-gt\n dataroot_lq: /root/dataset/NTIRE2020-Track1/track1-valid-input\n io_backend:\n type: disk\n\n\n# network structures\nnetwork_g:\n type: BlindSR_ST\n n_feats: 128\n n_encoder_res: 6\n scale: 4\n n_sr_blocks: 42\n\n# network structures\nnetwork_TA:\n type: BlindSR_TA\n n_feats: 128\n n_encoder_res: 6\n scale: 4\n n_sr_blocks: 42\n\nnetwork_d:\n type: UNetDiscriminatorSN\n num_in_ch: 3\n num_feat: 64\n skip_connection: True\n\n# path\npath:\n # use the pre-trained Real-ESRNet model\n pretrain_network_TA: experiments/KDSRT-rec.pth\n pretrain_network_g: experiments/KDSRS-GAN.pth\n # pretrain_network_g: experiments/KDSRS-GANV2.pth\n param_key_g: params_ema\n strict_load_g: False\n ignore_resume_networks: network_TA\n\n\n\nval:\n save_img: True\n suffix: ~ # add suffix to saved images, if None, use exp name\n\n metrics:\n psnr: # metric name\n type: calculate_psnr\n crop_border: 4\n test_y_channel: true\n\n\n\n"
},
{
"path": "KDSR-GAN/options/test_kdsrnet_x4TA.yml",
"content": "# general settings\nname: train_KDSRNetTAx4plus_1000k_B12G4\nmodel_type: KDSRNetTAModel\nscale: 4\nnum_gpu: auto # auto: can infer from your visible devices automatically. official: 4 GPUs\nmanual_seed: 0\n\n# ----------------- options for synthesizing training data in RealESRNetModel ----------------- #\ngt_usm: True # USM the ground-truth\n\n# dataset and data loader settings\ndatasets:\n # Uncomment these for validation\n test_1:\n name: NTIRE2020-Track1\n type: PairedImageDataset\n dataroot_gt: /root/dataset/NTIRE2020-Track1/track1-valid-gt\n dataroot_lq: /root/dataset/NTIRE2020-Track1/track1-valid-input\n io_backend:\n type: disk\n\n# network structures\nnetwork_g:\n type: BlindSR_TA\n n_feats: 128\n n_encoder_res: 6\n scale: 4\n n_sr_blocks: 42\n\n# path\npath:\n pretrain_network_g: ./experiments/KDSRT-rec.pth\n param_key_g: params_ema\n strict_load_g: true\n\n# Uncomment these for validation\n# validation settings\nval:\n save_img: True\n suffix: ~ # add suffix to saved images, if None, use exp name\n\n metrics:\n psnr: # metric name\n type: calculate_psnr\n crop_border: 4\n test_y_channel: true\n \n ssim:\n type: calculate_ssim\n crop_border: 4\n test_y_channel: true\n\n\n"
},
{
"path": "KDSR-GAN/options/train_kdsrgan_x4ST.yml",
"content": "# general settings\nname: train_KDSRGANx4STplus_400k_B12G4\nmodel_type: KDSRGANSTModel\nscale: 4\nnum_gpu: auto # auto: can infer from your visible devices automatically. official: 4 GPUs\nmanual_seed: 0\n\n# ----------------- options for synthesizing training data in RealESRGANModel ----------------- #\nl1_gt_usm: False\npercep_gt_usm: False\ngan_gt_usm: False\n\n# the first degradation process\nresize_prob: [0.2, 0.7, 0.1] # up, down, keep\nresize_range: [0.15, 1.5]\ngaussian_noise_prob: 0.5\nnoise_range: [1, 30]\npoisson_scale_range: [0.05, 3]\ngray_noise_prob: 0.4\njpeg_range: [30, 95]\n\n# the second degradation process\nsecond_blur_prob: 0.8\nresize_prob2: [0.3, 0.4, 0.3] # up, down, keep\nresize_range2: [0.3, 1.2]\ngaussian_noise_prob2: 0.5\nnoise_range2: [1, 25]\npoisson_scale_range2: [0.05, 2.5]\ngray_noise_prob2: 0.4\njpeg_range2: [30, 95]\n\ngt_size: 256\nqueue_size: 180\n\n# dataset and data loader settings\ndatasets:\n train:\n name: DF2K+OST\n type: RealESRGANDataset\n dataroot_gt: /root/dataset\n meta_info: datasets/meta_info/meta_info_DF2Kmultiscale+OST_sub.txt\n io_backend:\n type: disk\n\n blur_kernel_size: 21\n kernel_list: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']\n kernel_prob: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]\n sinc_prob: 0.1\n blur_sigma: [0.2, 3]\n betag_range: [0.5, 4]\n betap_range: [1, 2]\n\n blur_kernel_size2: 21\n kernel_list2: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']\n kernel_prob2: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]\n sinc_prob2: 0.1\n blur_sigma2: [0.2, 1.5]\n betag_range2: [0.5, 4]\n betap_range2: [1, 2]\n\n final_sinc_prob: 0.8\n\n gt_size: 256\n use_hflip: True\n use_rot: False\n\n # data loader\n use_shuffle: true\n num_worker_per_gpu: 6\n batch_size_per_gpu: 9\n dataset_enlarge_ratio: 1\n prefetch_mode: ~\n\n # Uncomment these for validation\n val_1:\n name: NTIRE2020-Track1\n type: PairedImageDataset\n dataroot_gt: /root/dataset/NTIRE2020-Track1/track1-valid-gt\n dataroot_lq: /root/dataset/NTIRE2020-Track1/track1-valid-input\n io_backend:\n type: disk\n\n val_2:\n name: AIM2019-Track2\n type: PairedImageDataset\n dataroot_gt: /root/dataset/AIM2019-Track2/valid-gt-clean\n dataroot_lq: /root/dataset/AIM2019-Track2/valid-input-noisy\n io_backend:\n type: disk\n\n val_3:\n name: RealSR\n type: PairedImageDataset\n dataroot_lq: /root/dataset/RealSR/Canon/Test/4/LR\n dataroot_gt: /root/dataset/RealSR/Canon/Test/4/HR\n io_backend:\n type: disk\n \n\n\n# network structures\nnetwork_g:\n type: BlindSR_ST\n n_feats: 128\n n_encoder_res: 6\n scale: 4\n n_sr_blocks: 42\n\n# network structures\nnetwork_TA:\n type: BlindSR_TA\n n_feats: 128\n n_encoder_res: 6\n scale: 4\n n_sr_blocks: 42\n\nnetwork_d:\n type: UNetDiscriminatorSN\n num_in_ch: 3\n num_feat: 64\n skip_connection: True\n\n# path\npath:\n # use the pre-trained Real-ESRNet model\n pretrain_network_TA: experiments/KDSRT-rec.pth\n pretrain_network_g: experiments/KDSRS-rec.pth\n param_key_g: params_ema\n strict_load_g: True\n resume_state: ~\n ignore_resume_networks: network_TA\n\n# training settings\ntrain:\n ema_decay: 0.999\n optim_g:\n type: Adam\n lr: !!float 1e-4\n weight_decay: 0\n betas: [0.9, 0.99]\n optim_d:\n type: Adam\n lr: !!float 1e-4\n weight_decay: 0\n betas: [0.9, 0.99]\n\n scheduler:\n type: MultiStepLR\n milestones: [400000]\n gamma: 0.5\n \n encoder_iter: 100000\n total_iter: 400000\n lr_encoder: !!float 2e-4\n lr_sr: !!float 1e-4\n gamma_encoder: 0.1\n gamma_sr: 0.5\n lr_decay_encoder: 60000\n lr_decay_sr: 400000\n warmup_iter: -1 # no warm up\n\n\n # losses\n pixel_opt:\n type: L1Loss\n loss_weight: 1.0\n reduction: mean\n # perceptual loss (content and style losses)\n perceptual_opt:\n type: PerceptualLoss\n layer_weights:\n # before relu\n 'conv1_2': 0.1\n 'conv2_2': 0.1\n 'conv3_4': 1\n 'conv4_4': 1\n 'conv5_4': 1\n vgg_type: vgg19\n use_input_norm: true\n perceptual_weight: !!float 1.0\n style_weight: 0\n range_norm: false\n criterion: l1\n # gan loss\n gan_opt:\n type: GANLoss\n gan_type: vanilla\n real_label_val: 1.0\n fake_label_val: 0.0\n loss_weight: !!float 1.0\n \n kd_opt:\n type: KDLoss\n loss_weight: 1\n temperature: 0.15\n\n net_d_iters: 1\n net_d_init_iters: 0\n\n# Uncomment these for validation\n# validation settings\n\nval:\n val_freq: !!float 1e4\n save_img: False\n\n metrics:\n psnr: # metric name\n type: calculate_psnr\n crop_border: 4\n test_y_channel: true\n\n\n# logging settings\nlogger:\n print_freq: 100\n save_checkpoint_freq: !!float 1e4\n use_tb_logger: true\n wandb:\n project: ~\n resume_id: ~\n\n# dist training settings\ndist_params:\n backend: nccl\n port: 29500\n"
},
{
"path": "KDSR-GAN/options/train_kdsrgan_x4STV2.yml",
"content": "# general settings\nname: train_KDSRGANx4STplus_400k_V2\nmodel_type: KDSRGANSTModel\nscale: 4\nnum_gpu: auto # auto: can infer from your visible devices automatically. official: 4 GPUs\nmanual_seed: 0\n\n# ----------------- options for synthesizing training data in RealESRGANModel ----------------- #\nl1_gt_usm: True\npercep_gt_usm: True\ngan_gt_usm: False\n\n# the first degradation process\nresize_prob: [0.2, 0.7, 0.1] # up, down, keep\nresize_range: [0.15, 1.5]\ngaussian_noise_prob: 0.5\nnoise_range: [1, 30]\npoisson_scale_range: [0.05, 3]\ngray_noise_prob: 0.4\njpeg_range: [30, 95]\n\n# the second degradation process\nsecond_blur_prob: 0.8\nresize_prob2: [0.3, 0.4, 0.3] # up, down, keep\nresize_range2: [0.3, 1.2]\ngaussian_noise_prob2: 0.5\nnoise_range2: [1, 25]\npoisson_scale_range2: [0.05, 2.5]\ngray_noise_prob2: 0.4\njpeg_range2: [30, 95]\n\ngt_size: 256\nqueue_size: 180\n\n# dataset and data loader settings\ndatasets:\n train:\n name: DF2K+OST\n type: RealESRGANDataset\n dataroot_gt: /root/dataset\n meta_info: datasets/meta_info/meta_info_DF2Kmultiscale+OST_sub.txt\n io_backend:\n type: disk\n\n blur_kernel_size: 21\n kernel_list: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']\n kernel_prob: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]\n sinc_prob: 0.1\n blur_sigma: [0.2, 3]\n betag_range: [0.5, 4]\n betap_range: [1, 2]\n\n blur_kernel_size2: 21\n kernel_list2: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']\n kernel_prob2: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]\n sinc_prob2: 0.1\n blur_sigma2: [0.2, 1.5]\n betag_range2: [0.5, 4]\n betap_range2: [1, 2]\n\n final_sinc_prob: 0.8\n\n gt_size: 256\n use_hflip: True\n use_rot: False\n\n # data loader\n use_shuffle: true\n num_worker_per_gpu: 6\n batch_size_per_gpu: 9\n dataset_enlarge_ratio: 1\n prefetch_mode: ~\n\n # Uncomment these for validation\n val_1:\n name: NTIRE2020-Track1\n type: PairedImageDataset\n dataroot_gt: /root/dataset/NTIRE2020-Track1/track1-valid-gt\n dataroot_lq: /root/dataset/NTIRE2020-Track1/track1-valid-input\n io_backend:\n type: disk\n\n val_2:\n name: AIM2019-Track2\n type: PairedImageDataset\n dataroot_gt: /root/dataset/AIM2019-Track2/valid-gt-clean\n dataroot_lq: /root/dataset/AIM2019-Track2/valid-input-noisy\n io_backend:\n type: disk\n\n val_3:\n name: RealSR\n type: PairedImageDataset\n dataroot_lq: /root/dataset/RealSR/Canon/Test/4/LR\n dataroot_gt: /root/dataset/RealSR/Canon/Test/4/HR\n io_backend:\n type: disk\n \n\n\n# network structures\nnetwork_g:\n type: BlindSR_ST\n n_feats: 128\n n_encoder_res: 6\n scale: 4\n n_sr_blocks: 42\n\n# network structures\nnetwork_TA:\n type: BlindSR_TA\n n_feats: 128\n n_encoder_res: 6\n scale: 4\n n_sr_blocks: 42\n\nnetwork_d:\n type: UNetDiscriminatorSN\n num_in_ch: 3\n num_feat: 64\n skip_connection: True\n\n# path\npath:\n # use the pre-trained Real-ESRNet model\n pretrain_network_TA: experiments/KDSRT-rec.pth\n pretrain_network_g: experiments/KDSRS-rec.pth\n param_key_g: params_ema\n strict_load_g: True\n resume_state: ~\n ignore_resume_networks: network_TA\n\n# training settings\ntrain:\n ema_decay: 0.999\n optim_g:\n type: Adam\n lr: !!float 1e-4\n weight_decay: 0\n betas: [0.9, 0.99]\n optim_d:\n type: Adam\n lr: !!float 1e-4\n weight_decay: 0\n betas: [0.9, 0.99]\n\n scheduler:\n type: MultiStepLR\n milestones: [400000]\n gamma: 0.5\n \n encoder_iter: 100000\n total_iter: 400000\n lr_encoder: !!float 2e-4\n lr_sr: !!float 1e-4\n gamma_encoder: 0.1\n gamma_sr: 0.5\n lr_decay_encoder: 60000\n lr_decay_sr: 400000\n warmup_iter: -1 # no warm up\n\n\n # losses\n pixel_opt:\n type: L1Loss\n loss_weight: 1.0\n reduction: mean\n # perceptual loss (content and style losses)\n perceptual_opt:\n type: PerceptualLoss\n layer_weights:\n # before relu\n 'conv1_2': 0.1\n 'conv2_2': 0.1\n 'conv3_4': 1\n 'conv4_4': 1\n 'conv5_4': 1\n vgg_type: vgg19\n use_input_norm: true\n perceptual_weight: !!float 1.0\n style_weight: 0\n range_norm: false\n criterion: l1\n # gan loss\n gan_opt:\n type: GANLoss\n gan_type: vanilla\n real_label_val: 1.0\n fake_label_val: 0.0\n loss_weight: !!float 1.0\n \n kd_opt:\n type: KDLoss\n loss_weight: 1\n temperature: 0.15\n\n net_d_iters: 1\n net_d_init_iters: 0\n\n# Uncomment these for validation\n# validation settings\n\nval:\n val_freq: !!float 1e4\n save_img: False\n\n metrics:\n psnr: # metric name\n type: calculate_psnr\n crop_border: 4\n test_y_channel: true\n\n\n# logging settings\nlogger:\n print_freq: 100\n save_checkpoint_freq: !!float 1e4\n use_tb_logger: true\n wandb:\n project: ~\n resume_id: ~\n\n# dist training settings\ndist_params:\n backend: nccl\n port: 29500\n"
},
{
"path": "KDSR-GAN/options/train_kdsrnet_x4ST.yml",
"content": "# general settings\nname: train_KDSRNetSTx4_1000k\nmodel_type: KDSRNetSTModel\nscale: 4\nnum_gpu: auto # auto: can infer from your visible devices automatically. official: 4 GPUs\nmanual_seed: 0\n\n# ----------------- options for synthesizing training data in RealESRNetModel ----------------- #\ngt_usm: False # USM the ground-truth\n\n# the first degradation process\nresize_prob: [0.2, 0.7, 0.1] # up, down, keep\nresize_range: [0.15, 1.5]\ngaussian_noise_prob: 0.5\nnoise_range: [1, 30]\npoisson_scale_range: [0.05, 3]\ngray_noise_prob: 0.4\njpeg_range: [30, 95]\n\n# the second degradation process\nsecond_blur_prob: 0.8\nresize_prob2: [0.3, 0.4, 0.3] # up, down, keep\nresize_range2: [0.3, 1.2]\ngaussian_noise_prob2: 0.5\nnoise_range2: [1, 25]\npoisson_scale_range2: [0.05, 2.5]\ngray_noise_prob2: 0.4\njpeg_range2: [30, 95]\n\ngt_size: 256\nqueue_size: 180\n\n# dataset and data loader settings\ndatasets:\n train:\n name: DF2K+OST\n type: RealESRGANDataset\n dataroot_gt: /root/dataset\n meta_info: datasets/meta_info/meta_info_DF2Kmultiscale+OST_sub.txt\n io_backend:\n type: disk\n\n blur_kernel_size: 21\n kernel_list: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']\n kernel_prob: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]\n sinc_prob: 0.1\n blur_sigma: [0.2, 3]\n betag_range: [0.5, 4]\n betap_range: [1, 2]\n\n blur_kernel_size2: 21\n kernel_list2: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']\n kernel_prob2: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]\n sinc_prob2: 0.1\n blur_sigma2: [0.2, 1.5]\n betag_range2: [0.5, 4]\n betap_range2: [1, 2]\n\n final_sinc_prob: 0.8\n\n gt_size: 256\n use_hflip: True\n use_rot: False\n\n # data loader\n use_shuffle: true\n num_worker_per_gpu: 6\n batch_size_per_gpu: 9\n dataset_enlarge_ratio: 1\n prefetch_mode: ~\n\n # Uncomment these for validation\n val:\n name: NTIRE2020-Track1\n type: PairedImageDataset\n dataroot_gt: /root/dataset/NTIRE2020-Track1/track1-valid-gt\n dataroot_lq: /root/dataset/NTIRE2020-Track1/track1-valid-input\n io_backend:\n type: disk\n\n# network structures\nnetwork_g:\n type: BlindSR_ST\n n_feats: 128\n n_encoder_res: 6\n scale: 4\n n_sr_blocks: 42\n\n# network structures\nnetwork_TA:\n type: BlindSR_TA\n n_feats: 128\n n_encoder_res: 6\n scale: 4\n n_sr_blocks: 42\n\n# path\npath:\n pretrain_network_g: experiments/KDSRT-rec.pth\n pretrain_network_TA: experiments/KDSRT-rec.pth\n param_key_g: params_ema\n strict_load_g: False\n resume_state: ~\n ignore_resume_networks: network_TA\n\n\n# training settings\ntrain:\n ema_decay: 0.999\n\n optim_g:\n type: Adam\n lr: !!float 2e-4\n weight_decay: 0\n betas: [0.9, 0.99]\n\n scheduler:\n type: MultiStepLR\n milestones: [1000000]\n gamma: 0.5\n\n encoder_iter: 100000\n total_iter: 1000000\n lr_encoder: !!float 2e-4\n lr_sr: !!float 2e-4\n gamma_encoder: 0.1\n gamma_sr: 0.5\n lr_decay_encoder: 60000\n lr_decay_sr: 600000\n warmup_iter: -1 # no warm up\n\n # losses\n pixel_opt:\n type: L1Loss\n loss_weight: 1.0\n reduction: mean\n\n kd_opt:\n type: KDLoss\n loss_weight: 1\n temperature: 0.15\n\n# Uncomment these for validation\n# validation settings\nval:\n val_freq: !!float 1e4\n save_img: False\n\n metrics:\n psnr: # metric name\n type: calculate_psnr\n crop_border: 4\n test_y_channel: true\n\n# logging settings\nlogger:\n print_freq: 100\n save_checkpoint_freq: !!float 1e4\n use_tb_logger: true\n wandb:\n project: ~\n resume_id: ~\n\n# dist training settings\ndist_params:\n backend: nccl\n port: 29500\n"
},
{
"path": "KDSR-GAN/options/train_kdsrnet_x4TA.yml",
"content": "# general settings\nname: train_KDSRNetTAx4plus_1000k_B12G4\nmodel_type: KDSRNetTAModel\nscale: 4\nnum_gpu: auto # auto: can infer from your visible devices automatically. official: 4 GPUs\nmanual_seed: 0\n\n# ----------------- options for synthesizing training data in RealESRNetModel ----------------- #\ngt_usm: False # USM the ground-truth\n\n# the first degradation process\nresize_prob: [0.2, 0.7, 0.1] # up, down, keep\nresize_range: [0.15, 1.5]\ngaussian_noise_prob: 0.5\nnoise_range: [1, 30]\npoisson_scale_range: [0.05, 3]\ngray_noise_prob: 0.4\njpeg_range: [30, 95]\n\n# the second degradation process\nsecond_blur_prob: 0.8\nresize_prob2: [0.3, 0.4, 0.3] # up, down, keep\nresize_range2: [0.3, 1.2]\ngaussian_noise_prob2: 0.5\nnoise_range2: [1, 25]\npoisson_scale_range2: [0.05, 2.5]\ngray_noise_prob2: 0.4\njpeg_range2: [30, 95]\n\ngt_size: 256\nqueue_size: 180\n\n# dataset and data loader settings\ndatasets:\n train:\n name: DF2K+OST\n type: RealESRGANDataset\n dataroot_gt: /root/dataset\n meta_info: datasets/meta_info/meta_info_DF2Kmultiscale+OST_sub.txt\n io_backend:\n type: disk\n\n blur_kernel_size: 21\n kernel_list: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']\n kernel_prob: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]\n sinc_prob: 0.1\n blur_sigma: [0.2, 3]\n betag_range: [0.5, 4]\n betap_range: [1, 2]\n\n blur_kernel_size2: 21\n kernel_list2: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']\n kernel_prob2: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]\n sinc_prob2: 0.1\n blur_sigma2: [0.2, 1.5]\n betag_range2: [0.5, 4]\n betap_range2: [1, 2]\n\n final_sinc_prob: 0.8\n\n gt_size: 256\n use_hflip: True\n use_rot: False\n\n # data loader\n use_shuffle: true\n num_worker_per_gpu: 6\n batch_size_per_gpu: 9\n dataset_enlarge_ratio: 1\n prefetch_mode: ~\n\n # Uncomment these for validation\n val:\n name: NTIRE2020-Track1\n type: PairedImageDataset\n dataroot_gt: /root/dataset/NTIRE2020-Track1/track1-valid-gt\n dataroot_lq: /root/dataset/NTIRE2020-Track1/track1-valid-input\n io_backend:\n type: disk\n\n# network structures\nnetwork_g:\n type: BlindSR_TA\n n_feats: 128\n n_encoder_res: 6\n scale: 4\n n_sr_blocks: 42\n\n# path\npath:\n pretrain_network_g: ~\n param_key_g: params_ema\n strict_load_g: true\n resume_state: ~\n\n# training settings\ntrain:\n ema_decay: 0.999\n optim_g:\n type: Adam\n lr: !!float 1e-4\n weight_decay: 0\n betas: [0.9, 0.99]\n\n scheduler:\n type: MultiStepLR\n milestones: [4000000]\n gamma: 0.5\n\n total_iter: 1000000\n warmup_iter: -1 # no warm up\n\n # losses\n pixel_opt:\n type: L1Loss\n loss_weight: 1.0\n reduction: mean\n\n# Uncomment these for validation\n# validation settings\nval:\n val_freq: !!float 1e4\n save_img: False\n\n metrics:\n psnr: # metric name\n type: calculate_psnr\n crop_border: 4\n test_y_channel: true\n\n# logging settings\nlogger:\n print_freq: 100\n save_checkpoint_freq: !!float 1e4\n use_tb_logger: true\n wandb:\n project: ~\n resume_id: ~\n\n# dist training settings\ndist_params:\n backend: nccl\n port: 29500"
},
{
"path": "KDSR-GAN/pip.sh",
"content": "pip install basicsr \npip install -r requirements.txt\npip install pandas \nsudo python3 setup.py develop\n"
},
{
"path": "KDSR-GAN/requirements.txt",
"content": "basicsr>=1.3.3.11\nfacexlib>=0.2.0.3\ngfpgan>=0.2.1\nnumpy\nopencv-python\nPillow\ntorch>=1.7\ntorchvision\ntqdm\n"
},
{
"path": "KDSR-GAN/scripts/extract_subimages.py",
"content": "import argparse\nimport cv2\nimport numpy as np\nimport os\nimport sys\nfrom basicsr.utils import scandir\nfrom multiprocessing import Pool\nfrom os import path as osp\nfrom tqdm import tqdm\n\n\ndef main(args):\n \"\"\"A multi-thread tool to crop large images to sub-images for faster IO.\n\n opt (dict): Configuration dict. It contains:\n n_thread (int): Thread number.\n compression_level (int): CV_IMWRITE_PNG_COMPRESSION from 0 to 9. A higher value means a smaller size\n and longer compression time. Use 0 for faster CPU decompression. Default: 3, same in cv2.\n input_folder (str): Path to the input folder.\n save_folder (str): Path to save folder.\n crop_size (int): Crop size.\n step (int): Step for overlapped sliding window.\n thresh_size (int): Threshold size. Patches whose size is lower than thresh_size will be dropped.\n\n Usage:\n For each folder, run this script.\n Typically, there are GT folder and LQ folder to be processed for DIV2K dataset.\n After process, each sub_folder should have the same number of subimages.\n Remember to modify opt configurations according to your settings.\n \"\"\"\n\n opt = {}\n opt['n_thread'] = args.n_thread\n opt['compression_level'] = args.compression_level\n opt['input_folder'] = args.input\n opt['save_folder'] = args.output\n opt['crop_size'] = args.crop_size\n opt['step'] = args.step\n opt['thresh_size'] = args.thresh_size\n extract_subimages(opt)\n\n\ndef extract_subimages(opt):\n \"\"\"Crop images to subimages.\n\n Args:\n opt (dict): Configuration dict. It contains:\n input_folder (str): Path to the input folder.\n save_folder (str): Path to save folder.\n n_thread (int): Thread number.\n \"\"\"\n input_folder = opt['input_folder']\n save_folder = opt['save_folder']\n if not osp.exists(save_folder):\n os.makedirs(save_folder)\n print(f'mkdir {save_folder} ...')\n else:\n print(f'Folder {save_folder} already exists. Exit.')\n sys.exit(1)\n\n # scan all images\n img_list = list(scandir(input_folder, full_path=True))\n\n pbar = tqdm(total=len(img_list), unit='image', desc='Extract')\n pool = Pool(opt['n_thread'])\n for path in img_list:\n pool.apply_async(worker, args=(path, opt), callback=lambda arg: pbar.update(1))\n pool.close()\n pool.join()\n pbar.close()\n print('All processes done.')\n\n\ndef worker(path, opt):\n \"\"\"Worker for each process.\n\n Args:\n path (str): Image path.\n opt (dict): Configuration dict. It contains:\n crop_size (int): Crop size.\n step (int): Step for overlapped sliding window.\n thresh_size (int): Threshold size. Patches whose size is lower than thresh_size will be dropped.\n save_folder (str): Path to save folder.\n compression_level (int): for cv2.IMWRITE_PNG_COMPRESSION.\n\n Returns:\n process_info (str): Process information displayed in progress bar.\n \"\"\"\n crop_size = opt['crop_size']\n step = opt['step']\n thresh_size = opt['thresh_size']\n img_name, extension = osp.splitext(osp.basename(path))\n\n # remove the x2, x3, x4 and x8 in the filename for DIV2K\n img_name = img_name.replace('x2', '').replace('x3', '').replace('x4', '').replace('x8', '')\n\n img = cv2.imread(path, cv2.IMREAD_UNCHANGED)\n\n h, w = img.shape[0:2]\n h_space = np.arange(0, h - crop_size + 1, step)\n if h - (h_space[-1] + crop_size) > thresh_size:\n h_space = np.append(h_space, h - crop_size)\n w_space = np.arange(0, w - crop_size + 1, step)\n if w - (w_space[-1] + crop_size) > thresh_size:\n w_space = np.append(w_space, w - crop_size)\n\n index = 0\n for x in h_space:\n for y in w_space:\n index += 1\n cropped_img = img[x:x + crop_size, y:y + crop_size, ...]\n cropped_img = np.ascontiguousarray(cropped_img)\n cv2.imwrite(\n osp.join(opt['save_folder'], f'{img_name}_s{index:03d}{extension}'), cropped_img,\n [cv2.IMWRITE_PNG_COMPRESSION, opt['compression_level']])\n process_info = f'Processing {img_name} ...'\n return process_info\n\n\nif __name__ == '__main__':\n parser = argparse.ArgumentParser()\n parser.add_argument('--input', type=str, default='/mnt/bn/xiabinpaint/dataset/DF2K/DF2K_multiscale', help='Input folder')\n parser.add_argument('--output', type=str, default='/mnt/bn/xiabinpaint/dataset/DF2K/DF2K_multiscale_sub', help='Output folder')\n parser.add_argument('--crop_size', type=int, default=400, help='Crop size')\n parser.add_argument('--step', type=int, default=200, help='Step for overlapped sliding window')\n parser.add_argument(\n '--thresh_size',\n type=int,\n default=0,\n help='Threshold size. Patches whose size is lower than thresh_size will be dropped.')\n parser.add_argument('--n_thread', type=int, default=20, help='Thread number.')\n parser.add_argument('--compression_level', type=int, default=3, help='Compression level')\n args = parser.parse_args()\n\n main(args)\n"
},
{
"path": "KDSR-GAN/scripts/extract_subimages_DF2K.py",
"content": "import argparse\nimport cv2\nimport numpy as np\nimport os\nimport sys\nfrom basicsr.utils import scandir\nfrom multiprocessing import Pool\nfrom os import path as osp\nfrom tqdm import tqdm\n\n\ndef main(args):\n \"\"\"A multi-thread tool to crop large images to sub-images for faster IO.\n\n opt (dict): Configuration dict. It contains:\n n_thread (int): Thread number.\n compression_level (int): CV_IMWRITE_PNG_COMPRESSION from 0 to 9. A higher value means a smaller size\n and longer compression time. Use 0 for faster CPU decompression. Default: 3, same in cv2.\n input_folder (str): Path to the input folder.\n save_folder (str): Path to save folder.\n crop_size (int): Crop size.\n step (int): Step for overlapped sliding window.\n thresh_size (int): Threshold size. Patches whose size is lower than thresh_size will be dropped.\n\n Usage:\n For each folder, run this script.\n Typically, there are GT folder and LQ folder to be processed for DIV2K dataset.\n After process, each sub_folder should have the same number of subimages.\n Remember to modify opt configurations according to your settings.\n \"\"\"\n\n opt = {}\n opt['n_thread'] = args.n_thread\n opt['compression_level'] = args.compression_level\n opt['input_folder'] = args.input\n opt['save_folder'] = args.output\n opt['crop_size'] = args.crop_size\n opt['step'] = args.step\n opt['thresh_size'] = args.thresh_size\n extract_subimages(opt)\n\n\ndef extract_subimages(opt):\n \"\"\"Crop images to subimages.\n\n Args:\n opt (dict): Configuration dict. It contains:\n input_folder (str): Path to the input folder.\n save_folder (str): Path to save folder.\n n_thread (int): Thread number.\n \"\"\"\n input_folder = opt['input_folder']\n save_folder = opt['save_folder']\n if not osp.exists(save_folder):\n os.makedirs(save_folder)\n print(f'mkdir {save_folder} ...')\n else:\n print(f'Folder {save_folder} already exists. Exit.')\n sys.exit(1)\n\n # scan all images\n img_list = list(scandir(input_folder, full_path=True))\n\n pbar = tqdm(total=len(img_list), unit='image', desc='Extract')\n pool = Pool(opt['n_thread'])\n for path in img_list:\n pool.apply_async(worker, args=(path, opt), callback=lambda arg: pbar.update(1))\n pool.close()\n pool.join()\n pbar.close()\n print('All processes done.')\n\n\ndef worker(path, opt):\n \"\"\"Worker for each process.\n\n Args:\n path (str): Image path.\n opt (dict): Configuration dict. It contains:\n crop_size (int): Crop size.\n step (int): Step for overlapped sliding window.\n thresh_size (int): Threshold size. Patches whose size is lower than thresh_size will be dropped.\n save_folder (str): Path to save folder.\n compression_level (int): for cv2.IMWRITE_PNG_COMPRESSION.\n\n Returns:\n process_info (str): Process information displayed in progress bar.\n \"\"\"\n crop_size = opt['crop_size']\n step = opt['step']\n thresh_size = opt['thresh_size']\n img_name, extension = osp.splitext(osp.basename(path))\n\n # remove the x2, x3, x4 and x8 in the filename for DIV2K\n img_name = img_name.replace('x2', '').replace('x3', '').replace('x4', '').replace('x8', '')\n\n img = cv2.imread(path, cv2.IMREAD_UNCHANGED)\n\n h, w = img.shape[0:2]\n h_space = np.arange(0, h - crop_size + 1, step)\n if h - (h_space[-1] + crop_size) > thresh_size:\n h_space = np.append(h_space, h - crop_size)\n w_space = np.arange(0, w - crop_size + 1, step)\n if w - (w_space[-1] + crop_size) > thresh_size:\n w_space = np.append(w_space, w - crop_size)\n\n index = 0\n for x in h_space:\n for y in w_space:\n index += 1\n cropped_img = img[x:x + crop_size, y:y + crop_size, ...]\n cropped_img = np.ascontiguousarray(cropped_img)\n cv2.imwrite(\n osp.join(opt['save_folder'], f'{img_name}_s{index:03d}{extension}'), cropped_img,\n [cv2.IMWRITE_PNG_COMPRESSION, opt['compression_level']])\n process_info = f'Processing {img_name} ...'\n return process_info\n\n\nif __name__ == '__main__':\n parser = argparse.ArgumentParser()\n parser.add_argument('--input', type=str, default='/mnt/bn/xiabinpaint/dataset/DF2K/HR', help='Input folder')\n parser.add_argument('--output', type=str, default='/mnt/bn/xiabinpaint/dataset/DF2K/DF2K_sub', help='Output folder')\n parser.add_argument('--crop_size', type=int, default=480, help='Crop size')\n parser.add_argument('--step', type=int, default=240, help='Step for overlapped sliding window')\n parser.add_argument(\n '--thresh_size',\n type=int,\n default=0,\n help='Threshold size. Patches whose size is lower than thresh_size will be dropped.')\n parser.add_argument('--n_thread', type=int, default=20, help='Thread number.')\n parser.add_argument('--compression_level', type=int, default=3, help='Compression level')\n args = parser.parse_args()\n\n main(args)\n"
},
{
"path": "KDSR-GAN/scripts/generate_meta_info.py",
"content": "import argparse\nimport cv2\nimport glob\nimport os\n\n\ndef main(args):\n txt_file = open(args.meta_info, 'w')\n for folder, root in zip(args.input, args.root):\n img_paths = sorted(glob.glob(os.path.join(folder, '*')))\n for img_path in img_paths:\n status = True\n if args.check:\n # read the image once for check, as some images may have errors\n try:\n img = cv2.imread(img_path)\n except (IOError, OSError) as error:\n print(f'Read {img_path} error: {error}')\n status = False\n if img is None:\n status = False\n print(f'Img is None: {img_path}')\n if status:\n # get the relative path\n img_name = os.path.relpath(img_path, root)\n print(img_name)\n txt_file.write(f'{img_name}\\n')\n\n\nif __name__ == '__main__':\n \"\"\"Generate meta info (txt file) for only Ground-Truth images.\n\n It can also generate meta info from several folders into one txt file.\n \"\"\"\n parser = argparse.ArgumentParser()\n parser.add_argument(\n '--input',\n nargs='+',\n default=['datasets/DF2K/DF2K_HR', 'datasets/DF2K/DF2K_multiscale'],\n help='Input folder, can be a list')\n parser.add_argument(\n '--root',\n nargs='+',\n default=['datasets/DF2K', 'datasets/DF2K'],\n help='Folder root, should have the length as input folders')\n parser.add_argument(\n '--meta_info',\n type=str,\n default='datasets/DF2K/meta_info/meta_info_DF2Kmultiscale.txt',\n help='txt path for meta info')\n parser.add_argument('--check', action='store_true', help='Read image to check whether it is ok')\n args = parser.parse_args()\n\n assert len(args.input) == len(args.root), ('Input folder and folder root should have the same length, but got '\n f'{len(args.input)} and {len(args.root)}.')\n os.makedirs(os.path.dirname(args.meta_info), exist_ok=True)\n\n main(args)\n"
},
{
"path": "KDSR-GAN/scripts/generate_meta_info_DF2K.py",
"content": "import argparse\nimport cv2\nimport glob\nimport os\n\n\ndef main(args):\n txt_file = open(args.meta_info, 'w')\n for folder, root in zip(args.input, args.root):\n img_paths = sorted(glob.glob(os.path.join(folder, '*')))\n for img_path in img_paths:\n status = True\n if args.check:\n # read the image once for check, as some images may have errors\n try:\n img = cv2.imread(img_path)\n except (IOError, OSError) as error:\n print(f'Read {img_path} error: {error}')\n status = False\n if img is None:\n status = False\n print(f'Img is None: {img_path}')\n if status:\n # get the relative path\n img_name = os.path.relpath(img_path, root)\n print(img_name)\n txt_file.write(f'{img_name}\\n')\n\n\nif __name__ == '__main__':\n \"\"\"Generate meta info (txt file) for only Ground-Truth images.\n\n It can also generate meta info from several folders into one txt file.\n \"\"\"\n parser = argparse.ArgumentParser()\n parser.add_argument(\n '--input',\n nargs='+',\n default=['/mnt/bn/xiabinpaint/dataset/DF2K/DF2K_sub', '/mnt/bn/xiabinpaint/dataset/DF2K/DF2K_multiscale_sub'],\n help='Input folder, can be a list')\n parser.add_argument(\n '--root',\n nargs='+',\n default=['/mnt/bn/xiabinpaint/dataset', '/mnt/bn/xiabinpaint/dataset'],\n help='Folder root, should have the length as input folders')\n parser.add_argument(\n '--meta_info',\n type=str,\n default='/mnt/bn/xiabinpaint/ICCV-SR/KDSR-GAN/datasets/meta_info/meta_info_DF2Kmultiscale_sub.txt',\n help='txt path for meta info')\n parser.add_argument('--check', action='store_true', help='Read image to check whether it is ok')\n args = parser.parse_args()\n\n assert len(args.input) == len(args.root), ('Input folder and folder root should have the same length, but got '\n f'{len(args.input)} and {len(args.root)}.')\n os.makedirs(os.path.dirname(args.meta_info), exist_ok=True)\n\n main(args)\n"
},
{
"path": "KDSR-GAN/scripts/generate_meta_info_OST.py",
"content": "import argparse\nimport cv2\nimport glob\nimport os\n\n\ndef main(args):\n txt_file = open(args.meta_info, 'w')\n for folder, root in zip(args.input, args.root):\n img_paths = sorted(glob.glob(os.path.join(folder, '*')))\n for img_path in img_paths:\n status = True\n if args.check:\n # read the image once for check, as some images may have errors\n try:\n img = cv2.imread(img_path)\n except (IOError, OSError) as error:\n print(f'Read {img_path} error: {error}')\n status = False\n if img is None:\n status = False\n print(f'Img is None: {img_path}')\n if status:\n # get the relative path\n img_name = os.path.relpath(img_path, root)\n print(img_name)\n txt_file.write(f'{img_name}\\n')\n\n\nif __name__ == '__main__':\n \"\"\"Generate meta info (txt file) for only Ground-Truth images.\n\n It can also generate meta info from several folders into one txt file.\n \"\"\"\n parser = argparse.ArgumentParser()\n parser.add_argument(\n '--input',\n nargs='+',\n default=['/mnt/bn/xiabinpaint/dataset/OST/train/HR'],\n help='Input folder, can be a list')\n parser.add_argument(\n '--root',\n nargs='+',\n default=['/mnt/bn/xiabinpaint/dataset'],\n help='Folder root, should have the length as input folders')\n parser.add_argument(\n '--meta_info',\n type=str,\n default='/mnt/bn/xiabinpaint/ICCV-SR/KDSR-GAN/datasets/meta_info/meta_info_OST.txt',\n help='txt path for meta info')\n parser.add_argument('--check', action='store_true', help='Read image to check whether it is ok')\n args = parser.parse_args()\n\n assert len(args.input) == len(args.root), ('Input folder and folder root should have the same length, but got '\n f'{len(args.input)} and {len(args.root)}.')\n os.makedirs(os.path.dirname(args.meta_info), exist_ok=True)\n\n main(args)\n"
},
{
"path": "KDSR-GAN/scripts/generate_meta_info_pairdata.py",
"content": "import argparse\nimport glob\nimport os\n\n\ndef main(args):\n txt_file = open(args.meta_info, 'w')\n # sca images\n img_paths_gt = sorted(glob.glob(os.path.join(args.input[0], '*')))\n img_paths_lq = sorted(glob.glob(os.path.join(args.input[1], '*')))\n\n assert len(img_paths_gt) == len(img_paths_lq), ('GT folder and LQ folder should have the same length, but got '\n f'{len(img_paths_gt)} and {len(img_paths_lq)}.')\n\n for img_path_gt, img_path_lq in zip(img_paths_gt, img_paths_lq):\n # get the relative paths\n img_name_gt = os.path.relpath(img_path_gt, args.root[0])\n img_name_lq = os.path.relpath(img_path_lq, args.root[1])\n print(f'{img_name_gt}, {img_name_lq}')\n txt_file.write(f'{img_name_gt}, {img_name_lq}\\n')\n\n\nif __name__ == '__main__':\n \"\"\"This script is used to generate meta info (txt file) for paired images.\n \"\"\"\n parser = argparse.ArgumentParser()\n parser.add_argument(\n '--input',\n nargs='+',\n default=['datasets/DF2K/DIV2K_train_HR_sub', 'datasets/DF2K/DIV2K_train_LR_bicubic_X4_sub'],\n help='Input folder, should be [gt_folder, lq_folder]')\n parser.add_argument('--root', nargs='+', default=[None, None], help='Folder root, will use the ')\n parser.add_argument(\n '--meta_info',\n type=str,\n default='datasets/DF2K/meta_info/meta_info_DIV2K_sub_pair.txt',\n help='txt path for meta info')\n args = parser.parse_args()\n\n assert len(args.input) == 2, 'Input folder should have two elements: gt folder and lq folder'\n assert len(args.root) == 2, 'Root path should have two elements: root for gt folder and lq folder'\n os.makedirs(os.path.dirname(args.meta_info), exist_ok=True)\n for i in range(2):\n if args.input[i].endswith('/'):\n args.input[i] = args.input[i][:-1]\n if args.root[i] is None:\n args.root[i] = os.path.dirname(args.input[i])\n\n main(args)\n"
},
{
"path": "KDSR-GAN/scripts/generate_multiscale_DF2K.py",
"content": "import argparse\nimport glob\nimport os\nfrom PIL import Image\n\n\ndef main(args):\n # For DF2K, we consider the following three scales,\n # and the smallest image whose shortest edge is 400\n scale_list = [0.75, 0.5, 1 / 3]\n shortest_edge = 400\n\n path_list = sorted(glob.glob(os.path.join(args.input, '*')))\n for path in path_list:\n print(path)\n basename = os.path.splitext(os.path.basename(path))[0]\n\n img = Image.open(path)\n width, height = img.size\n for idx, scale in enumerate(scale_list):\n print(f'\\t{scale:.2f}')\n rlt = img.resize((int(width * scale), int(height * scale)), resample=Image.LANCZOS)\n rlt.save(os.path.join(args.output, f'{basename}T{idx}.png'))\n\n # save the smallest image which the shortest edge is 400\n if width < height:\n ratio = height / width\n width = shortest_edge\n height = int(width * ratio)\n else:\n ratio = width / height\n height = shortest_edge\n width = int(height * ratio)\n rlt = img.resize((int(width), int(height)), resample=Image.LANCZOS)\n rlt.save(os.path.join(args.output, f'{basename}T{idx+1}.png'))\n\n\nif __name__ == '__main__':\n \"\"\"Generate multi-scale versions for GT images with LANCZOS resampling.\n It is now used for DF2K dataset (DIV2K + Flickr 2K)\n \"\"\"\n parser = argparse.ArgumentParser()\n parser.add_argument('--input', type=str, default='/mnt/bd/dlspace-hl-256g-0001/datasets/DF2K/HR', help='Input folder')\n parser.add_argument('--output', type=str, default='/mnt/bd/dlspace-hl-256g-0001/datasets/DF2K/DF2K_multiscale', help='Output folder')\n args = parser.parse_args()\n os.makedirs(args.output, exist_ok=True)\n main(args)\n"
},
{
"path": "KDSR-GAN/scripts/pytorch2onnx.py",
"content": "import argparse\nimport torch\nimport torch.onnx\nfrom basicsr.archs.rrdbnet_arch import RRDBNet\n\n\ndef main(args):\n # An instance of the model\n model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4)\n if args.params:\n keyname = 'params'\n else:\n keyname = 'params_ema'\n model.load_state_dict(torch.load(args.input)[keyname])\n # set the train mode to false since we will only run the forward pass.\n model.train(False)\n model.cpu().eval()\n\n # An example input\n x = torch.rand(1, 3, 64, 64)\n # Export the model\n with torch.no_grad():\n torch_out = torch.onnx._export(model, x, args.output, opset_version=11, export_params=True)\n print(torch_out.shape)\n\n\nif __name__ == '__main__':\n \"\"\"Convert pytorch model to onnx models\"\"\"\n parser = argparse.ArgumentParser()\n parser.add_argument(\n '--input', type=str, default='experiments/pretrained_models/RealESRGAN_x4plus.pth', help='Input model path')\n parser.add_argument('--output', type=str, default='realesrgan-x4.onnx', help='Output onnx path')\n parser.add_argument('--params', action='store_false', help='Use params instead of params_ema')\n args = parser.parse_args()\n\n main(args)\n"
},
{
"path": "KDSR-GAN/setup.cfg",
"content": "[flake8]\nignore =\n # line break before binary operator (W503)\n W503,\n # line break after binary operator (W504)\n W504,\nmax-line-length=120\n\n[yapf]\nbased_on_style = pep8\ncolumn_limit = 120\nblank_line_before_nested_class_or_def = true\nsplit_before_expression_after_opening_paren = true\n\n[isort]\nline_length = 120\nmulti_line_output = 0\nknown_standard_library = pkg_resources,setuptools\nknown_first_party = realesrgan\nknown_third_party = PIL,basicsr,cv2,numpy,pytest,torch,torchvision,tqdm,yaml\nno_lines_before = STDLIB,LOCALFOLDER\ndefault_section = THIRDPARTY\n\n[codespell]\nskip = .git,./docs/build\ncount =\nquiet-level = 3\n\n[aliases]\ntest=pytest\n\n[tool:pytest]\naddopts=tests/\n"
},
{
"path": "KDSR-GAN/setup.py",
"content": "#!/usr/bin/env python\n\nfrom setuptools import find_packages, setup\n\nimport os\nimport subprocess\nimport time\n\nversion_file = 'kdsrgan/version.py'\n\n\ndef readme():\n with open('README.md', encoding='utf-8') as f:\n content = f.read()\n return content\n\n\ndef get_git_hash():\n\n def _minimal_ext_cmd(cmd):\n # construct minimal environment\n env = {}\n for k in ['SYSTEMROOT', 'PATH', 'HOME']:\n v = os.environ.get(k)\n if v is not None:\n env[k] = v\n # LANGUAGE is used on win32\n env['LANGUAGE'] = 'C'\n env['LANG'] = 'C'\n env['LC_ALL'] = 'C'\n out = subprocess.Popen(cmd, stdout=subprocess.PIPE, env=env).communicate()[0]\n return out\n\n try:\n out = _minimal_ext_cmd(['git', 'rev-parse', 'HEAD'])\n sha = out.strip().decode('ascii')\n except OSError:\n sha = 'unknown'\n\n return sha\n\n\ndef get_hash():\n if os.path.exists('.git'):\n sha = get_git_hash()[:7]\n else:\n sha = 'unknown'\n\n return sha\n\n\ndef write_version_py():\n content = \"\"\"# GENERATED VERSION FILE\n# TIME: {}\n__version__ = '{}'\n__gitsha__ = '{}'\nversion_info = ({})\n\"\"\"\n sha = get_hash()\n with open('VERSION', 'r') as f:\n SHORT_VERSION = f.read().strip()\n VERSION_INFO = ', '.join([x if x.isdigit() else f'\"{x}\"' for x in SHORT_VERSION.split('.')])\n\n version_file_str = content.format(time.asctime(), SHORT_VERSION, sha, VERSION_INFO)\n with open(version_file, 'w') as f:\n f.write(version_file_str)\n\n\ndef get_version():\n with open(version_file, 'r') as f:\n exec(compile(f.read(), version_file, 'exec'))\n return locals()['__version__']\n\n\ndef get_requirements(filename='requirements.txt'):\n here = os.path.dirname(os.path.realpath(__file__))\n with open(os.path.join(here, filename), 'r') as f:\n requires = [line.replace('\\n', '') for line in f.readlines()]\n return requires\n\n\nif __name__ == '__main__':\n write_version_py()\n setup(\n name='realesrgan',\n version=get_version(),\n description='Real-ESRGAN aims at developing Practical Algorithms for General Image Restoration',\n long_description=readme(),\n long_description_content_type='text/markdown',\n author='Xintao Wang',\n author_email='xintao.wang@outlook.com',\n keywords='computer vision, pytorch, image restoration, super-resolution, esrgan, real-esrgan',\n url='https://github.com/xinntao/Real-ESRGAN',\n include_package_data=True,\n packages=find_packages(exclude=('options', 'datasets', 'experiments', 'results', 'tb_logger', 'wandb')),\n classifiers=[\n 'Development Status :: 4 - Beta',\n 'License :: OSI Approved :: Apache Software License',\n 'Operating System :: OS Independent',\n 'Programming Language :: Python :: 3',\n 'Programming Language :: Python :: 3.7',\n 'Programming Language :: Python :: 3.8',\n ],\n license='BSD-3-Clause License',\n setup_requires=['cython', 'numpy'],\n install_requires=get_requirements(),\n zip_safe=False)\n"
},
{
"path": "KDSR-GAN/test.sh",
"content": "\n\n\n\npython3 kdsrgan/test.py -opt options/test_kdsrgan_x4ST.yml "
},
{
"path": "KDSR-GAN/tests/data/gt.lmdb/meta_info.txt",
"content": "baboon.png (480,500,3) 1\ncomic.png (360,240,3) 1\n"
},
{
"path": "KDSR-GAN/tests/data/lq.lmdb/meta_info.txt",
"content": "baboon.png (120,125,3) 1\ncomic.png (80,60,3) 1\n"
},
{
"path": "KDSR-GAN/tests/data/meta_info_gt.txt",
"content": "baboon.png\ncomic.png\n"
},
{
"path": "KDSR-GAN/tests/data/meta_info_pair.txt",
"content": "gt/baboon.png, lq/baboon.png\ngt/comic.png, lq/comic.png\n"
},
{
"path": "KDSR-GAN/tests/data/test_realesrgan_dataset.yml",
"content": "name: Demo\ntype: RealESRGANDataset\ndataroot_gt: tests/data/gt\nmeta_info: tests/data/meta_info_gt.txt\nio_backend:\n type: disk\n\nblur_kernel_size: 21\nkernel_list: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']\nkernel_prob: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]\nsinc_prob: 1\nblur_sigma: [0.2, 3]\nbetag_range: [0.5, 4]\nbetap_range: [1, 2]\n\nblur_kernel_size2: 21\nkernel_list2: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']\nkernel_prob2: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]\nsinc_prob2: 1\nblur_sigma2: [0.2, 1.5]\nbetag_range2: [0.5, 4]\nbetap_range2: [1, 2]\n\nfinal_sinc_prob: 1\n\ngt_size: 128\nuse_hflip: True\nuse_rot: False\n"
},
{
"path": "KDSR-GAN/tests/data/test_realesrgan_model.yml",
"content": "scale: 4\nnum_gpu: 1\nmanual_seed: 0\nis_train: True\ndist: False\n\n# ----------------- options for synthesizing training data ----------------- #\n# USM the ground-truth\nl1_gt_usm: True\npercep_gt_usm: True\ngan_gt_usm: False\n\n# the first degradation process\nresize_prob: [0.2, 0.7, 0.1] # up, down, keep\nresize_range: [0.15, 1.5]\ngaussian_noise_prob: 1\nnoise_range: [1, 30]\npoisson_scale_range: [0.05, 3]\ngray_noise_prob: 1\njpeg_range: [30, 95]\n\n# the second degradation process\nsecond_blur_prob: 1\nresize_prob2: [0.3, 0.4, 0.3] # up, down, keep\nresize_range2: [0.3, 1.2]\ngaussian_noise_prob2: 1\nnoise_range2: [1, 25]\npoisson_scale_range2: [0.05, 2.5]\ngray_noise_prob2: 1\njpeg_range2: [30, 95]\n\ngt_size: 32\nqueue_size: 1\n\n# network structures\nnetwork_g:\n type: RRDBNet\n num_in_ch: 3\n num_out_ch: 3\n num_feat: 4\n num_block: 1\n num_grow_ch: 2\n\nnetwork_d:\n type: UNetDiscriminatorSN\n num_in_ch: 3\n num_feat: 2\n skip_connection: True\n\n# path\npath:\n pretrain_network_g: ~\n param_key_g: params_ema\n strict_load_g: true\n resume_state: ~\n\n# training settings\ntrain:\n ema_decay: 0.999\n optim_g:\n type: Adam\n lr: !!float 1e-4\n weight_decay: 0\n betas: [0.9, 0.99]\n optim_d:\n type: Adam\n lr: !!float 1e-4\n weight_decay: 0\n betas: [0.9, 0.99]\n\n scheduler:\n type: MultiStepLR\n milestones: [400000]\n gamma: 0.5\n\n total_iter: 400000\n warmup_iter: -1 # no warm up\n\n # losses\n pixel_opt:\n type: L1Loss\n loss_weight: 1.0\n reduction: mean\n # perceptual loss (content and style losses)\n perceptual_opt:\n type: PerceptualLoss\n layer_weights:\n # before relu\n 'conv1_2': 0.1\n 'conv2_2': 0.1\n 'conv3_4': 1\n 'conv4_4': 1\n 'conv5_4': 1\n vgg_type: vgg19\n use_input_norm: true\n perceptual_weight: !!float 1.0\n style_weight: 0\n range_norm: false\n criterion: l1\n # gan loss\n gan_opt:\n type: GANLoss\n gan_type: vanilla\n real_label_val: 1.0\n fake_label_val: 0.0\n loss_weight: !!float 1e-1\n\n net_d_iters: 1\n net_d_init_iters: 0\n\n\n# validation settings\nval:\n val_freq: !!float 5e3\n save_img: False\n"
},
{
"path": "KDSR-GAN/tests/data/test_realesrgan_paired_dataset.yml",
"content": "name: Demo\ntype: RealESRGANPairedDataset\nscale: 4\ndataroot_gt: tests/data\ndataroot_lq: tests/data\nmeta_info: tests/data/meta_info_pair.txt\nio_backend:\n type: disk\n\nphase: train\ngt_size: 128\nuse_hflip: True\nuse_rot: False\n"
},
{
"path": "KDSR-GAN/tests/data/test_realesrnet_model.yml",
"content": "scale: 4\nnum_gpu: 1\nmanual_seed: 0\nis_train: True\ndist: False\n\n# ----------------- options for synthesizing training data ----------------- #\ngt_usm: True # USM the ground-truth\n\n# the first degradation process\nresize_prob: [0.2, 0.7, 0.1] # up, down, keep\nresize_range: [0.15, 1.5]\ngaussian_noise_prob: 1\nnoise_range: [1, 30]\npoisson_scale_range: [0.05, 3]\ngray_noise_prob: 1\njpeg_range: [30, 95]\n\n# the second degradation process\nsecond_blur_prob: 1\nresize_prob2: [0.3, 0.4, 0.3] # up, down, keep\nresize_range2: [0.3, 1.2]\ngaussian_noise_prob2: 1\nnoise_range2: [1, 25]\npoisson_scale_range2: [0.05, 2.5]\ngray_noise_prob2: 1\njpeg_range2: [30, 95]\n\ngt_size: 32\nqueue_size: 1\n\n# network structures\nnetwork_g:\n type: RRDBNet\n num_in_ch: 3\n num_out_ch: 3\n num_feat: 4\n num_block: 1\n num_grow_ch: 2\n\n# path\npath:\n pretrain_network_g: ~\n param_key_g: params_ema\n strict_load_g: true\n resume_state: ~\n\n# training settings\ntrain:\n ema_decay: 0.999\n optim_g:\n type: Adam\n lr: !!float 2e-4\n weight_decay: 0\n betas: [0.9, 0.99]\n\n scheduler:\n type: MultiStepLR\n milestones: [1000000]\n gamma: 0.5\n\n total_iter: 1000000\n warmup_iter: -1 # no warm up\n\n # losses\n pixel_opt:\n type: L1Loss\n loss_weight: 1.0\n reduction: mean\n\n\n# validation settings\nval:\n val_freq: !!float 5e3\n save_img: False\n"
},
{
"path": "KDSR-GAN/tests/test_dataset.py",
"content": "import pytest\nimport yaml\n\nfrom realesrgan.data.realesrgan_dataset import RealESRGANDataset\nfrom realesrgan.data.realesrgan_paired_dataset import RealESRGANPairedDataset\n\n\ndef test_realesrgan_dataset():\n\n with open('tests/data/test_realesrgan_dataset.yml', mode='r') as f:\n opt = yaml.load(f, Loader=yaml.FullLoader)\n\n dataset = RealESRGANDataset(opt)\n assert dataset.io_backend_opt['type'] == 'disk' # io backend\n assert len(dataset) == 2 # whether to read correct meta info\n assert dataset.kernel_list == [\n 'iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso'\n ] # correct initialization the degradation configurations\n assert dataset.betag_range2 == [0.5, 4]\n\n # test __getitem__\n result = dataset.__getitem__(0)\n # check returned keys\n expected_keys = ['gt', 'kernel1', 'kernel2', 'sinc_kernel', 'gt_path']\n assert set(expected_keys).issubset(set(result.keys()))\n # check shape and contents\n assert result['gt'].shape == (3, 400, 400)\n assert result['kernel1'].shape == (21, 21)\n assert result['kernel2'].shape == (21, 21)\n assert result['sinc_kernel'].shape == (21, 21)\n assert result['gt_path'] == 'tests/data/gt/baboon.png'\n\n # ------------------ test lmdb backend -------------------- #\n opt['dataroot_gt'] = 'tests/data/gt.lmdb'\n opt['io_backend']['type'] = 'lmdb'\n\n dataset = RealESRGANDataset(opt)\n assert dataset.io_backend_opt['type'] == 'lmdb' # io backend\n assert len(dataset.paths) == 2 # whether to read correct meta info\n assert dataset.kernel_list == [\n 'iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso'\n ] # correct initialization the degradation configurations\n assert dataset.betag_range2 == [0.5, 4]\n\n # test __getitem__\n result = dataset.__getitem__(1)\n # check returned keys\n expected_keys = ['gt', 'kernel1', 'kernel2', 'sinc_kernel', 'gt_path']\n assert set(expected_keys).issubset(set(result.keys()))\n # check shape and contents\n assert result['gt'].shape == (3, 400, 400)\n assert result['kernel1'].shape == (21, 21)\n assert result['kernel2'].shape == (21, 21)\n assert result['sinc_kernel'].shape == (21, 21)\n assert result['gt_path'] == 'comic'\n\n # ------------------ test with sinc_prob = 0 -------------------- #\n opt['dataroot_gt'] = 'tests/data/gt.lmdb'\n opt['io_backend']['type'] = 'lmdb'\n opt['sinc_prob'] = 0\n opt['sinc_prob2'] = 0\n opt['final_sinc_prob'] = 0\n dataset = RealESRGANDataset(opt)\n result = dataset.__getitem__(0)\n # check returned keys\n expected_keys = ['gt', 'kernel1', 'kernel2', 'sinc_kernel', 'gt_path']\n assert set(expected_keys).issubset(set(result.keys()))\n # check shape and contents\n assert result['gt'].shape == (3, 400, 400)\n assert result['kernel1'].shape == (21, 21)\n assert result['kernel2'].shape == (21, 21)\n assert result['sinc_kernel'].shape == (21, 21)\n assert result['gt_path'] == 'baboon'\n\n # ------------------ lmdb backend should have paths ends with lmdb -------------------- #\n with pytest.raises(ValueError):\n opt['dataroot_gt'] = 'tests/data/gt'\n opt['io_backend']['type'] = 'lmdb'\n dataset = RealESRGANDataset(opt)\n\n\ndef test_realesrgan_paired_dataset():\n\n with open('tests/data/test_realesrgan_paired_dataset.yml', mode='r') as f:\n opt = yaml.load(f, Loader=yaml.FullLoader)\n\n dataset = RealESRGANPairedDataset(opt)\n assert dataset.io_backend_opt['type'] == 'disk' # io backend\n assert len(dataset) == 2 # whether to read correct meta info\n\n # test __getitem__\n result = dataset.__getitem__(0)\n # check returned keys\n expected_keys = ['gt', 'lq', 'gt_path', 'lq_path']\n assert set(expected_keys).issubset(set(result.keys()))\n # check shape and contents\n assert result['gt'].shape == (3, 128, 128)\n assert result['lq'].shape == (3, 32, 32)\n assert result['gt_path'] == 'tests/data/gt/baboon.png'\n assert result['lq_path'] == 'tests/data/lq/baboon.png'\n\n # ------------------ test lmdb backend -------------------- #\n opt['dataroot_gt'] = 'tests/data/gt.lmdb'\n opt['dataroot_lq'] = 'tests/data/lq.lmdb'\n opt['io_backend']['type'] = 'lmdb'\n\n dataset = RealESRGANPairedDataset(opt)\n assert dataset.io_backend_opt['type'] == 'lmdb' # io backend\n assert len(dataset) == 2 # whether to read correct meta info\n\n # test __getitem__\n result = dataset.__getitem__(1)\n # check returned keys\n expected_keys = ['gt', 'lq', 'gt_path', 'lq_path']\n assert set(expected_keys).issubset(set(result.keys()))\n # check shape and contents\n assert result['gt'].shape == (3, 128, 128)\n assert result['lq'].shape == (3, 32, 32)\n assert result['gt_path'] == 'comic'\n assert result['lq_path'] == 'comic'\n\n # ------------------ test paired_paths_from_folder -------------------- #\n opt['dataroot_gt'] = 'tests/data/gt'\n opt['dataroot_lq'] = 'tests/data/lq'\n opt['io_backend'] = dict(type='disk')\n opt['meta_info'] = None\n\n dataset = RealESRGANPairedDataset(opt)\n assert dataset.io_backend_opt['type'] == 'disk' # io backend\n assert len(dataset) == 2 # whether to read correct meta info\n\n # test __getitem__\n result = dataset.__getitem__(0)\n # check returned keys\n expected_keys = ['gt', 'lq', 'gt_path', 'lq_path']\n assert set(expected_keys).issubset(set(result.keys()))\n # check shape and contents\n assert result['gt'].shape == (3, 128, 128)\n assert result['lq'].shape == (3, 32, 32)\n\n # ------------------ test normalization -------------------- #\n dataset.mean = [0.5, 0.5, 0.5]\n dataset.std = [0.5, 0.5, 0.5]\n # test __getitem__\n result = dataset.__getitem__(0)\n # check returned keys\n expected_keys = ['gt', 'lq', 'gt_path', 'lq_path']\n assert set(expected_keys).issubset(set(result.keys()))\n # check shape and contents\n assert result['gt'].shape == (3, 128, 128)\n assert result['lq'].shape == (3, 32, 32)\n"
},
{
"path": "KDSR-GAN/tests/test_discriminator_arch.py",
"content": "import torch\n\nfrom realesrgan.archs.discriminator_arch import UNetDiscriminatorSN\n\n\ndef test_unetdiscriminatorsn():\n \"\"\"Test arch: UNetDiscriminatorSN.\"\"\"\n\n # model init and forward (cpu)\n net = UNetDiscriminatorSN(num_in_ch=3, num_feat=4, skip_connection=True)\n img = torch.rand((1, 3, 32, 32), dtype=torch.float32)\n output = net(img)\n assert output.shape == (1, 1, 32, 32)\n\n # model init and forward (gpu)\n if torch.cuda.is_available():\n net.cuda()\n output = net(img.cuda())\n assert output.shape == (1, 1, 32, 32)\n"
},
{
"path": "KDSR-GAN/tests/test_model.py",
"content": "import torch\nimport yaml\nfrom basicsr.archs.rrdbnet_arch import RRDBNet\nfrom basicsr.data.paired_image_dataset import PairedImageDataset\nfrom basicsr.losses.losses import GANLoss, L1Loss, PerceptualLoss\n\nfrom realesrgan.archs.discriminator_arch import UNetDiscriminatorSN\nfrom realesrgan.models.realesrgan_model import RealESRGANModel\nfrom realesrgan.models.realesrnet_model import RealESRNetModel\n\n\ndef test_realesrnet_model():\n with open('tests/data/test_realesrnet_model.yml', mode='r') as f:\n opt = yaml.load(f, Loader=yaml.FullLoader)\n\n # build model\n model = RealESRNetModel(opt)\n # test attributes\n assert model.__class__.__name__ == 'RealESRNetModel'\n assert isinstance(model.net_g, RRDBNet)\n assert isinstance(model.cri_pix, L1Loss)\n assert isinstance(model.optimizers[0], torch.optim.Adam)\n\n # prepare data\n gt = torch.rand((1, 3, 32, 32), dtype=torch.float32)\n kernel1 = torch.rand((1, 5, 5), dtype=torch.float32)\n kernel2 = torch.rand((1, 5, 5), dtype=torch.float32)\n sinc_kernel = torch.rand((1, 5, 5), dtype=torch.float32)\n data = dict(gt=gt, kernel1=kernel1, kernel2=kernel2, sinc_kernel=sinc_kernel)\n model.feed_data(data)\n # check dequeue\n model.feed_data(data)\n # check data shape\n assert model.lq.shape == (1, 3, 8, 8)\n assert model.gt.shape == (1, 3, 32, 32)\n\n # change probability to test if-else\n model.opt['gaussian_noise_prob'] = 0\n model.opt['gray_noise_prob'] = 0\n model.opt['second_blur_prob'] = 0\n model.opt['gaussian_noise_prob2'] = 0\n model.opt['gray_noise_prob2'] = 0\n model.feed_data(data)\n # check data shape\n assert model.lq.shape == (1, 3, 8, 8)\n assert model.gt.shape == (1, 3, 32, 32)\n\n # ----------------- test nondist_validation -------------------- #\n # construct dataloader\n dataset_opt = dict(\n name='Demo',\n dataroot_gt='tests/data/gt',\n dataroot_lq='tests/data/lq',\n io_backend=dict(type='disk'),\n scale=4,\n phase='val')\n dataset = PairedImageDataset(dataset_opt)\n dataloader = torch.utils.data.DataLoader(dataset=dataset, batch_size=1, shuffle=False, num_workers=0)\n assert model.is_train is True\n model.nondist_validation(dataloader, 1, None, False)\n assert model.is_train is True\n\n\ndef test_realesrgan_model():\n with open('tests/data/test_realesrgan_model.yml', mode='r') as f:\n opt = yaml.load(f, Loader=yaml.FullLoader)\n\n # build model\n model = RealESRGANModel(opt)\n # test attributes\n assert model.__class__.__name__ == 'RealESRGANModel'\n assert isinstance(model.net_g, RRDBNet) # generator\n assert isinstance(model.net_d, UNetDiscriminatorSN) # discriminator\n assert isinstance(model.cri_pix, L1Loss)\n assert isinstance(model.cri_perceptual, PerceptualLoss)\n assert isinstance(model.cri_gan, GANLoss)\n assert isinstance(model.optimizers[0], torch.optim.Adam)\n assert isinstance(model.optimizers[1], torch.optim.Adam)\n\n # prepare data\n gt = torch.rand((1, 3, 32, 32), dtype=torch.float32)\n kernel1 = torch.rand((1, 5, 5), dtype=torch.float32)\n kernel2 = torch.rand((1, 5, 5), dtype=torch.float32)\n sinc_kernel = torch.rand((1, 5, 5), dtype=torch.float32)\n data = dict(gt=gt, kernel1=kernel1, kernel2=kernel2, sinc_kernel=sinc_kernel)\n model.feed_data(data)\n # check dequeue\n model.feed_data(data)\n # check data shape\n assert model.lq.shape == (1, 3, 8, 8)\n assert model.gt.shape == (1, 3, 32, 32)\n\n # change probability to test if-else\n model.opt['gaussian_noise_prob'] = 0\n model.opt['gray_noise_prob'] = 0\n model.opt['second_blur_prob'] = 0\n model.opt['gaussian_noise_prob2'] = 0\n model.opt['gray_noise_prob2'] = 0\n model.feed_data(data)\n # check data shape\n assert model.lq.shape == (1, 3, 8, 8)\n assert model.gt.shape == (1, 3, 32, 32)\n\n # ----------------- test nondist_validation -------------------- #\n # construct dataloader\n dataset_opt = dict(\n name='Demo',\n dataroot_gt='tests/data/gt',\n dataroot_lq='tests/data/lq',\n io_backend=dict(type='disk'),\n scale=4,\n phase='val')\n dataset = PairedImageDataset(dataset_opt)\n dataloader = torch.utils.data.DataLoader(dataset=dataset, batch_size=1, shuffle=False, num_workers=0)\n assert model.is_train is True\n model.nondist_validation(dataloader, 1, None, False)\n assert model.is_train is True\n\n # ----------------- test optimize_parameters -------------------- #\n model.feed_data(data)\n model.optimize_parameters(1)\n assert model.output.shape == (1, 3, 32, 32)\n assert isinstance(model.log_dict, dict)\n # check returned keys\n expected_keys = ['l_g_pix', 'l_g_percep', 'l_g_gan', 'l_d_real', 'out_d_real', 'l_d_fake', 'out_d_fake']\n assert set(expected_keys).issubset(set(model.log_dict.keys()))\n"
},
{
"path": "KDSR-GAN/tests/test_utils.py",
"content": "import numpy as np\nfrom basicsr.archs.rrdbnet_arch import RRDBNet\n\nfrom realesrgan.utils import RealESRGANer\n\n\ndef test_realesrganer():\n # initialize with default model\n restorer = RealESRGANer(\n scale=4,\n model_path='experiments/pretrained_models/RealESRGAN_x4plus.pth',\n model=None,\n tile=10,\n tile_pad=10,\n pre_pad=2,\n half=False)\n assert isinstance(restorer.model, RRDBNet)\n assert restorer.half is False\n # initialize with user-defined model\n model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=6, num_grow_ch=32, scale=4)\n restorer = RealESRGANer(\n scale=4,\n model_path='experiments/pretrained_models/RealESRGAN_x4plus_anime_6B.pth',\n model=model,\n tile=10,\n tile_pad=10,\n pre_pad=2,\n half=True)\n # test attribute\n assert isinstance(restorer.model, RRDBNet)\n assert restorer.half is True\n\n # ------------------ test pre_process ---------------- #\n img = np.random.random((12, 12, 3)).astype(np.float32)\n restorer.pre_process(img)\n assert restorer.img.shape == (1, 3, 14, 14)\n # with modcrop\n restorer.scale = 1\n restorer.pre_process(img)\n assert restorer.img.shape == (1, 3, 16, 16)\n\n # ------------------ test process ---------------- #\n restorer.process()\n assert restorer.output.shape == (1, 3, 64, 64)\n\n # ------------------ test post_process ---------------- #\n restorer.mod_scale = 4\n output = restorer.post_process()\n assert output.shape == (1, 3, 60, 60)\n\n # ------------------ test tile_process ---------------- #\n restorer.scale = 4\n img = np.random.random((12, 12, 3)).astype(np.float32)\n restorer.pre_process(img)\n restorer.tile_process()\n assert restorer.output.shape == (1, 3, 64, 64)\n\n # ------------------ test enhance ---------------- #\n img = np.random.random((12, 12, 3)).astype(np.float32)\n result = restorer.enhance(img, outscale=2)\n assert result[0].shape == (24, 24, 3)\n assert result[1] == 'RGB'\n\n # ------------------ test enhance with 16-bit image---------------- #\n img = np.random.random((4, 4, 3)).astype(np.uint16) + 512\n result = restorer.enhance(img, outscale=2)\n assert result[0].shape == (8, 8, 3)\n assert result[1] == 'RGB'\n\n # ------------------ test enhance with gray image---------------- #\n img = np.random.random((4, 4)).astype(np.float32)\n result = restorer.enhance(img, outscale=2)\n assert result[0].shape == (8, 8)\n assert result[1] == 'L'\n\n # ------------------ test enhance with RGBA---------------- #\n img = np.random.random((4, 4, 4)).astype(np.float32)\n result = restorer.enhance(img, outscale=2)\n assert result[0].shape == (8, 8, 4)\n assert result[1] == 'RGBA'\n\n # ------------------ test enhance with RGBA, alpha_upsampler---------------- #\n restorer.tile_size = 0\n img = np.random.random((4, 4, 4)).astype(np.float32)\n result = restorer.enhance(img, outscale=2, alpha_upsampler=None)\n assert result[0].shape == (8, 8, 4)\n assert result[1] == 'RGBA'\n"
},
{
"path": "KDSR-GAN/train_KDSRGAN_S.sh",
"content": "\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \\\npython3 -m torch.distributed.launch --nproc_per_node=8 --master_port=4397 kdsrgan/train.py -opt options/train_kdsrgan_x4ST.yml --launcher pytorch \n\n"
},
{
"path": "KDSR-GAN/train_KDSRNet_S.sh",
"content": "\n\n\n\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \\\npython3 -m torch.distributed.launch --nproc_per_node=8 --master_port=4349 kdsrgan/train.py -opt options/train_kdsrnet_x4ST.yml --launcher pytorch "
},
{
"path": "KDSR-GAN/train_KDSRNet_T.sh",
"content": "\n\n\n\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \\\npython3 -m torch.distributed.launch --nproc_per_node=8 --master_port=3309 kdsrgan/train.py -opt options/train_kdsrnet_x4TA.yml --launcher pytorch "
},
{
"path": "KDSR-classic/README.md",
"content": "# KDSR-classic\n\n\nThis project is the official implementation of 'Knowledge Distillation based Degradation Estimation for Blind Super-Resolution', ICLR2023\n> **Knowledge Distillation based Degradation Estimation for Blind Super-Resolution [[Paper](https://arxiv.org/pdf/2211.16928.pdf)] [[Project](https://github.com/Zj-BinXia/KDSR)]**\n\nThis is code for KDSR-class (for classic degradation model, ie y=kx+n)\n\n\n \n
\n\n---\n\n## Dependencies and Installation\n\n- Python >= 3.8 (Recommend to use [Anaconda](https://www.anaconda.com/download/#linux) or [Miniconda](https://docs.conda.io/en/latest/miniconda.html))\n- [PyTorch >= 1.10](https://pytorch.org/)\n\n## Dataset Preparation\n\nWe use DF2K, which combines [DIV2K](https://data.vision.ee.ethz.ch/cvl/DIV2K/) (800 images) and [Flickr2K](https://github.com/LimBee/NTIRE2017) (2650 images).\n\n---\n\n## Training (4 V100 GPUs)\n\n### Isotropic Gaussian Kernels\n\n1. We train KDSRT-M ( using L1 loss)\n\n```bash\nsh main_iso_KDSRsMx4_stage3.sh \n```\n\n2. we train KDSRS-M (using L1 loss and KD loss). **It is notable that modify the ''pre_train_TA'' and ''pre_train_ST'' of main_iso_KDSRsMx4_stage4.sh to the path of trained KDSRT-M checkpoint.** Then, we run\n\n```bash\nsh main_iso_KDSRsMx4_stage4.sh \n```\n\n### Anisotropic Gaussian Kernels plus noise\n\n1. We train KDSRT ( using L1 loss)\n\n```bash\nsh main_anisonoise_KDSRsMx4_stage3.sh\n```\n\n2. we train KDSRS (using L1 loss and KD loss). **It is notable that modify the ''pre_train_TA'' and ''pre_train_ST'' of main_anisonoise_KDSRsMx4_stage4.sh to the path of trained KDSRT checkpoint.** Then, we run\n\n```bash\nsh main_anisonoise_KDSRsMx4_stage4.sh\n```\n\n---\n\n## :european_castle: Model Zoo\n\nPlease download checkpoints from [Google Drive](https://drive.google.com/drive/folders/113NBvfcrCedvend96KqDiRYVy3N8yprl).\n\n---\n\n## Testing\n\n### Isotropic Gaussian Kernels\n\nTest KDSRsM\n```bash\nsh test_iso_KDSRsMx4.sh\n```\nTest KDSRsL\n\n```bash\nsh test_iso_KDSRsLx4.sh\n```\n### Anisotropic Gaussian Kernels plus noise\n\n```bash\nsh test_anisonoise_KDSRsMx4.sh\n```\n\n---\n\n## Results\n\n ![](\"images/iso-quan.jpg\")
\n
\n\n\n ![](\"images/aniso-quan.jpg\")
\n
\n\n\n---\n\n## BibTeX\n\n @InProceedings{xia2022knowledge,\n title={Knowledge Distillation based Degradation Estimation for Blind Super-Resolution},\n author={Xia, Bin and Zhang, Yulun and Wang, Yitong and Tian, Yapeng and Yang, Wenming and Timofte, Radu and Van Gool, Luc},\n journal={ICLR},\n year={2023}\n }\n\n\n## Acknowledgements\nThis code is built on [EDSR (PyTorch)](https://github.com/thstkdgus35/EDSR-PyTorch), [IKC](https://github.com/yuanjunchai/IKC). We thank the authors for sharing the codes.\n\n## 📧 Contact\nIf you have any question, please email `zjbinxia@gmail.com`.\n\n"
},
{
"path": "KDSR-classic/cal_parms.py",
"content": "def count_param(model):\n param_count = 0\n for param in model.parameters():\n param_count += param.view(-1).size()[0]\n return param_count\n\nfrom option import args\nfrom model_ST.blindsr import make_model\n# args.n_feats=64\nargs.scale=[4]\nnet = make_model(args)\nprint(\"KDSRs-M parameters (M):\",count_param(net))\n\nargs.scale=[4]\nargs.n_feats=128 \nargs.n_blocks=28 \nargs.n_resblocks=5\nnet = make_model(args)\nprint(\"KDSRs-L parameters (M):\",count_param(net))\n"
},
{
"path": "KDSR-classic/data/__init__.py",
"content": "from importlib import import_module\n#from dataloader import MSDataLoader\nfrom torch.utils.data import dataloader\nfrom torch.utils.data import ConcatDataset\n\n# This is a simple wrapper function for ConcatDataset\nclass MyConcatDataset(ConcatDataset):\n def __init__(self, datasets):\n super(MyConcatDataset, self).__init__(datasets)\n self.train = datasets[0].train\n\n def set_scale(self, idx_scale):\n for d in self.datasets:\n if hasattr(d, 'set_scale'): d.set_scale(idx_scale)\n\nclass Data:\n def __init__(self, args):\n self.loader_train = None\n if not args.test_only:\n datasets = []\n for d in args.data_train:\n print(d)\n module_name = d if d.find('DIV2K-Q') < 0 else 'DIV2KJPEG'\n m = import_module('data.' + module_name.lower())\n datasets.append(getattr(m, module_name)(args, name=d))\n\n self.loader_train = dataloader.DataLoader(\n MyConcatDataset(datasets),\n batch_size=args.batch_size,\n shuffle=True,\n pin_memory=not args.cpu,\n num_workers=args.n_threads,\n )\n\n self.loader_test = []\n for d in args.data_test:\n if d in ['Set5', 'Set14', 'B100', 'Urban100',\"MANGA109\"]:\n m = import_module('data.benchmark')\n testset = getattr(m, 'Benchmark')(args, train=False, name=d)\n else:\n module_name = d if d.find('DIV2K-Q') < 0 else 'DIV2KJPEG'\n m = import_module('data.' + module_name.lower())\n testset = getattr(m, module_name)(args, train=False, name=d)\n\n self.loader_test.append(\n dataloader.DataLoader(\n testset,\n batch_size=1,\n shuffle=False,\n pin_memory=not args.cpu,\n num_workers=args.n_threads,\n )\n )\n\n"
},
{
"path": "KDSR-classic/data/benchmark.py",
"content": "import os\nfrom data import common\nfrom data import multiscalesrdata as srdata\n\n\nclass Benchmark(srdata.SRData):\n def __init__(self, args, name='', train=True):\n super(Benchmark, self).__init__(\n args, name=name, train=train, benchmark=True\n )\n\n def _set_filesystem(self, dir_data):\n self.apath = os.path.join(dir_data,'benchmark', self.name)\n self.dir_hr = os.path.join(self.apath, 'HR')\n self.dir_lr = os.path.join(self.apath, 'LR_bicubic')\n self.ext = ('.png','.png')\n print(self.dir_hr)\n print(self.dir_lr)\n"
},
{
"path": "KDSR-classic/data/common.py",
"content": "import random\n\nimport numpy as np\nimport skimage.color as sc\n\nimport torch\n\n\ndef get_patch(hr, patch_size=48, scale=2, multi=False, input_large=False):\n ih, iw = hr.shape[:2]\n\n ip = scale * patch_size\n\n ix = random.randrange(0, iw - ip + 1)\n iy = random.randrange(0, ih - ip + 1)\n\n hr = hr[iy:iy + ip, ix:ix + ip, :]\n\n return hr\n\n\ndef set_channel(hr, n_channels=3):\n def _set_channel(img):\n if img.ndim == 2:\n img = np.expand_dims(img, axis=2)\n\n c = img.shape[2]\n if n_channels == 1 and c == 3:\n img = np.expand_dims(sc.rgb2ycbcr(img)[:, :, 0], 2)\n elif n_channels == 3 and c == 1:\n img = np.concatenate([img] * n_channels, 2)\n\n return img\n\n return _set_channel(hr)\n\n\ndef np2Tensor(hr, rgb_range=255):\n def _np2Tensor(img):\n np_transpose = np.ascontiguousarray(img.transpose((2, 0, 1)))\n tensor = torch.from_numpy(np_transpose).float()\n tensor.mul_(rgb_range / 255)\n\n return tensor\n\n return _np2Tensor(hr)\n\n\ndef augment(hr, hflip=True, rot=True):\n hflip = hflip and random.random() < 0.5\n vflip = rot and random.random() < 0.5\n rot90 = rot and random.random() < 0.5\n\n def _augment(img):\n if hflip: img = img[:, ::-1, :]\n if vflip: img = img[::-1, :, :]\n if rot90: img = img.transpose(1, 0, 2)\n\n return img\n\n return _augment(hr)\n\n"
},
{
"path": "KDSR-classic/data/common2.py",
"content": "import random\n\nimport numpy as np\nimport skimage.color as sc\n\nimport torch\n\n\ndef get_patch(*args, patch_size=96, scale=2, multi=False, input_large=False):\n ih, iw = args[0].shape[:2]\n\n if not input_large:\n p = scale if multi else 1\n tp = p * patch_size\n ip = tp // scale\n else:\n tp = patch_size\n ip = patch_size\n\n ix = random.randrange(0, iw - ip + 1)\n iy = random.randrange(0, ih - ip + 1)\n\n if not input_large:\n tx, ty = scale * ix, scale * iy\n else:\n tx, ty = ix, iy\n\n ret = [\n args[0][iy:iy + ip, ix:ix + ip, :],\n *[a[ty:ty + tp, tx:tx + tp, :] for a in args[1:]]\n ]\n\n return ret\n\n\ndef set_channel(*args, n_channels=3):\n def _set_channel(img):\n if img.ndim == 2:\n img = np.expand_dims(img, axis=2)\n\n c = img.shape[2]\n if n_channels == 1 and c == 3:\n img = np.expand_dims(sc.rgb2ycbcr(img)[:, :, 0], 2)\n elif n_channels == 3 and c == 1:\n img = np.concatenate([img] * n_channels, 2)\n\n return img\n\n return [_set_channel(a) for a in args]\n\n\ndef np2Tensor(*args, rgb_range=255):\n def _np2Tensor(img):\n np_transpose = np.ascontiguousarray(img.transpose((2, 0, 1)))\n tensor = torch.from_numpy(np_transpose).float()\n tensor.mul_(rgb_range / 255)\n\n return tensor\n\n return [_np2Tensor(a) for a in args]\n\n\ndef augment(*args, hflip=True, rot=True):\n hflip = hflip and random.random() < 0.5\n vflip = rot and random.random() < 0.5\n rot90 = rot and random.random() < 0.5\n\n def _augment(img):\n if hflip: img = img[:, ::-1, :]\n if vflip: img = img[::-1, :, :]\n if rot90: img = img.transpose(1, 0, 2)\n\n return img\n\n return [_augment(a) for a in args]"
},
{
"path": "KDSR-classic/data/df2k.py",
"content": "import os\nfrom data import multiscalesrdata\n\n\nclass DF2K(multiscalesrdata.SRData):\n def __init__(self, args, name='DF2K', train=True, benchmark=False):\n data_range = [r.split('-') for r in args.data_range.split('/')]\n if train:\n data_range = data_range[0]\n else:\n if args.test_only and len(data_range) == 1:\n data_range = data_range[0]\n else:\n data_range = data_range[1]\n\n self.begin, self.end = list(map(lambda x: int(x), data_range))\n super(DF2K, self).__init__(args, name=name, train=train, benchmark=benchmark)\n\n def _scan(self):\n names_hr = super(DF2K, self)._scan()\n names_hr = names_hr[self.begin - 1:self.end]\n\n return names_hr\n\n def _set_filesystem(self, dir_data):\n super(DF2K, self)._set_filesystem(dir_data)\n self.dir_hr = os.path.join(self.apath, 'HR')\n self.dir_lr = os.path.join(self.apath, 'LR_bicubic')\n\n"
},
{
"path": "KDSR-classic/data/multiscalesrdata.py",
"content": "import os\nimport glob\nimport random\nimport pickle\n\nfrom data import common\n\nimport numpy as np\nimport imageio\nimport torch\nimport torch.utils.data as data\n\n\nclass SRData(data.Dataset):\n def __init__(self, args, name='', train=True, benchmark=False):\n self.args = args\n self.name = name\n self.train = train\n self.split = 'train' if train else 'test'\n self.do_eval = True\n self.benchmark = benchmark\n self.input_large = (args.model == 'VDSR')\n self.scale = args.scale\n self.idx_scale = 0\n\n self._set_filesystem(args.dir_data)\n if args.ext.find('img') < 0:\n path_bin = os.path.join(self.apath, 'bin')\n os.makedirs(path_bin, exist_ok=True)\n\n list_hr, list_lr = self._scan()\n if args.ext.find('img') >= 0 or benchmark:\n self.images_hr, self.images_lr = list_hr, list_lr\n elif args.ext.find('sep') >= 0:\n os.makedirs(\n self.dir_hr.replace(self.apath, path_bin),\n exist_ok=True\n )\n for s in self.scale:\n os.makedirs(\n os.path.join(\n self.dir_lr.replace(self.apath, path_bin),\n 'X{}'.format(s)\n ),\n exist_ok=True\n )\n\n self.images_hr, self.images_lr = [], [[] for _ in self.scale]\n for h in list_hr:\n b = h.replace(self.apath, path_bin)\n b = b.replace(self.ext[0], '.pt')\n self.images_hr.append(b)\n self._check_and_load(args.ext, h, b, verbose=True)\n for i, ll in enumerate(list_lr):\n for l in ll:\n b = l.replace(self.apath, path_bin)\n b = b.replace(self.ext[1], '.pt')\n self.images_lr[i].append(b)\n self._check_and_load(args.ext, l, b, verbose=True)\n if train:\n n_patches = args.batch_size * args.test_every\n n_images = len(args.data_train) * len(self.images_hr)\n if n_images == 0:\n self.repeat = 0\n else:\n self.repeat = max(n_patches // n_images, 1)\n\n # Below functions as used to prepare images\n def _scan(self):\n names_hr = sorted(\n glob.glob(os.path.join(self.dir_hr, '*' + self.ext[0]))\n )\n names_lr = [[] for _ in self.scale]\n for f in names_hr:\n filename, _ = os.path.splitext(os.path.basename(f))\n for si, s in enumerate(self.scale):\n names_lr[si].append(os.path.join(\n self.dir_lr, 'X{}/{}x{}{}'.format(\n s, filename, s, self.ext[1]\n )\n ))\n\n return names_hr, names_lr\n\n def _set_filesystem(self, dir_data):\n self.apath = os.path.join(dir_data, self.name)\n self.dir_hr = os.path.join(self.apath, 'HR')\n self.dir_lr = os.path.join(self.apath, 'LR_bicubic')\n if self.input_large: self.dir_lr += 'L'\n self.ext = ('.png', '.png')\n\n def _check_and_load(self, ext, img, f, verbose=True):\n if not os.path.isfile(f) or ext.find('reset') >= 0:\n if verbose:\n print('Making a binary: {}'.format(f))\n with open(f, 'wb') as _f:\n pickle.dump(imageio.imread(img), _f)\n\n def __getitem__(self, idx):\n hr, filename = self._load_file(idx)\n hr = self.get_patch( hr)\n hr = common.set_channel(hr, n_channels=self.args.n_colors)\n hr = common.np2Tensor(hr, rgb_range=self.args.rgb_range)\n\n return hr, filename\n\n def __len__(self):\n if self.train:\n return len(self.images_hr) * self.repeat\n else:\n return len(self.images_hr)\n\n def _get_index(self, idx):\n if self.train:\n return idx % len(self.images_hr)\n else:\n return idx\n\n def _load_file(self, idx):\n idx = self._get_index(idx)\n f_hr = self.images_hr[idx]\n\n filename, _ = os.path.splitext(os.path.basename(f_hr))\n if self.args.ext == 'img' or self.benchmark:\n hr = imageio.imread(f_hr)\n elif self.args.ext.find('sep') >= 0:\n with open(f_hr, 'rb') as _f:\n hr = pickle.load(_f)\n\n return hr, filename\n\n def get_patch(self, hr):\n scale = self.scale[self.idx_scale]\n if self.train:\n hr = common.get_patch(\n hr,\n patch_size=self.args.patch_size,\n scale=scale,\n multi=(len(self.scale) > 1),\n input_large=self.input_large\n )\n if not self.args.no_augment: hr = common.augment( hr)\n else:\n ih, iw = hr.shape[:2]\n ih = ih//scale\n iw = iw // scale\n hr = hr[0:ih * scale, 0:iw * scale]\n\n return hr\n\n def set_scale(self, idx_scale):\n if not self.input_large:\n self.idx_scale = idx_scale\n else:\n self.idx_scale = random.randint(0, len(self.scale) - 1)\n\n"
},
{
"path": "KDSR-classic/data/multiscalesrdata2.py",
"content": "import os\nimport glob\nimport random\nimport pickle\n\nfrom data import common2 as common\n\nimport numpy as np\nimport imageio\nimport torch\nimport torch.utils.data as data\n\n\nclass SRData(data.Dataset):\n def __init__(self, args, name='', train=True, benchmark=False):\n self.args = args\n self.name = name\n self.train = train\n self.split = 'train' if train else 'test'\n self.do_eval = True\n self.benchmark = benchmark\n self.input_large = (args.model == 'VDSR')\n self.scale = args.scale\n self.idx_scale = 0\n\n self._set_filesystem(args.dir_data)\n if args.ext.find('img') < 0:\n path_bin = os.path.join(self.apath, 'bin')\n os.makedirs(path_bin, exist_ok=True)\n\n list_hr, list_lr = self._scan()\n if args.ext.find('img') >= 0 or benchmark:\n self.images_hr, self.images_lr = list_hr, list_lr\n elif args.ext.find('sep') >= 0:\n os.makedirs(\n self.dir_hr.replace(self.apath, path_bin),\n exist_ok=True\n )\n for s in self.scale:\n os.makedirs(\n os.path.join(\n self.dir_lr.replace(self.apath, path_bin),\n 'X{}'.format(s)\n ),\n exist_ok=True\n )\n\n self.images_hr, self.images_lr = [], [[] for _ in self.scale]\n for h in list_hr:\n b = h.replace(self.apath, path_bin)\n b = b.replace(self.ext[0], '.pt')\n self.images_hr.append(b)\n self._check_and_load(args.ext, h, b, verbose=True)\n for i, ll in enumerate(list_lr):\n for l in ll:\n b = l.replace(self.apath, path_bin)\n b = b.replace(self.ext[1], '.pt')\n self.images_lr[i].append(b)\n self._check_and_load(args.ext, l, b, verbose=True)\n if train:\n n_patches = args.batch_size * args.test_every\n n_images = len(args.data_train) * len(self.images_hr)\n if n_images == 0:\n self.repeat = 0\n else:\n self.repeat = max(n_patches // n_images, 1)\n\n # Below functions as used to prepare images\n def _scan(self):\n names_hr = sorted(\n glob.glob(os.path.join(self.dir_hr, '*' + self.ext[0]))\n )\n\n names_lr = [[] for _ in self.scale]\n for f in names_hr:\n filename, _ = os.path.splitext(os.path.basename(f))\n for si, s in enumerate(self.scale):\n names_lr[si].append(os.path.join(\n self.dir_lr, 'X{}/{}x{}{}'.format(\n s, filename, s, self.ext[1]\n )\n ))\n\n return names_hr, names_lr\n\n def _set_filesystem(self, dir_data):\n self.apath = os.path.join(dir_data, self.name)\n self.dir_hr = os.path.join(self.apath, 'HR')\n self.dir_lr = os.path.join(self.apath, 'LR_bicubic')\n if self.input_large: self.dir_lr += 'L'\n self.ext = ('.png', '.png')\n\n def _check_and_load(self, ext, img, f, verbose=True):\n if not os.path.isfile(f) or ext.find('reset') >= 0:\n if verbose:\n print('Making a binary: {}'.format(f))\n with open(f, 'wb') as _f:\n pickle.dump(imageio.imread(img,pilmode=\"RGB\"), _f)\n\n def __getitem__(self, idx):\n lr, hr, filename = self._load_file(idx)\n pair = self.get_patch(lr, hr)\n pair = common.set_channel(*pair, n_channels=self.args.n_colors)\n pair_t = common.np2Tensor(*pair, rgb_range=self.args.rgb_range)\n\n return pair_t[0], pair_t[1], filename\n\n def __len__(self):\n if self.train:\n return len(self.images_hr) * self.repeat\n else:\n return len(self.images_hr)\n\n def _get_index(self, idx):\n if self.train:\n return idx % len(self.images_hr)\n else:\n return idx\n\n def _load_file(self, idx):\n idx = self._get_index(idx)\n f_hr = self.images_hr[idx]\n f_lr = self.images_lr[self.idx_scale][idx]\n\n filename, _ = os.path.splitext(os.path.basename(f_hr))\n if self.args.ext == 'img' or self.benchmark:\n hr = imageio.imread(f_hr,pilmode=\"RGB\")\n lr = imageio.imread(f_lr,pilmode=\"RGB\")\n elif self.args.ext.find('sep') >= 0:\n with open(f_hr, 'rb') as _f:\n hr = pickle.load(_f)\n with open(f_lr, 'rb') as _f:\n lr = pickle.load(_f)\n\n return lr, hr, filename\n\n def get_patch(self, lr, hr):\n scale = self.scale[self.idx_scale]\n if self.train:\n lr, hr = common.get_patch(\n lr, hr,\n patch_size=self.args.patch_size,\n scale=scale,\n multi=(len(self.scale) > 1),\n input_large=self.input_large\n )\n if not self.args.no_augment: lr, hr = common.augment(lr, hr)\n else:\n ih, iw = lr.shape[:2]\n hr = hr[0:ih * scale, 0:iw * scale]\n\n return lr, hr\n\n def set_scale(self, idx_scale):\n if not self.input_large:\n self.idx_scale = idx_scale\n else:\n self.idx_scale = random.randint(0, len(self.scale) - 1)"
},
{
"path": "KDSR-classic/dataloader.py",
"content": "import sys\nimport threading\nimport queue\nimport random\nimport collections\n\nimport torch\nimport torch.multiprocessing as multiprocessing\n\nfrom torch._C import _set_worker_signal_handlers\nfrom torch.utils.data.dataloader import DataLoader\nfrom torch.utils.data.dataloader import _DataLoaderIter\nfrom torch.utils.data import _utils\n\nif sys.version_info[0] == 2:\n import Queue as queue\nelse:\n import queue\n\ndef _ms_loop(dataset, index_queue, data_queue, collate_fn, scale, seed, init_fn, worker_id):\n global _use_shared_memory\n _use_shared_memory = True\n _set_worker_signal_handlers()\n\n torch.set_num_threads(1)\n torch.manual_seed(seed)\n while True:\n r = index_queue.get()\n if r is None:\n break\n idx, batch_indices = r\n try:\n idx_scale = 0\n if len(scale) > 1 and dataset.train:\n idx_scale = random.randrange(0, len(scale))\n dataset.set_scale(idx_scale)\n\n samples = collate_fn([dataset[i] for i in batch_indices])\n samples.append(idx_scale)\n\n except Exception:\n data_queue.put((idx, _utils.ExceptionWrapper(sys.exc_info())))\n else:\n data_queue.put((idx, samples))\n\nclass _MSDataLoaderIter(_DataLoaderIter):\n def __init__(self, loader):\n self.dataset = loader.dataset\n self.scale = loader.scale\n self.collate_fn = loader.collate_fn\n self.batch_sampler = loader.batch_sampler\n self.num_workers = loader.num_workers\n self.pin_memory = loader.pin_memory and torch.cuda.is_available()\n self.timeout = loader.timeout\n self.done_event = threading.Event()\n\n self.sample_iter = iter(self.batch_sampler)\n\n if self.num_workers > 0:\n self.worker_init_fn = loader.worker_init_fn\n self.index_queues = [\n multiprocessing.Queue() for _ in range(self.num_workers)\n ]\n self.worker_queue_idx = 0\n self.worker_result_queue = multiprocessing.Queue()\n self.batches_outstanding = 0\n self.worker_pids_set = False\n self.shutdown = False\n self.send_idx = 0\n self.rcvd_idx = 0\n self.reorder_dict = {}\n\n base_seed = torch.LongTensor(1).random_()[0]\n self.workers = [\n multiprocessing.Process(\n target=_ms_loop,\n args=(\n self.dataset,\n self.index_queues[i],\n self.worker_result_queue,\n self.collate_fn,\n self.scale,\n base_seed + i,\n self.worker_init_fn,\n i\n )\n )\n for i in range(self.num_workers)]\n\n if self.pin_memory or self.timeout > 0:\n self.data_queue = queue.Queue()\n if self.pin_memory:\n maybe_device_id = torch.cuda.current_device()\n else:\n # do not initialize cuda context if not necessary\n maybe_device_id = None\n self.pin_memory_thread = threading.Thread(\n target=_utils.pin_memory._pin_memory_loop,\n args=(self.worker_result_queue, self.data_queue, maybe_device_id, self.done_event))\n self.pin_memory_thread.daemon = True\n self.pin_memory_thread.start()\n else:\n self.data_queue = self.worker_result_queue\n\n for w in self.workers:\n w.daemon = True # ensure that the worker exits on process exit\n w.start()\n\n _utils.signal_handling._set_worker_pids(id(self), tuple(w.pid for w in self.workers))\n _utils.signal_handling._set_SIGCHLD_handler()\n self.worker_pids_set = True\n\n # prime the prefetch loop\n for _ in range(2 * self.num_workers):\n self._put_indices()\n\nclass MSDataLoader(DataLoader):\n def __init__(\n self, args, dataset, batch_size=1, shuffle=False,\n sampler=None, batch_sampler=None,\n collate_fn=_utils.collate.default_collate, pin_memory=False, drop_last=True,\n timeout=0, worker_init_fn=None):\n\n super(MSDataLoader, self).__init__(\n dataset, batch_size=batch_size, shuffle=shuffle,\n sampler=sampler, batch_sampler=batch_sampler,\n num_workers=args.n_threads, collate_fn=collate_fn,\n pin_memory=pin_memory, drop_last=drop_last,\n timeout=timeout, worker_init_fn=worker_init_fn)\n\n self.scale = args.scale\n\n def __iter__(self):\n return _MSDataLoaderIter(self)\n"
},
{
"path": "KDSR-classic/loss/__init__.py",
"content": "import os\nfrom importlib import import_module\n\nimport matplotlib.pyplot as plt\n\nimport numpy as np\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\n\nclass Loss(nn.modules.loss._Loss):\n def __init__(self, args, ckp):\n super(Loss, self).__init__()\n print('Preparing loss function:')\n\n self.n_GPUs = args.n_GPUs\n self.loss = []\n self.loss_module = nn.ModuleList()\n for loss in args.loss.split('+'):\n weight, loss_type = loss.split('*')\n if loss_type == 'MSE':\n loss_function = nn.MSELoss()\n elif loss_type == 'L1':\n loss_function = nn.L1Loss()\n elif loss_type == 'CE':\n loss_function = nn.CrossEntropyLoss()\n elif loss_type.find('VGG') >= 0:\n module = import_module('loss.vgg')\n loss_function = getattr(module, 'VGG')(\n loss_type[3:],\n rgb_range=args.rgb_range\n )\n elif loss_type.find('GAN') >= 0:\n module = import_module('loss.adversarial')\n loss_function = getattr(module, 'Adversarial')(\n args,\n loss_type\n )\n \n self.loss.append({\n 'type': loss_type,\n 'weight': float(weight),\n 'function': loss_function}\n )\n if loss_type.find('GAN') >= 0:\n self.loss.append({'type': 'DIS', 'weight': 1, 'function': None})\n\n if len(self.loss) > 1:\n self.loss.append({'type': 'Total', 'weight': 0, 'function': None})\n\n for l in self.loss:\n if l['function'] is not None:\n print('{:.3f} * {}'.format(l['weight'], l['type']))\n self.loss_module.append(l['function'])\n\n self.log = torch.Tensor()\n\n device = torch.device('cpu' if args.cpu else 'cuda')\n self.loss_module.to(device)\n if args.precision == 'half': self.loss_module.half()\n if not args.cpu and args.n_GPUs > 1:\n self.loss_module = nn.DataParallel(\n self.loss_module, range(args.n_GPUs)\n )\n\n if args.load != '.': self.load(ckp.dir, cpu=args.cpu)\n\n def forward(self, sr, hr):\n losses = []\n for i, l in enumerate(self.loss):\n if l['function'] is not None:\n loss = l['function'](sr, hr)\n effective_loss = l['weight'] * loss\n losses.append(effective_loss)\n self.log[-1, i] += effective_loss.item()\n elif l['type'] == 'DIS':\n self.log[-1, i] += self.loss[i - 1]['function'].loss\n\n loss_sum = sum(losses)\n if len(self.loss) > 1:\n self.log[-1, -1] += loss_sum.item()\n\n return loss_sum\n\n def step(self):\n for l in self.get_loss_module():\n if hasattr(l, 'scheduler'):\n l.scheduler.step()\n\n def start_log(self):\n self.log = torch.cat((self.log, torch.zeros(1, len(self.loss))))\n\n def end_log(self, n_batches):\n self.log[-1].div_(n_batches)\n\n def display_loss(self, batch):\n n_samples = batch + 1\n log = []\n for l, c in zip(self.loss, self.log[-1]):\n log.append('[{}: {:.4f}]'.format(l['type'], c / n_samples))\n\n return ''.join(log)\n\n def plot_loss(self, apath, epoch):\n axis = np.linspace(1, epoch, epoch)\n for i, l in enumerate(self.loss):\n label = '{} Loss'.format(l['type'])\n fig = plt.figure()\n plt.title(label)\n plt.plot(axis, self.log[:, i].numpy(), label=label)\n plt.legend()\n plt.xlabel('Epochs')\n plt.ylabel('Loss')\n plt.grid(True)\n plt.savefig('{}/loss_{}.pdf'.format(apath, l['type']))\n plt.close(fig)\n\n def get_loss_module(self):\n if self.n_GPUs == 1:\n return self.loss_module\n else:\n return self.loss_module.module\n\n def save(self, apath):\n torch.save(self.state_dict(), os.path.join(apath, 'loss.pt'))\n torch.save(self.log, os.path.join(apath, 'loss_log.pt'))\n\n def load(self, apath, cpu=False):\n if cpu:\n kwargs = {'map_location': lambda storage, loc: storage}\n else:\n kwargs = {}\n\n self.load_state_dict(torch.load(\n os.path.join(apath, 'loss.pt'),\n **kwargs\n ))\n self.log = torch.load(os.path.join(apath, 'loss_log.pt'))\n for l in self.loss_module:\n if hasattr(l, 'scheduler'):\n for _ in range(len(self.log)): l.scheduler.step()"
},
{
"path": "KDSR-classic/loss/adversarial.py",
"content": "import utility\nfrom model import common\nfrom loss import discriminator\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.optim as optim\nfrom torch.autograd import Variable\n\nclass Adversarial(nn.Module):\n def __init__(self, args, gan_type):\n super(Adversarial, self).__init__()\n self.gan_type = gan_type\n self.gan_k = args.gan_k\n self.discriminator = discriminator.Discriminator(args, gan_type)\n if gan_type != 'WGAN_GP':\n self.optimizer = utility.make_optimizer(args, self.discriminator)\n else:\n self.optimizer = optim.Adam(\n self.discriminator.parameters(),\n betas=(0, 0.9), eps=1e-8, lr=1e-5\n )\n self.scheduler = utility.make_scheduler(args, self.optimizer)\n\n def forward(self, fake, real):\n fake_detach = fake.detach()\n\n self.loss = 0\n for _ in range(self.gan_k):\n self.optimizer.zero_grad()\n d_fake = self.discriminator(fake_detach)\n d_real = self.discriminator(real)\n if self.gan_type == 'GAN':\n label_fake = torch.zeros_like(d_fake)\n label_real = torch.ones_like(d_real)\n loss_d \\\n = F.binary_cross_entropy_with_logits(d_fake, label_fake) \\\n + F.binary_cross_entropy_with_logits(d_real, label_real)\n elif self.gan_type.find('WGAN') >= 0:\n loss_d = (d_fake - d_real).mean()\n if self.gan_type.find('GP') >= 0:\n epsilon = torch.rand_like(fake).view(-1, 1, 1, 1)\n hat = fake_detach.mul(1 - epsilon) + real.mul(epsilon)\n hat.requires_grad = True\n d_hat = self.discriminator(hat)\n gradients = torch.autograd.grad(\n outputs=d_hat.sum(), inputs=hat,\n retain_graph=True, create_graph=True, only_inputs=True\n )[0]\n gradients = gradients.view(gradients.size(0), -1)\n gradient_norm = gradients.norm(2, dim=1)\n gradient_penalty = 10 * gradient_norm.sub(1).pow(2).mean()\n loss_d += gradient_penalty\n\n # Discriminator update\n self.loss += loss_d.item()\n loss_d.backward()\n self.optimizer.step()\n\n if self.gan_type == 'WGAN':\n for p in self.discriminator.parameters():\n p.data.clamp_(-1, 1)\n\n self.loss /= self.gan_k\n\n d_fake_for_g = self.discriminator(fake)\n if self.gan_type == 'GAN':\n loss_g = F.binary_cross_entropy_with_logits(\n d_fake_for_g, label_real\n )\n elif self.gan_type.find('WGAN') >= 0:\n loss_g = -d_fake_for_g.mean()\n\n # Generator loss\n return loss_g\n \n def state_dict(self, *args, **kwargs):\n state_discriminator = self.discriminator.state_dict(*args, **kwargs)\n state_optimizer = self.optimizer.state_dict()\n\n return dict(**state_discriminator, **state_optimizer)\n \n# Some references\n# https://github.com/kuc2477/pytorch-wgan-gp/blob/master/model.py\n# OR\n# https://github.com/caogang/wgan-gp/blob/master/gan_cifar10.py\n"
},
{
"path": "KDSR-classic/loss/discriminator.py",
"content": "from model import common\n\nimport torch.nn as nn\n\nclass Discriminator(nn.Module):\n def __init__(self, args, gan_type='GAN'):\n super(Discriminator, self).__init__()\n\n in_channels = 3\n out_channels = 64\n depth = 7\n #bn = not gan_type == 'WGAN_GP'\n bn = True\n act = nn.LeakyReLU(negative_slope=0.2, inplace=True)\n\n m_features = [\n common.BasicBlock(args.n_colors, out_channels, 3, bn=bn, act=act)\n ]\n for i in range(depth):\n in_channels = out_channels\n if i % 2 == 1:\n stride = 1\n out_channels *= 2\n else:\n stride = 2\n m_features.append(common.BasicBlock(\n in_channels, out_channels, 3, stride=stride, bn=bn, act=act\n ))\n\n self.features = nn.Sequential(*m_features)\n\n patch_size = args.patch_size // (2**((depth + 1) // 2))\n m_classifier = [\n nn.Linear(out_channels * patch_size**2, 1024),\n act,\n nn.Linear(1024, 1)\n ]\n self.classifier = nn.Sequential(*m_classifier)\n\n def forward(self, x):\n features = self.features(x)\n output = self.classifier(features.view(features.size(0), -1))\n\n return output\n\n"
},
{
"path": "KDSR-classic/loss/vgg.py",
"content": "from model import common\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torchvision.models as models\nfrom torch.autograd import Variable\n\nclass VGG(nn.Module):\n def __init__(self, conv_index, rgb_range=1):\n super(VGG, self).__init__()\n vgg_features = models.vgg19(pretrained=True).features\n modules = [m for m in vgg_features]\n if conv_index == '22':\n self.vgg = nn.Sequential(*modules[:8])\n elif conv_index == '54':\n self.vgg = nn.Sequential(*modules[:35])\n\n vgg_mean = (0.485, 0.456, 0.406)\n vgg_std = (0.229 * rgb_range, 0.224 * rgb_range, 0.225 * rgb_range)\n self.sub_mean = common.MeanShift(rgb_range, vgg_mean, vgg_std)\n self.vgg.requires_grad = False\n\n def forward(self, sr, hr):\n def _forward(x):\n x = self.sub_mean(x)\n x = self.vgg(x)\n return x\n \n vgg_sr = _forward(sr)\n with torch.no_grad():\n vgg_hr = _forward(hr.detach())\n\n loss = F.mse_loss(vgg_sr, vgg_hr)\n\n return loss\n"
},
{
"path": "KDSR-classic/main_anisonoise_KDSRsMx4_stage3.sh",
"content": "## noise-free degradations with isotropic Gaussian blurs\n# training knowledge distillation\nCUDA_VISIBLE_DEVICES=0,1,2,3 python3 main_anisonoise_stage3.py --dir_data='/root/datasets' \\\n --model='blindsr' \\\n --scale='4' \\\n --blur_type='aniso_gaussian' \\\n --noise=25.0 \\\n --lambda_min=0.2 \\\n --lambda_max=4.0 \\\n --lambda_1 4.0 \\\n --lambda_2 1.5 \\\n --theta 120 \\\n --n_GPUs=4 \\\n --epochs_encoder 0 \\\n --epochs_sr 500 \\\n --data_test Set14 \\\n --st_save_epoch 495 \\\n --data_train DF2K \\\n --save 'KDSRsMx4_anisonoise_stage3'\\\n --patch_size 64 \\\n --batch_size 64 \\\n\n\n\n"
},
{
"path": "KDSR-classic/main_anisonoise_KDSRsMx4_stage4.sh",
"content": "## noise-free degradations with isotropic Gaussian blurs\n# training knowledge distillation\nCUDA_VISIBLE_DEVICES=0,1,2,3 python3 main_anisonoise_stage4.py --dir_data='/root/datasets' \\\n --model='blindsr' \\\n --scale='4' \\\n --blur_type='aniso_gaussian' \\\n --noise=25.0 \\\n --lambda_min=0.2 \\\n --lambda_max=4.0 \\\n --lambda_1 4.0 \\\n --lambda_2 1.5 \\\n --theta 120 \\\n --n_GPUs=4 \\\n --epochs_encoder 100 \\\n --epochs_sr 500 \\\n --data_test Set14 \\\n --st_save_epoch 495 \\\n --data_train DF2K \\\n --save 'KDSRsMx4_anisonoise_stage4'\\\n --pre_train_TA=\"./experiment/KDSRsMx4_anisonoise_stage3_x4_bicubic_aniso/model/model_TA_last.pt\" \\\n --pre_train_ST=\"./experiment/KDSRsMx4_anisonoise_stage3_x4_bicubic_aniso/model/model_TA_last.pt\" \\\n --resume 0 \\\n --lr_encoder 2e-4 \\\n --patch_size 64 \\\n --batch_size 64 \\\n\n\n"
},
{
"path": "KDSR-classic/main_anisonoise_stage3.py",
"content": "from option import args\nimport torch\nimport utility\nimport data\nimport model_TA\nimport loss\nfrom trainer_anisonoise_stage3 import Trainer\n\n\ndef count_param(model):\n param_count = 0\n for param in model.parameters():\n param_count += param.view(-1).size()[0]\n return param_count\n\nif __name__ == '__main__':\n torch.manual_seed(args.seed)\n checkpoint = utility.checkpoint(args)\n if checkpoint.ok:\n loader = data.Data(args)\n model_TA = model_TA.Model(args, checkpoint)\n print(count_param(model_TA))\n loss = loss.Loss(args, checkpoint) if not args.test_only else None\n t = Trainer(args, loader, model_TA, loss, checkpoint)\n while not t.terminate():\n t.train()\n t.test()\n\n\n\n checkpoint.done()"
},
{
"path": "KDSR-classic/main_anisonoise_stage4.py",
"content": "from option import args\nimport torch\nimport utility\nimport data\nimport model_TA\nimport model_ST\nimport loss\nfrom trainer_anisonoise_stage4 import Trainer\n\n\n# def count_param(model):\n# param_count = 0\n# for param in model.parameters():\n# param_count += param.view(-1).size()[0]\n# return param_count\n\nif __name__ == '__main__':\n torch.manual_seed(args.seed)\n checkpoint = utility.checkpoint(args)\n if checkpoint.ok:\n loader = data.Data(args)\n model_TA = model_TA.Model(args, checkpoint)\n model_ST = model_ST.Model(args, checkpoint)\n loss = loss.Loss(args, checkpoint) if not args.test_only else None\n t = Trainer(args, loader, model_ST, model_TA, loss, checkpoint)\n while not t.terminate():\n t.train()\n t.test()\n\n\n\n checkpoint.done()"
},
{
"path": "KDSR-classic/main_iso_KDSRsLx4_stage3.sh",
"content": "## noise-free degradations with isotropic Gaussian blurs\n# training knowledge distillation\nCUDA_VISIBLE_DEVICES=0,1,2,3 python3 main_iso_stage3.py --dir_data='/root/datasets' \\\n --model='blindsr' \\\n --scale='4' \\\n --blur_type='iso_gaussian' \\\n --noise=0 \\\n --sig_min=0.2 \\\n --sig_max=4.0 \\\n --sig=3.6 \\\n --n_GPUs=4 \\\n --epochs_encoder 0 \\\n --epochs_sr 500 \\\n --data_test Set14 \\\n --st_save_epoch 495 \\\n --data_train DF2K \\\n --save 'KDSRsLx4_iso_stage3'\\\n --patch_size 64 \\\n --batch_size 64 \\\n --n_feats 128 \\\n --n_blocks 28 \\\n --n_resblocks 5\n\n\n"
},
{
"path": "KDSR-classic/main_iso_KDSRsLx4_stage4.sh",
"content": "## noise-free degradations with isotropic Gaussian blurs\n# training knowledge distillation\nCUDA_VISIBLE_DEVICES=0,1,2,3 python3 main_iso_stage4.py --dir_data='/root/datasets' \\\n --model='blindsr' \\\n --scale='4' \\\n --blur_type='iso_gaussian' \\\n --noise=0 \\\n --sig_min=0.2 \\\n --sig_max=4.0 \\\n --sig=3.6 \\\n --n_GPUs=4 \\\n --epochs_encoder 0 \\\n --epochs_sr 600 \\\n --lr_decay_sr 150 \\\n --data_test Set14 \\\n --st_save_epoch 0 \\\n --data_train DF2K \\\n --save 'KDSRsLx4_iso_stage4'\\\n --pre_train_TA=\"experiment/KDSRsLx4_iso_stage3_x4_bicubic_iso/model/model_TA_last.pt\" \\\n --pre_train_ST=\"experiment/KDSRsLx4_iso_stage3_x4_bicubic_iso/model/model_TA_last.pt\" \\\n --lr_encoder 2e-4 \\\n --patch_size 64 \\\n --batch_size 64 \\\n --resume 0 \\\n --n_feats 128 \\\n --n_blocks 28 \\\n --n_resblocks 5\n\n\n"
},
{
"path": "KDSR-classic/main_iso_KDSRsMx4_stage3.sh",
"content": "## noise-free degradations with isotropic Gaussian blurs\n# training knowledge distillation\nCUDA_VISIBLE_DEVICES=4,5,6,7 python3 main_iso_stage3.py --dir_data='/root/datasets' \\\n --model='blindsr' \\\n --scale='4' \\\n --blur_type='iso_gaussian' \\\n --noise=0 \\\n --sig_min=0.2 \\\n --sig_max=4.0 \\\n --sig=3.6 \\\n --n_GPUs=4 \\\n --epochs_encoder 0 \\\n --epochs_sr 500 \\\n --data_test Set14 \\\n --st_save_epoch 495 \\\n --data_train DF2K \\\n --save 'KDSRsMx4_iso_stage3'\\\n --patch_size 64 \\\n --batch_size 64 \\\n\n\n"
},
{
"path": "KDSR-classic/main_iso_KDSRsMx4_stage4.sh",
"content": "## noise-free degradations with isotropic Gaussian blurs\n# training knowledge distillation\nCUDA_VISIBLE_DEVICES=4,5,6,7 python3 main_iso_stage4.py --dir_data='/root/datasets' \\\n --model='blindsr' \\\n --scale='4' \\\n --blur_type='iso_gaussian' \\\n --noise=0 \\\n --sig_min=0.2 \\\n --sig_max=4.0 \\\n --sig=3.6 \\\n --n_GPUs=4 \\\n --epochs_encoder 0 \\\n --epochs_sr 600 \\\n --lr_decay_sr 150 \\\n --data_test Set14 \\\n --st_save_epoch 0 \\\n --data_train DF2K \\\n --save 'KDSRsMx4_iso_stage4'\\\n --pre_train_TA=\"experiment/KDSRsMx4_iso_stage3_x4_bicubic_iso/model/model_TA_last.pt\" \\\n --pre_train_ST=\"experiment/KDSRsMx4_iso_stage3_x4_bicubic_iso/model/model_TA_last.pt\" \\\n --lr_encoder 2e-4 \\\n --patch_size 64 \\\n --batch_size 64 \\\n --resume 0\n\n\n"
},
{
"path": "KDSR-classic/main_iso_stage3.py",
"content": "from option import args\nimport torch\nimport utility\nimport data\nimport model_TA\nimport loss\nfrom trainer_iso_stage3 import Trainer\n\n\ndef count_param(model):\n param_count = 0\n for param in model.parameters():\n param_count += param.view(-1).size()[0]\n return param_count\n\nif __name__ == '__main__':\n torch.manual_seed(args.seed)\n checkpoint = utility.checkpoint(args)\n if checkpoint.ok:\n loader = data.Data(args)\n model_TA = model_TA.Model(args, checkpoint)\n print(count_param(model_TA))\n loss = loss.Loss(args, checkpoint) if not args.test_only else None\n t = Trainer(args, loader, model_TA, loss, checkpoint)\n while not t.terminate():\n t.train()\n t.test()\n\n\n\n checkpoint.done()"
},
{
"path": "KDSR-classic/main_iso_stage4.py",
"content": "from option import args\nimport torch\nimport utility\nimport data\nimport model_TA\nimport model_ST\nimport loss\nfrom trainer_iso_stage4 import Trainer\n\n\n# def count_param(model):\n# param_count = 0\n# for param in model.parameters():\n# param_count += param.view(-1).size()[0]\n# return param_count\n\nif __name__ == '__main__':\n torch.manual_seed(args.seed)\n checkpoint = utility.checkpoint(args)\n if checkpoint.ok:\n loader = data.Data(args)\n model_TA = model_TA.Model(args, checkpoint)\n model_ST = model_ST.Model(args, checkpoint)\n loss = loss.Loss(args, checkpoint) if not args.test_only else None\n t = Trainer(args, loader, model_ST, model_TA, loss, checkpoint)\n while not t.terminate():\n t.train()\n t.test()\n\n\n\n checkpoint.done()"
},
{
"path": "KDSR-classic/model/__init__.py",
"content": "import os\nfrom importlib import import_module\n\nimport torch\nimport torch.nn as nn\n\n\nclass Model(nn.Module):\n def __init__(self, args, ckp):\n super(Model, self).__init__()\n print('Making model...')\n self.args = args\n self.scale = args.scale\n self.idx_scale = 0\n self.self_ensemble = args.self_ensemble\n self.chop = args.chop\n self.precision = args.precision\n self.cpu = args.cpu\n self.device = torch.device('cpu' if args.cpu else 'cuda')\n self.n_GPUs = args.n_GPUs\n self.save_models = args.save_models\n self.save = args.save\n\n module = import_module('model.'+args.model)\n self.model = module.make_model(args).to(self.device)\n if args.precision == 'half': self.model.half()\n\n if not args.cpu and args.n_GPUs > 1:\n self.model = nn.DataParallel(self.model, range(args.n_GPUs))\n\n self.load(\n ckp.dir,\n pre_train=args.pre_train,\n resume=args.resume,\n cpu=args.cpu\n )\n\n def forward(self, lr,lr_bic):\n if self.self_ensemble and not self.training:\n if self.chop:\n forward_function = self.forward_chop\n else:\n forward_function = self.model.forward\n\n return self.forward_x8(lr, forward_function)\n elif self.chop and not self.training:\n return self.forward_chop(lr)\n else:\n return self.model(lr,lr_bic)\n\n def get_model(self):\n if self.n_GPUs <= 1 or self.cpu:\n return self.model\n else:\n return self.model.module\n\n def state_dict(self, **kwargs):\n target = self.get_model()\n return target.state_dict(**kwargs)\n\n def save(self, apath, epoch, is_best=False):\n target = self.get_model()\n torch.save(\n target.state_dict(),\n os.path.join(apath, 'model', 'model_latest.pt')\n )\n if is_best:\n torch.save(\n target.state_dict(),\n os.path.join(apath, 'model', 'model_best.pt')\n )\n\n if self.save_models:\n torch.save(\n target.state_dict(),\n os.path.join(apath, 'model', 'model_{}.pt'.format(epoch))\n )\n\n def load(self, apath, pre_train='.', resume=-1, cpu=False):\n if cpu:\n kwargs = {'map_location': lambda storage, loc: storage}\n else:\n kwargs = {}\n\n if resume == -1:\n self.get_model().load_state_dict(\n torch.load(os.path.join(apath, 'model', 'model_latest.pt'), **kwargs),\n strict=True\n )\n\n elif resume == 0:\n if pre_train != '.':\n self.get_model().load_state_dict(\n torch.load(pre_train, **kwargs),\n strict=True\n )\n\n elif resume > 0:\n self.get_model().load_state_dict(\n torch.load(os.path.join(apath, 'model', 'model_{}.pt'.format(resume)), **kwargs),\n strict=False\n )\n\n def forward_chop(self, x, shave=10, min_size=160000):\n scale = self.scale[self.idx_scale]\n n_GPUs = min(self.n_GPUs, 4)\n b, c, h, w = x.size()\n h_half, w_half = h // 2, w // 2\n h_size, w_size = h_half + shave, w_half + shave\n lr_list = [\n x[:, :, 0:h_size, 0:w_size],\n x[:, :, 0:h_size, (w - w_size):w],\n x[:, :, (h - h_size):h, 0:w_size],\n x[:, :, (h - h_size):h, (w - w_size):w]]\n\n if w_size * h_size < min_size:\n sr_list = []\n for i in range(0, 4, n_GPUs):\n lr_batch = torch.cat(lr_list[i:(i + n_GPUs)], dim=0)\n sr_batch = self.model(lr_batch)\n sr_list.extend(sr_batch.chunk(n_GPUs, dim=0))\n else:\n sr_list = [\n self.forward_chop(patch, shave=shave, min_size=min_size) \\\n for patch in lr_list\n ]\n\n h, w = scale * h, scale * w\n h_half, w_half = scale * h_half, scale * w_half\n h_size, w_size = scale * h_size, scale * w_size\n shave *= scale\n\n output = x.new(b, c, h, w)\n output[:, :, 0:h_half, 0:w_half] \\\n = sr_list[0][:, :, 0:h_half, 0:w_half]\n output[:, :, 0:h_half, w_half:w] \\\n = sr_list[1][:, :, 0:h_half, (w_size - w + w_half):w_size]\n output[:, :, h_half:h, 0:w_half] \\\n = sr_list[2][:, :, (h_size - h + h_half):h_size, 0:w_half]\n output[:, :, h_half:h, w_half:w] \\\n = sr_list[3][:, :, (h_size - h + h_half):h_size, (w_size - w + w_half):w_size]\n\n return output\n\n def forward_x8(self, x, forward_function):\n def _transform(v, op):\n if self.precision != 'single': v = v.float()\n\n v2np = v.data.cpu().numpy()\n if op == 'v':\n tfnp = v2np[:, :, :, ::-1].copy()\n elif op == 'h':\n tfnp = v2np[:, :, ::-1, :].copy()\n elif op == 't':\n tfnp = v2np.transpose((0, 1, 3, 2)).copy()\n\n ret = torch.Tensor(tfnp).to(self.device)\n if self.precision == 'half': ret = ret.half()\n\n return ret\n\n lr_list = [x]\n for tf in 'v', 'h', 't':\n lr_list.extend([_transform(t, tf) for t in lr_list])\n\n sr_list = [forward_function(aug) for aug in lr_list]\n for i in range(len(sr_list)):\n if i > 3:\n sr_list[i] = _transform(sr_list[i], 't')\n if i % 4 > 1:\n sr_list[i] = _transform(sr_list[i], 'h')\n if (i % 4) % 2 == 1:\n sr_list[i] = _transform(sr_list[i], 'v')\n\n output_cat = torch.cat(sr_list, dim=0)\n output = output_cat.mean(dim=0, keepdim=True)\n\n return output\n\n"
},
{
"path": "KDSR-classic/model/blindsr.py",
"content": "import torch\nfrom torch import nn\nimport model.common as common\nimport torch.nn.functional as F\nfrom moco.builder import MoCo\n\n\ndef make_model(args):\n return CZSR(args)\n\n\nclass CZSR(nn.Module):\n def __init__(self, args, conv=common.default_conv):\n super(CZSR, self).__init__()\n\n n_feats = args.n_feats\n kernel_size = 3\n scale = args.scale[0]\n act = nn.LeakyReLU(0.1, True)\n self.head = nn.Sequential(\n nn.Conv2d(3, n_feats, kernel_size=3, padding=1),\n act\n )\n # m_body =[\n # common.ResBlock(\n # conv, n_feats, kernel_size, act=act, res_scale=args.res_scale),\n # common.ResBlock(\n # conv, n_feats, kernel_size, act=act, res_scale=args.res_scale),\n # common.ResBlock(\n # conv, n_feats, kernel_size, act=act, res_scale=args.res_scale),\n # common.ResBlock(\n # conv, n_feats, kernel_size, act=act, res_scale=args.res_scale),\n # ]\n m_body = [\n nn.Conv2d(n_feats, n_feats, kernel_size=3, padding=1),\n act,\n nn.Conv2d(n_feats, n_feats, kernel_size=3, padding=1),\n act,\n nn.Conv2d(n_feats, n_feats, kernel_size=3, padding=1),\n act,\n nn.Conv2d(n_feats, n_feats, kernel_size=3, padding=1),\n act,\n nn.Conv2d(n_feats, n_feats, kernel_size=3, padding=1),\n act,\n nn.Conv2d(n_feats, n_feats, kernel_size=3, padding=1),\n act\n ]\n m_tail = [\n common.Upsampler(conv, scale, n_feats, act=False),\n nn.Conv2d(\n n_feats, args.n_colors, kernel_size,\n padding=(kernel_size // 2)\n )\n ]\n self.tail = nn.Sequential(*m_tail)\n\n self.body = nn.Sequential(*m_body)\n\n\n def forward(self, lr,lr_bic):\n res = self.head(lr)\n res = self.body(res)\n res = self.tail(res)\n res +=lr_bic\n\n return res\n\n\n\n\n"
},
{
"path": "KDSR-classic/model/common.py",
"content": "import math\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\n\ndef default_conv(in_channels, out_channels, kernel_size, bias=True):\n return nn.Conv2d(in_channels, out_channels, kernel_size, padding=(kernel_size//2), bias=bias)\n\nclass ResBlock(nn.Module):\n def __init__(\n self, conv, n_feats, kernel_size,\n bias=True, bn=False, act=nn.LeakyReLU(0.1, inplace=True), res_scale=1):\n\n super(ResBlock, self).__init__()\n m = []\n for i in range(2):\n m.append(conv(n_feats, n_feats, kernel_size, bias=bias))\n if bn:\n m.append(nn.BatchNorm2d(n_feats))\n if i == 0:\n m.append(act)\n\n self.body = nn.Sequential(*m)\n # self.res_scale = res_scale\n\n def forward(self, x):\n res = self.body(x)\n res += x\n\n return res\n\nclass MeanShift(nn.Conv2d):\n def __init__(self, rgb_range, rgb_mean, rgb_std, sign=-1):\n super(MeanShift, self).__init__(3, 3, kernel_size=1)\n std = torch.Tensor(rgb_std)\n self.weight.data = torch.eye(3).view(3, 3, 1, 1)\n self.weight.data.div_(std.view(3, 1, 1, 1))\n self.bias.data = sign * rgb_range * torch.Tensor(rgb_mean)\n self.bias.data.div_(std)\n self.weight.requires_grad = False\n self.bias.requires_grad = False\n\n\nclass Upsampler(nn.Sequential):\n def __init__(self, conv, scale, n_feat, act=False, bias=True):\n m = []\n if (int(scale) & (int(scale) - 1)) == 0: # Is scale = 2^n?\n for _ in range(int(math.log(scale, 2))):\n m.append(conv(n_feat, 4 * n_feat, 3, bias))\n m.append(nn.PixelShuffle(2))\n if act: m.append(act())\n elif scale == 3:\n m.append(conv(n_feat, 9 * n_feat, 3, bias))\n m.append(nn.PixelShuffle(3))\n if act: m.append(act())\n else:\n raise NotImplementedError\n\n super(Upsampler, self).__init__(*m)\n\n"
},
{
"path": "KDSR-classic/model_ST/__init__.py",
"content": "import os\nfrom importlib import import_module\n\nimport torch\nimport torch.nn as nn\n\n\nclass Model(nn.Module):\n def __init__(self, args, ckp):\n super(Model, self).__init__()\n print('Making model...')\n self.args = args\n self.scale = args.scale\n self.idx_scale = 0\n self.self_ensemble = args.self_ensemble\n self.chop = args.chop\n self.precision = args.precision\n self.cpu = args.cpu\n self.device = torch.device('cpu' if args.cpu else 'cuda')\n self.n_GPUs = args.n_GPUs\n self.save_models = args.save_models\n self.save = args.save\n\n module = import_module('model_ST.'+args.model)\n self.model = module.make_model(args).to(self.device)\n if args.precision == 'half': self.model.half()\n\n if not args.cpu and args.n_GPUs > 1:\n self.model = nn.DataParallel(self.model, range(args.n_GPUs))\n\n self.load(\n ckp.dir,\n pre_train=args.pre_train_ST,\n resume=args.resume,\n cpu=args.cpu\n )\n\n def forward(self, x):\n if self.self_ensemble and not self.training:\n if self.chop:\n forward_function = self.forward_chop\n else:\n forward_function = self.model.forward\n\n return self.forward_x8(x, forward_function)\n elif self.chop and not self.training:\n return self.forward_chop(x)\n else:\n return self.model(x)\n\n def get_model(self):\n if self.n_GPUs <= 1 or self.cpu:\n return self.model\n else:\n return self.model.module\n\n def state_dict(self, **kwargs):\n target = self.get_model()\n return target.state_dict(**kwargs)\n\n def save(self, apath, epoch, is_best=False):\n target = self.get_model()\n torch.save(\n target.state_dict(),\n os.path.join(apath, 'model', 'model_ST_latest.pt')\n )\n if is_best:\n torch.save(\n target.state_dict(),\n os.path.join(apath, 'model', 'model_ST_best.pt')\n )\n\n if self.save_models:\n torch.save(\n target.state_dict(),\n os.path.join(apath, 'model', 'model_ST_{}.pt'.format(epoch))\n )\n\n def load(self, apath, pre_train='.', resume=-1, cpu=False):\n if cpu:\n kwargs = {'map_location': lambda storage, loc: storage}\n else:\n kwargs = {}\n\n if resume == -1:\n self.get_model().load_state_dict(\n torch.load(os.path.join(apath, 'model', 'model_ST_latest.pt'), **kwargs),\n strict=True\n )\n\n elif resume == 0:\n if pre_train != '.':\n self.get_model().load_state_dict(\n torch.load(pre_train, **kwargs),\n strict=False\n )\n\n elif resume > 0:\n self.get_model().load_state_dict(\n torch.load(os.path.join(apath, 'model', 'model_ST_{}.pt'.format(resume)), **kwargs),\n strict=True\n )\n\n def forward_chop(self, x, shave=10, min_size=160000):\n scale = self.scale[self.idx_scale]\n n_GPUs = min(self.n_GPUs, 4)\n b, c, h, w = x.size()\n h_half, w_half = h // 2, w // 2\n h_size, w_size = h_half + shave, w_half + shave\n lr_list = [\n x[:, :, 0:h_size, 0:w_size],\n x[:, :, 0:h_size, (w - w_size):w],\n x[:, :, (h - h_size):h, 0:w_size],\n x[:, :, (h - h_size):h, (w - w_size):w]]\n\n if w_size * h_size < min_size:\n sr_list = []\n for i in range(0, 4, n_GPUs):\n lr_batch = torch.cat(lr_list[i:(i + n_GPUs)], dim=0)\n sr_batch = self.model(lr_batch)\n sr_list.extend(sr_batch.chunk(n_GPUs, dim=0))\n else:\n sr_list = [\n self.forward_chop(patch, shave=shave, min_size=min_size) \\\n for patch in lr_list\n ]\n\n h, w = scale * h, scale * w\n h_half, w_half = scale * h_half, scale * w_half\n h_size, w_size = scale * h_size, scale * w_size\n shave *= scale\n\n output = x.new(b, c, h, w)\n output[:, :, 0:h_half, 0:w_half] \\\n = sr_list[0][:, :, 0:h_half, 0:w_half]\n output[:, :, 0:h_half, w_half:w] \\\n = sr_list[1][:, :, 0:h_half, (w_size - w + w_half):w_size]\n output[:, :, h_half:h, 0:w_half] \\\n = sr_list[2][:, :, (h_size - h + h_half):h_size, 0:w_half]\n output[:, :, h_half:h, w_half:w] \\\n = sr_list[3][:, :, (h_size - h + h_half):h_size, (w_size - w + w_half):w_size]\n\n return output\n\n def forward_x8(self, x, forward_function):\n def _transform(v, op):\n if self.precision != 'single': v = v.float()\n\n v2np = v.data.cpu().numpy()\n if op == 'v':\n tfnp = v2np[:, :, :, ::-1].copy()\n elif op == 'h':\n tfnp = v2np[:, :, ::-1, :].copy()\n elif op == 't':\n tfnp = v2np.transpose((0, 1, 3, 2)).copy()\n\n ret = torch.Tensor(tfnp).to(self.device)\n if self.precision == 'half': ret = ret.half()\n\n return ret\n\n lr_list = [x]\n for tf in 'v', 'h', 't':\n lr_list.extend([_transform(t, tf) for t in lr_list])\n\n sr_list = [forward_function(aug) for aug in lr_list]\n for i in range(len(sr_list)):\n if i > 3:\n sr_list[i] = _transform(sr_list[i], 't')\n if i % 4 > 1:\n sr_list[i] = _transform(sr_list[i], 'h')\n if (i % 4) % 2 == 1:\n sr_list[i] = _transform(sr_list[i], 'v')\n\n output_cat = torch.cat(sr_list, dim=0)\n output = output_cat.mean(dim=0, keepdim=True)\n\n return output\n\n"
},
{
"path": "KDSR-classic/model_ST/blindsr.py",
"content": "import torch\nfrom torch import nn\nimport model.common as common\nimport torch.nn.functional as F\n\n\ndef make_model(args):\n return BlindSR(args)\n\nclass IDR_DDC(nn.Module):\n def __init__(self, channels_in, channels_out, kernel_size, reduction):\n super(IDR_DDC, self).__init__()\n self.channels_out = channels_out\n self.channels_in = channels_in\n self.kernel_size = kernel_size\n\n self.kernel = nn.Sequential(\n nn.Linear(channels_in, channels_in, bias=False),\n nn.LeakyReLU(0.1, True),\n nn.Linear(channels_in, channels_in * self.kernel_size * self.kernel_size, bias=False)\n )\n\n def forward(self, x):\n '''\n :param x[0]: feature map: B * C * H * W\n :param x[1]: degradation representation: B * C\n '''\n b, c, h, w = x[0].size()\n\n # branch 1\n kernel = self.kernel(x[1]).view(-1, 1, self.kernel_size, self.kernel_size)\n out = F.conv2d(x[0].view(1, -1, h, w), kernel, groups=b*c, padding=(self.kernel_size-1)//2)\n out = out.view(b, -1, h, w)\n\n\n return out\n\n\n\nclass IDR_DCRB(nn.Module):\n def __init__(self, conv, n_feat, kernel_size, reduction):\n super(IDR_DCRB, self).__init__()\n\n self.da_conv1 = IDR_DDC(n_feat, n_feat, kernel_size, reduction)\n self.conv1 = conv(n_feat, n_feat, kernel_size)\n self.relu = nn.LeakyReLU(0.1, True)\n\n def forward(self, x):\n '''\n :param x[0]: feature map: B * C * H * W\n :param x[1]: degradation representation: B * C\n '''\n\n out = self.relu(self.da_conv1(x))\n out = self.conv1(out)\n out = out + x[0]\n\n return out\n\n\nclass DAG(nn.Module):\n def __init__(self, conv, n_feat, kernel_size, reduction, n_blocks):\n super(DAG, self).__init__()\n self.n_blocks = n_blocks\n modules_body = [\n IDR_DCRB(conv, n_feat, kernel_size, reduction) \\\n for _ in range(n_blocks)\n ]\n # modules_body.append(conv(n_feat, n_feat, kernel_size))\n\n self.body = nn.Sequential(*modules_body)\n\n def forward(self, x):\n '''\n :param x[0]: feature map: B * C * H * W\n :param x[1]: degradation representation: B * C\n '''\n res = x[0]\n for i in range(self.n_blocks):\n res = self.body[i]([res, x[1]])\n\n return res\n\n\nclass KDSR(nn.Module):\n def __init__(self, args, conv=common.default_conv):\n super(KDSR, self).__init__()\n n_blocks = args.n_blocks \n n_feats = args.n_feats\n kernel_size = 3\n reduction = 8\n scale = int(args.scale[0])\n\n # RGB mean for DIV2K\n rgb_mean = (0.4488, 0.4371, 0.4040)\n rgb_std = (1.0, 1.0, 1.0)\n self.sub_mean = common.MeanShift(args.rgb_range, rgb_mean, rgb_std)\n self.add_mean = common.MeanShift(args.rgb_range, rgb_mean, rgb_std, 1)\n\n # head module\n modules_head = [conv(3, n_feats, kernel_size)]\n self.head = nn.Sequential(*modules_head)\n\n \n\n # body\n modules_body = [\n DAG(common.default_conv, n_feats, kernel_size, reduction, n_blocks)\n ]\n modules_body.append(conv(n_feats, n_feats, kernel_size))\n self.body = nn.Sequential(*modules_body)\n\n # tail\n modules_tail = [common.Upsampler(conv, scale, n_feats, act=False),\n conv(n_feats, 3, kernel_size)]\n self.tail = nn.Sequential(*modules_tail)\n\n def forward(self, x, k_v):\n\n # sub mean\n x = self.sub_mean(x)\n\n # head\n x = self.head(x)\n\n # body\n res = x\n res = self.body[0]([res, k_v])\n res = self.body[-1](res)\n res = res + x\n\n # tail\n x = self.tail(res)\n\n # add mean\n x = self.add_mean(x)\n\n return x\n\n\n\nclass KD_IDE(nn.Module):\n def __init__(self,args):\n super(KD_IDE, self).__init__()\n n_feats = args.n_feats\n n_resblocks = args.n_resblocks\n E1=[nn.Conv2d(3, n_feats, kernel_size=3, padding=1),\n nn.LeakyReLU(0.1, True)]\n E2=[\n common.ResBlock(\n common.default_conv, n_feats, kernel_size=3\n ) for _ in range(n_resblocks)\n ]\n E3=[\n nn.Conv2d(n_feats, n_feats * 2, kernel_size=3, padding=1),\n nn.LeakyReLU(0.1, True),\n nn.Conv2d(n_feats * 2, n_feats * 2, kernel_size=3, padding=1),\n nn.LeakyReLU(0.1, True),\n nn.Conv2d(n_feats * 2, n_feats * 4, kernel_size=3, padding=1),\n nn.LeakyReLU(0.1, True),\n nn.AdaptiveAvgPool2d(1),\n ]\n E=E1+E2+E3\n self.E = nn.Sequential(\n *E\n )\n self.mlp = nn.Sequential(\n nn.Linear(n_feats * 4, n_feats * 4),\n nn.LeakyReLU(0.1, True),\n nn.Linear(n_feats * 4, n_feats * 4),\n nn.LeakyReLU(0.1, True)\n )\n\n # compress\n self.compress = nn.Sequential(\n nn.Linear(n_feats * 4, n_feats),\n nn.LeakyReLU(0.1, True)\n )\n\n def forward(self, x):\n fea = self.E(x).squeeze(-1).squeeze(-1)\n S_fea = []\n # for i in range(len(self.mlp)):\n # fea = self.mlp[i](fea)\n # if i==2:\n # T_fea.append(fea)\n fea1 = self.mlp(fea)\n fea = self.compress(fea1)\n S_fea.append(fea1)\n return fea,S_fea\n\n\nclass BlindSR(nn.Module):\n def __init__(self, args):\n super(BlindSR, self).__init__()\n\n # Generator\n self.G = KDSR(args)\n\n self.E_st = KD_IDE(args)\n\n\n def forward(self, x):\n if self.training:\n\n # degradation-aware represenetion learning\n deg_repre, S_fea = self.E_st(x)\n\n # degradation-aware SR\n sr = self.G(x, deg_repre)\n\n return sr, S_fea\n else:\n # degradation-aware represenetion learning\n deg_repre, _ = self.E_st(x)\n\n # degradation-aware SR\n sr = self.G(x, deg_repre)\n\n return sr\n\n"
},
{
"path": "KDSR-classic/model_ST/common.py",
"content": "import math\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\n\ndef default_conv(in_channels, out_channels, kernel_size, bias=True):\n return nn.Conv2d(in_channels, out_channels, kernel_size, padding=(kernel_size//2), bias=bias)\n\n\nclass MeanShift(nn.Conv2d):\n def __init__(self, rgb_range, rgb_mean, rgb_std, sign=-1):\n super(MeanShift, self).__init__(3, 3, kernel_size=1)\n std = torch.Tensor(rgb_std)\n self.weight.data = torch.eye(3).view(3, 3, 1, 1)\n self.weight.data.div_(std.view(3, 1, 1, 1))\n self.bias.data = sign * rgb_range * torch.Tensor(rgb_mean)\n self.bias.data.div_(std)\n self.weight.requires_grad = False\n self.bias.requires_grad = False\n\n\nclass Upsampler(nn.Sequential):\n def __init__(self, conv, scale, n_feat, act=False, bias=True):\n m = []\n if (scale & (scale - 1)) == 0: # Is scale = 2^n?\n for _ in range(int(math.log(scale, 2))):\n m.append(conv(n_feat, 4 * n_feat, 3, bias))\n m.append(nn.PixelShuffle(2))\n if act: m.append(act())\n elif scale == 3:\n m.append(conv(n_feat, 9 * n_feat, 3, bias))\n m.append(nn.PixelShuffle(3))\n if act: m.append(act())\n else:\n raise NotImplementedError\n\n super(Upsampler, self).__init__(*m)\n\n"
},
{
"path": "KDSR-classic/model_TA/__init__.py",
"content": "import os\nfrom importlib import import_module\n\nimport torch\nimport torch.nn as nn\n\n\nclass Model(nn.Module):\n def __init__(self, args, ckp):\n super(Model, self).__init__()\n print('Making model...')\n self.args = args\n self.scale = args.scale\n self.idx_scale = 0\n self.self_ensemble = args.self_ensemble\n self.chop = args.chop\n self.precision = args.precision\n self.cpu = args.cpu\n self.device = torch.device('cpu' if args.cpu else 'cuda')\n self.n_GPUs = args.n_GPUs\n self.save_models = args.save_models\n self.save = args.save\n\n module = import_module('model_TA.'+args.model)\n self.model = module.make_model(args).to(self.device)\n if args.precision == 'half': self.model.half()\n\n if not args.cpu and args.n_GPUs > 1:\n self.model = nn.DataParallel(self.model, range(args.n_GPUs))\n\n self.load(\n ckp.dir,\n pre_train=args.pre_train_TA,\n resume=args.resume,\n cpu=args.cpu\n )\n\n def forward(self, x, deg_repre):\n if self.self_ensemble and not self.training:\n if self.chop:\n forward_function = self.forward_chop\n else:\n forward_function = self.model.forward\n\n return self.forward_x8(x, forward_function)\n elif self.chop and not self.training:\n return self.forward_chop(x, deg_repre)\n else:\n return self.model(x, deg_repre)\n\n def get_model(self):\n if self.n_GPUs <= 1 or self.cpu:\n return self.model\n else:\n return self.model.module\n\n def state_dict(self, **kwargs):\n target = self.get_model()\n return target.state_dict(**kwargs)\n\n def save(self, apath, epoch, is_best=False):\n target = self.get_model()\n torch.save(\n target.state_dict(),\n os.path.join(apath, 'model', 'model_latest.pt')\n )\n if is_best:\n torch.save(\n target.state_dict(),\n os.path.join(apath, 'model', 'model_best.pt')\n )\n\n if self.save_models:\n torch.save(\n target.state_dict(),\n os.path.join(apath, 'model', 'model_{}.pt'.format(epoch))\n )\n\n def load(self, apath, pre_train='.', resume=-1, cpu=False):\n if cpu:\n kwargs = {'map_location': lambda storage, loc: storage}\n else:\n kwargs = {}\n\n if resume == -1:\n self.get_model().load_state_dict(\n torch.load(os.path.join(apath, 'model', 'model_latest.pt'), **kwargs),\n strict=True\n )\n\n elif resume == 0:\n if pre_train != '.':\n self.get_model().load_state_dict(\n torch.load(pre_train, **kwargs),\n strict=True\n )\n\n elif resume > 0:\n self.get_model().load_state_dict(\n torch.load(os.path.join(apath, 'model', 'model_{}.pt'.format(resume)), **kwargs),\n strict=False\n )\n\n def forward_chop(self, x, deg_repre, shave=10, min_size=160000):\n scale = self.scale[self.idx_scale]\n n_GPUs = min(self.n_GPUs, 4)\n b, c, h, w = x.size()\n h_half, w_half = h // 2, w // 2\n h_size, w_size = h_half + shave, w_half + shave\n lr_list = [\n x[:, :, 0:h_size, 0:w_size],\n x[:, :, 0:h_size, (w - w_size):w],\n x[:, :, (h - h_size):h, 0:w_size],\n x[:, :, (h - h_size):h, (w - w_size):w]]\n\n shave4=int(shave*scale)\n b, c4, h4, w4 = deg_repre.size()\n h4_half, w4_half = h4 // 2, w4 // 2\n h4_size, w4_size = h4_half + shave4, w4_half + shave4\n repre_list = [\n deg_repre[:, :, 0:h4_size, 0:w4_size],\n deg_repre[:, :, 0:h4_size, (w4 - w4_size):w4],\n deg_repre[:, :, (h4 - h4_size):h4, 0:w4_size],\n deg_repre[:, :, (h4 - h4_size):h4, (w4 - w4_size):w4]]\n\n if w_size * h_size < min_size:\n sr_list = []\n for i in range(0, 4, n_GPUs):\n lr_batch = torch.cat(lr_list[i:(i + n_GPUs)], dim=0)\n repre_batch = torch.cat(repre_list[i:(i + n_GPUs)], dim=0)\n sr_batch = self.model(lr_batch,repre_batch)\n sr_list.extend(sr_batch.chunk(n_GPUs, dim=0))\n else:\n sr_list = [\n self.forward_chop(lr_patch, repre_patch, shave=shave, min_size=min_size) \\\n for lr_patch, repre_patch in zip(lr_list,repre_list)\n ]\n\n h, w = scale * h, scale * w\n h_half, w_half = scale * h_half, scale * w_half\n h_size, w_size = scale * h_size, scale * w_size\n shave *= scale\n\n output = x.new(b, c, h, w)\n output[:, :, 0:h_half, 0:w_half] \\\n = sr_list[0][:, :, 0:h_half, 0:w_half]\n output[:, :, 0:h_half, w_half:w] \\\n = sr_list[1][:, :, 0:h_half, (w_size - w + w_half):w_size]\n output[:, :, h_half:h, 0:w_half] \\\n = sr_list[2][:, :, (h_size - h + h_half):h_size, 0:w_half]\n output[:, :, h_half:h, w_half:w] \\\n = sr_list[3][:, :, (h_size - h + h_half):h_size, (w_size - w + w_half):w_size]\n\n return output\n\n def forward_x8(self, x, forward_function):\n def _transform(v, op):\n if self.precision != 'single': v = v.float()\n\n v2np = v.data.cpu().numpy()\n if op == 'v':\n tfnp = v2np[:, :, :, ::-1].copy()\n elif op == 'h':\n tfnp = v2np[:, :, ::-1, :].copy()\n elif op == 't':\n tfnp = v2np.transpose((0, 1, 3, 2)).copy()\n\n ret = torch.Tensor(tfnp).to(self.device)\n if self.precision == 'half': ret = ret.half()\n\n return ret\n\n lr_list = [x]\n for tf in 'v', 'h', 't':\n lr_list.extend([_transform(t, tf) for t in lr_list])\n\n sr_list = [forward_function(aug) for aug in lr_list]\n for i in range(len(sr_list)):\n if i > 3:\n sr_list[i] = _transform(sr_list[i], 't')\n if i % 4 > 1:\n sr_list[i] = _transform(sr_list[i], 'h')\n if (i % 4) % 2 == 1:\n sr_list[i] = _transform(sr_list[i], 'v')\n\n output_cat = torch.cat(sr_list, dim=0)\n output = output_cat.mean(dim=0, keepdim=True)\n\n return output\n\n"
},
{
"path": "KDSR-classic/model_TA/blindsr.py",
"content": "import torch\nfrom torch import nn\nimport model.common as common\nimport torch.nn.functional as F\n\n\ndef make_model(args):\n return BlindSR(args)\n\nclass IDR_DDC(nn.Module):\n def __init__(self, channels_in, channels_out, kernel_size, reduction):\n super(IDR_DDC, self).__init__()\n self.channels_out = channels_out\n self.channels_in = channels_in\n self.kernel_size = kernel_size\n\n self.kernel = nn.Sequential(\n nn.Linear(channels_in, channels_in, bias=False),\n nn.LeakyReLU(0.1, True),\n nn.Linear(channels_in, channels_in * self.kernel_size * self.kernel_size, bias=False)\n )\n\n def forward(self, x):\n '''\n :param x[0]: feature map: B * C * H * W\n :param x[1]: degradation representation: B * C\n '''\n b, c, h, w = x[0].size()\n\n # branch 1\n kernel = self.kernel(x[1]).view(-1, 1, self.kernel_size, self.kernel_size)\n out = F.conv2d(x[0].view(1, -1, h, w), kernel, groups=b*c, padding=(self.kernel_size-1)//2)\n out = out.view(b, -1, h, w)\n\n\n return out\n\n\n\nclass IDR_DCRB(nn.Module):\n def __init__(self, conv, n_feat, kernel_size, reduction):\n super(IDR_DCRB, self).__init__()\n\n self.da_conv1 = IDR_DDC(n_feat, n_feat, kernel_size, reduction)\n self.conv1 = conv(n_feat, n_feat, kernel_size)\n self.relu = nn.LeakyReLU(0.1, True)\n\n def forward(self, x):\n '''\n :param x[0]: feature map: B * C * H * W\n :param x[1]: degradation representation: B * C\n '''\n\n out = self.relu(self.da_conv1(x))\n out = self.conv1(out)\n out = out + x[0]\n\n return out\n\n\nclass DAG(nn.Module):\n def __init__(self, conv, n_feat, kernel_size, reduction, n_blocks):\n super(DAG, self).__init__()\n self.n_blocks = n_blocks\n modules_body = [\n IDR_DCRB(conv, n_feat, kernel_size, reduction) \\\n for _ in range(n_blocks)\n ]\n # modules_body.append(conv(n_feat, n_feat, kernel_size))\n\n self.body = nn.Sequential(*modules_body)\n\n def forward(self, x):\n '''\n :param x[0]: feature map: B * C * H * W\n :param x[1]: degradation representation: B * C\n '''\n res = x[0]\n for i in range(self.n_blocks):\n res = self.body[i]([res, x[1]])\n\n return res\n\n\nclass KDSR(nn.Module):\n def __init__(self, args, conv=common.default_conv):\n super(KDSR, self).__init__()\n n_blocks = args.n_blocks \n n_feats = args.n_feats\n kernel_size = 3\n reduction = 8\n scale = int(args.scale[0])\n\n # RGB mean for DIV2K\n rgb_mean = (0.4488, 0.4371, 0.4040)\n rgb_std = (1.0, 1.0, 1.0)\n self.sub_mean = common.MeanShift(args.rgb_range, rgb_mean, rgb_std)\n self.add_mean = common.MeanShift(args.rgb_range, rgb_mean, rgb_std, 1)\n\n # head module\n modules_head = [conv(3, n_feats, kernel_size)]\n self.head = nn.Sequential(*modules_head)\n\n \n\n # body\n modules_body = [\n DAG(common.default_conv, n_feats, kernel_size, reduction, n_blocks)\n ]\n modules_body.append(conv(n_feats, n_feats, kernel_size))\n self.body = nn.Sequential(*modules_body)\n\n # tail\n modules_tail = [common.Upsampler(conv, scale, n_feats, act=False),\n conv(n_feats, 3, kernel_size)]\n self.tail = nn.Sequential(*modules_tail)\n\n def forward(self, x, k_v):\n\n # sub mean\n x = self.sub_mean(x)\n\n # head\n x = self.head(x)\n\n # body\n res = x\n res = self.body[0]([res, k_v])\n res = self.body[-1](res)\n res = res + x\n\n # tail\n x = self.tail(res)\n\n # add mean\n x = self.add_mean(x)\n\n return x\n\n\n\nclass KD_IDE(nn.Module):\n def __init__(self,args):\n super(KD_IDE, self).__init__()\n n_feats = args.n_feats\n n_resblocks = args.n_resblocks\n scale = int(args.scale[0])\n E1=[nn.Conv2d(3+scale*scale*3, n_feats, kernel_size=3, padding=1),\n nn.LeakyReLU(0.1, True)]\n E2=[\n common.ResBlock(\n common.default_conv, n_feats, kernel_size=3\n ) for _ in range(n_resblocks)\n ]\n E3=[\n nn.Conv2d(n_feats, n_feats * 2, kernel_size=3, padding=1),\n nn.LeakyReLU(0.1, True),\n nn.Conv2d(n_feats * 2, n_feats * 2, kernel_size=3, padding=1),\n nn.LeakyReLU(0.1, True),\n nn.Conv2d(n_feats * 2, n_feats * 4, kernel_size=3, padding=1),\n nn.LeakyReLU(0.1, True),\n nn.AdaptiveAvgPool2d(1),\n ]\n E=E1+E2+E3\n self.E = nn.Sequential(\n *E\n )\n self.mlp = nn.Sequential(\n nn.Linear(n_feats * 4, n_feats * 4),\n nn.LeakyReLU(0.1, True),\n nn.Linear(n_feats * 4, n_feats * 4),\n nn.LeakyReLU(0.1, True)\n )\n\n # compress\n self.compress = nn.Sequential(\n nn.Linear(n_feats * 4, n_feats),\n nn.LeakyReLU(0.1, True)\n )\n\n def forward(self, x):\n fea = self.E(x).squeeze(-1).squeeze(-1)\n T_fea = []\n fea1 = self.mlp(fea)\n fea = self.compress(fea1)\n T_fea.append(fea1)\n return fea,T_fea\n\n\nclass BlindSR(nn.Module):\n def __init__(self, args):\n super(BlindSR, self).__init__()\n\n # Generator\n self.G = KDSR(args)\n\n self.E = KD_IDE(args)\n\n\n def forward(self, x, deg_repre):\n if self.training:\n\n # degradation-aware represenetion learning\n deg_repre, T_fea = self.E(deg_repre)\n\n # degradation-aware SR\n sr = self.G(x, deg_repre)\n\n return sr, T_fea\n else:\n # degradation-aware represenetion learning\n deg_repre, _ = self.E(deg_repre)\n\n # degradation-aware SR\n sr = self.G(x, deg_repre)\n\n return sr\n"
},
{
"path": "KDSR-classic/model_TA/common.py",
"content": "import math\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\n\ndef default_conv(in_channels, out_channels, kernel_size, bias=True):\n return nn.Conv2d(in_channels, out_channels, kernel_size, padding=(kernel_size//2), bias=bias)\n\n\nclass MeanShift(nn.Conv2d):\n def __init__(self, rgb_range, rgb_mean, rgb_std, sign=-1):\n super(MeanShift, self).__init__(3, 3, kernel_size=1)\n std = torch.Tensor(rgb_std)\n self.weight.data = torch.eye(3).view(3, 3, 1, 1)\n self.weight.data.div_(std.view(3, 1, 1, 1))\n self.bias.data = sign * rgb_range * torch.Tensor(rgb_mean)\n self.bias.data.div_(std)\n self.weight.requires_grad = False\n self.bias.requires_grad = False\n\n\nclass Upsampler(nn.Sequential):\n def __init__(self, conv, scale, n_feat, act=False, bias=True):\n m = []\n if (scale & (scale - 1)) == 0: # Is scale = 2^n?\n for _ in range(int(math.log(scale, 2))):\n m.append(conv(n_feat, 4 * n_feat, 3, bias))\n m.append(nn.PixelShuffle(2))\n if act: m.append(act())\n elif scale == 3:\n m.append(conv(n_feat, 9 * n_feat, 3, bias))\n m.append(nn.PixelShuffle(3))\n if act: m.append(act())\n else:\n raise NotImplementedError\n\n super(Upsampler, self).__init__(*m)\n\n"
},
{
"path": "KDSR-classic/option.py",
"content": "import argparse\nimport template\n\nparser = argparse.ArgumentParser(description='EDSR and MDSR')\n\nparser.add_argument('--debug', action='store_true',\n help='Enables debug mode')\nparser.add_argument('--template', default='.',\n help='You can set various templates in option.py')\n\n# Hardware specifications\nparser.add_argument('--n_threads', type=int, default=4,\n help='number of threads for data loading')\nparser.add_argument('--cpu', type=bool, default=False,\n help='use cpu only')\nparser.add_argument('--n_GPUs', type=int, default=2,\n help='number of GPUs')\nparser.add_argument('--seed', type=int, default=1,\n help='random seed')\nparser.add_argument('--pre_train_meta', type=str, default= '.',\n help='pre-trained model directory')\nparser.add_argument('--pre_train_TA', type=str, default= '.',\n help='pre-trained model directory')\nparser.add_argument('--pre_train_ST', type=str, default= '.',\n help='pre-trained model directory')\nparser.add_argument('--is_stage3', action='store_true',\n help='set this option to test the model')\nparser.add_argument('--temperature', type=float, default=0.15,\n help='for konwledge distillation')\n\n# Meta-Learning\nparser.add_argument('--task_iter', type=int, default=20,\n help='each task iteration times')\nparser.add_argument('--test_iter', type=int, default=20,\n help='each task iteration times')\nparser.add_argument('--meta_batch_size', type=int, default=8,\n help='each task iteration times')\nparser.add_argument('--task_batch_size', type=int, default=16,\n help='each task iteration times')\nparser.add_argument('--lr_task', type=float, default=1e-3,\n help='learning rate to train the whole network')\n\n\n# Data specifications\nparser.add_argument('--dir_data', type=str, default='D:/LongguangWang/Data',\n help='dataset directory')\nparser.add_argument('--dir_demo', type=str, default='../test',\n help='demo image directory')\nparser.add_argument('--data_train', type=str, default='DF2K',\n help='train dataset name')\nparser.add_argument('--data_test', type=str, default='Set5',\n help='test dataset name')\nparser.add_argument('--data_range', type=str, default='1-3450/801-810',\n help='train/test data range')\nparser.add_argument('--ext', type=str, default='sep',\n help='dataset file extension')\nparser.add_argument('--scale', type=str, default='4',\n help='super resolution scale')\nparser.add_argument('--patch_size', type=int, default=36,\n help='output patch size')\nparser.add_argument('--rgb_range', type=int, default=1,\n help='maximum value of RGB')\nparser.add_argument('--n_colors', type=int, default=3,\n help='number of color channels to use')\nparser.add_argument('--chop', action='store_true',\n help='enable memory-efficient forward')\nparser.add_argument('--no_augment', action='store_true',\n help='do not use data augmentation')\n\n# Degradation specifications\nparser.add_argument('--blur_kernel', type=int, default=21,\n help='size of blur kernels')\nparser.add_argument('--blur_type', type=str, default='iso_gaussian',\n help='blur types (iso_gaussian | aniso_gaussian)')\nparser.add_argument('--mode', type=str, default='bicubic',\n help='downsampler (bicubic | s-fold)')\nparser.add_argument('--noise', type=float, default=0.0,\n help='noise level')\n## isotropic Gaussian blur\nparser.add_argument('--sig_min', type=float, default=0.2,\n help='minimum sigma of isotropic Gaussian blurs')\nparser.add_argument('--sig_max', type=float, default=4.0,\n help='maximum sigma of isotropic Gaussian blurs')\nparser.add_argument('--sig', type=float, default=4.0,\n help='specific sigma of isotropic Gaussian blurs')\n## anisotropic Gaussian blur\nparser.add_argument('--lambda_min', type=float, default=0.2,\n help='minimum value for the eigenvalue of anisotropic Gaussian blurs')\nparser.add_argument('--lambda_max', type=float, default=4.0,\n help='maximum value for the eigenvalue of anisotropic Gaussian blurs')\nparser.add_argument('--lambda_1', type=float, default=0.2,\n help='one eigenvalue of anisotropic Gaussian blurs')\nparser.add_argument('--lambda_2', type=float, default=4.0,\n help='another eigenvalue of anisotropic Gaussian blurs')\nparser.add_argument('--theta', type=float, default=0.0,\n help='rotation angle of anisotropic Gaussian blurs [0, 180]')\n\n\n# Model specifications\nparser.add_argument('--model', default='blindsr',\n help='model name')\nparser.add_argument('--pre_train', type=str, default= '.',\n help='pre-trained model directory')\nparser.add_argument('--extend', type=str, default='.',\n help='pre-trained model directory')\nparser.add_argument('--shift_mean', default=True,\n help='subtract pixel mean from the input')\nparser.add_argument('--dilation', action='store_true',\n help='use dilated convolution')\nparser.add_argument('--precision', type=str, default='single',\n choices=('single', 'half'),\n help='FP precision for test (single | half)')\nparser.add_argument('--n_resblocks', type=int, default=9,\n help='number of residual blocks')\nparser.add_argument('--n_feats', type=int, default=64,\n help='number of feature maps')\nparser.add_argument('--res_scale', type=float, default=1,\n help='residual scaling')\nparser.add_argument('--n_blocks', type=int, default=53,\n help='number of blocks in KDSR')\n\n# Training specifications\nparser.add_argument('--reset', action='store_true',\n help='reset the training')\nparser.add_argument('--test_every', type=int, default=1000,\n help='do test per every N batches')\nparser.add_argument('--epochs_encoder', type=int, default=100,\n help='number of epochs to train the degradation encoder')\nparser.add_argument('--epochs_sr', type=int, default=500,\n help='number of epochs to train the whole network')\nparser.add_argument('--st_save_epoch', type=int, default=550,\n help='number of epochs to save network')\nparser.add_argument('--batch_size', type=int, default=32,\n help='input batch size for training')\nparser.add_argument('--split_batch', type=int, default=1,\n help='split the batch into smaller chunks')\nparser.add_argument('--self_ensemble', action='store_true',\n help='use self-ensemble method for test')\nparser.add_argument('--test_only', action='store_true',\n help='set this option to test the model')\n\n# Optimization specifications\nparser.add_argument('--lr_encoder', type=float, default=1e-3,\n help='learning rate to train the degradation encoder')\nparser.add_argument('--lr_sr', type=float, default=1e-4,\n help='learning rate to train the whole network')\nparser.add_argument('--lr_decay_encoder', type=int, default=60,\n help='learning rate decay per N epochs')\nparser.add_argument('--lr_decay_sr', type=int, default=125,\n help='learning rate decay per N epochs')\nparser.add_argument('--decay_type', type=str, default='step',\n help='learning rate decay type')\nparser.add_argument('--gamma_encoder', type=float, default=0.1,\n help='learning rate decay factor for step decay')\nparser.add_argument('--gamma_sr', type=float, default=0.5,\n help='learning rate decay factor for step decay')\nparser.add_argument('--optimizer', default='ADAM',\n choices=('SGD', 'ADAM', 'RMSprop'),\n help='optimizer to use (SGD | ADAM | RMSprop)')\nparser.add_argument('--momentum', type=float, default=0.9,\n help='SGD momentum')\nparser.add_argument('--beta1', type=float, default=0.9,\n help='ADAM beta1')\nparser.add_argument('--beta2', type=float, default=0.999,\n help='ADAM beta2')\nparser.add_argument('--epsilon', type=float, default=1e-8,\n help='ADAM epsilon for numerical stability')\nparser.add_argument('--weight_decay', type=float, default=0,\n help='weight decay')\nparser.add_argument('--start_epoch', type=int, default=0,\n help='resume from the snapshot, and the start_epoch')\n\n# Loss specifications\nparser.add_argument('--loss', type=str, default='1*L1',\n help='loss function configuration')\nparser.add_argument('--skip_threshold', type=float, default='1e6',\n help='skipping batch that has large error')\n\n# Log specifications\nparser.add_argument('--save', type=str, default='blindsr',\n help='file name to save')\nparser.add_argument('--load', type=str, default='.',\n help='file name to load')\nparser.add_argument('--resume', type=int, default=0,\n help='resume from specific checkpoint')\nparser.add_argument('--save_models', action='store_true',\n help='save all intermediate models')\nparser.add_argument('--print_every', type=int, default=200,\n help='how many batches to wait before logging training status')\nparser.add_argument('--save_results', default=False,\n help='save output results')\n\nargs = parser.parse_args()\ntemplate.set_template(args)\n\nargs.scale = list(map(lambda x: int(x), args.scale.split('+')))\nargs.data_train = args.data_train.split('+')\nargs.data_test = args.data_test.split('+')\n\n\nfor arg in vars(args):\n if vars(args)[arg] == 'True':\n vars(args)[arg] = True\n elif vars(args)[arg] == 'False':\n vars(args)[arg] = False\n\n"
},
{
"path": "KDSR-classic/template.py",
"content": "def set_template(args):\n # Set the templates here\n if args.template.find('jpeg') >= 0:\n args.data_train = 'DIV2K_jpeg'\n args.data_test = 'DIV2K_jpeg'\n args.epochs = 200\n args.lr_decay = 100\n\n if args.template.find('EDSR_paper') >= 0:\n args.model = 'EDSR'\n args.n_resblocks = 32\n args.n_feats = 256\n args.res_scale = 0.1\n\n if args.template.find('MDSR') >= 0:\n args.model = 'MDSR'\n args.patch_size = 48\n args.epochs = 650\n\n if args.template.find('DDBPN') >= 0:\n args.model = 'DDBPN'\n args.patch_size = 128\n args.scale = '4'\n\n args.data_test = 'Set5'\n\n args.batch_size = 20\n args.epochs = 1000\n args.lr_decay = 500\n args.gamma = 0.1\n args.weight_decay = 1e-4\n\n args.loss = '1*MSE'\n\n if args.template.find('GAN') >= 0:\n args.epochs = 200\n args.lr = 5e-5\n args.lr_decay = 150\n\n if args.template.find('RCAN') >= 0:\n args.model = 'RCAN'\n args.n_resgroups = 10\n args.n_resblocks = 20\n args.n_feats = 64\n args.chop = True\n\n"
},
{
"path": "KDSR-classic/test.py",
"content": "from option import args\nimport torch\nimport utility\nimport data\nimport model\nimport loss\nfrom trainer import Trainer\n\n\nif __name__ == '__main__':\n torch.manual_seed(args.seed)\n checkpoint = utility.checkpoint(args)\n\n if checkpoint.ok:\n loader = data.Data(args)\n model = model.Model(args, checkpoint)\n loss = loss.Loss(args, checkpoint) if not args.test_only else None\n t = Trainer(args, loader, model, loss, checkpoint)\n while not t.terminate():\n t.test()\n\n checkpoint.done()\n"
},
{
"path": "KDSR-classic/test_anisonoise_KDSRsMx4.sh",
"content": "CUDA_VISIBLE_DEVICES=4 python3 main_anisonoise_stage4.py --dir_data='/root/datasets' \\\n --model='blindsr' \\\n --scale='4' \\\n --blur_type='aniso_gaussian' \\\n --lambda_1 3.5 \\\n --lambda_2 1.5 \\\n --theta 30 \\\n --n_GPUs=1 \\\n --data_test Set14 \\\n --save 'KDSRsM_anisonoise_test'\\\n --pre_train_ST=\"./experiment/KDSRsM_anisonoise_x4.pt\" \\\n --resume 0 \\\n --test_only \\\n --save_results False\n\n"
},
{
"path": "KDSR-classic/test_iso_KDSRsLx4.sh",
"content": "CUDA_VISIBLE_DEVICES=4 python3 main_iso_stage4.py --test_only \\\n --dir_data='/root/datasets' \\\n --data_test='Set5+Set14+B100+Urban100+MANGA109' \\\n --model='blindsr' \\\n --scale='4' \\\n --resume=0 \\\n --pre_train_ST 'experiment/KDSRsL_iso_x4.pt' \\\n --n_GPUs=1 \\\n --save 'KDSRsL_iso_test'\\\n --blur_type='iso_gaussian' \\\n --noise 0 \\\n --save_results False \\\n --n_feats 128 \\\n --n_blocks 28 \\\n --n_resblocks 5\n "
},
{
"path": "KDSR-classic/test_iso_KDSRsMx4.sh",
"content": "CUDA_VISIBLE_DEVICES=4 python3 main_iso_stage4.py --test_only \\\n --dir_data='/root/datasets' \\\n --data_test='Set5+Set14+B100+Urban100+MANGA109' \\\n --model='blindsr' \\\n --scale='4' \\\n --resume=0 \\\n --pre_train_ST 'experiment/KDSRsM_iso_x4.pt' \\\n --n_GPUs=1 \\\n --save 'KDSRsM_iso_test'\\\n --blur_type='iso_gaussian' \\\n --noise 0 \\\n --save_results False \\\n\n\n "
},
{
"path": "KDSR-classic/trainer_anisonoise_stage3.py",
"content": "import os\nimport utility2\nimport torch\nfrom decimal import Decimal\nimport torch.nn.functional as F\nfrom utils import util\nfrom utils import util2\nfrom collections import OrderedDict\nimport random\nimport numpy as np\nimport torch.nn as nn\nimport utility3\n\n\nclass Trainer():\n def __init__(self, args, loader, model_TA, my_loss, ckp):\n self.test_res_psnr = []\n self.test_res_ssim = []\n self.args = args\n self.scale = args.scale\n self.loss1= nn.L1Loss()\n self.ckp = ckp\n self.loader_train = loader.loader_train\n self.loader_test = loader.loader_test\n self.model_TA = model_TA\n self.loss = my_loss\n self.optimizer = utility2.make_optimizer(args, self.model_TA)\n self.scheduler = utility2.make_scheduler(args, self.optimizer)\n self.pixel_unshuffle = nn.PixelUnshuffle(self.scale[0])\n if self.args.load != '.':\n self.optimizer.load_state_dict(\n torch.load(os.path.join(ckp.dir, 'optimizer.pt'))\n )\n for _ in range(len(ckp.log)): self.scheduler.step()\n\n def train(self):\n self.scheduler.step()\n self.loss.step()\n epoch = self.scheduler.last_epoch + 1\n\n lr = self.args.lr_sr * (self.args.gamma_sr ** ((epoch - self.args.epochs_encoder) // self.args.lr_decay_sr))\n for param_group in self.optimizer.param_groups:\n param_group['lr'] = lr\n\n self.ckp.write_log('[Epoch {}]\\tLearning rate: {:.2e}'.format(epoch, Decimal(lr)))\n self.loss.start_log()\n self.model_TA.train()\n\n degrade = util2.SRMDPreprocessing(\n self.scale[0],\n kernel_size=self.args.blur_kernel,\n blur_type=self.args.blur_type,\n sig_min=self.args.sig_min,\n sig_max=self.args.sig_max,\n lambda_min=self.args.lambda_min,\n lambda_max=self.args.lambda_max,\n noise=self.args.noise\n )\n\n timer = utility2.timer()\n\n for batch, ( hr, _,) in enumerate(self.loader_train):\n hr = hr.cuda() # b, c, h, w\n timer.tic()\n loss_all = 0\n lr_blur, hr_blur = degrade(hr) # b, c, h, w\n hr2 = self.pixel_unshuffle(hr)\n sr, _ = self.model_TA(lr_blur, torch.cat([lr_blur,hr2], dim=1))\n # sr,_ = self.model_TA(lr_blur, torch.cat([hr_blur,hr],dim=1))\n loss_all += self.loss1(sr,hr)\n self.optimizer.zero_grad()\n loss_all.backward()\n self.optimizer.step()\n\n # Remove the hooks before next training phase\n timer.hold()\n\n if (batch + 1) % self.args.print_every == 0:\n self.ckp.write_log(\n 'Epoch: [{:04d}][{:04d}/{:04d}]\\t'\n 'Loss [SR loss:{:.3f}]\\t'\n 'Time [{:.1f}s]'.format(\n epoch, (batch + 1) * self.args.batch_size, len(self.loader_train.dataset),\n loss_all.item(),\n timer.release(),\n ))\n\n self.loss.end_log(len(self.loader_train))\n\n # save model\n if epoch > self.args.st_save_epoch or (epoch %20 ==0):\n target = self.model_TA.get_model()\n model_dict = target.state_dict()\n torch.save(\n model_dict,\n os.path.join(self.ckp.dir, 'model', 'model_TA_{}.pt'.format(epoch))\n )\n\n optimzer_dict = self.optimizer.state_dict()\n torch.save(\n optimzer_dict,\n os.path.join(self.ckp.dir, 'optimzer', 'optimzer_TA_{}.pt'.format(epoch))\n )\n\n target = self.model_TA.get_model()\n model_dict = target.state_dict()\n torch.save(\n model_dict,\n os.path.join(self.ckp.dir, 'model', 'model_TA_last.pt')\n )\n optimzer_dict = self.optimizer.state_dict()\n torch.save(\n optimzer_dict,\n os.path.join(self.ckp.dir, 'optimzer', 'optimzer_TA_last.pt')\n )\n\n\n def test(self):\n self.ckp.write_log('\\nEvaluation:')\n self.ckp.add_log(torch.zeros(1, len(self.scale)))\n\n timer_test = utility2.timer()\n self.model_TA.eval()\n\n degrade = util2.SRMDPreprocessing(\n self.scale[0],\n kernel_size=self.args.blur_kernel,\n blur_type=self.args.blur_type,\n sig=self.args.sig,\n lambda_1=self.args.lambda_1,\n lambda_2=self.args.lambda_2,\n theta=self.args.theta,\n noise=10\n )\n\n for idx_data, d in enumerate(self.loader_test):\n for idx_scale, scale in enumerate(self.scale):\n d.dataset.set_scale(idx_scale)\n eval_psnr = 0\n eval_ssim = 0\n for idx, (hr, filename) in enumerate(d):\n hr = hr.cuda() # b, c, h, w\n hr = self.crop_border(hr, scale)\n # inference\n timer_test.tic()\n lr_blur, hr_blur = degrade(hr, random=False)\n hr2 = self.pixel_unshuffle(hr)\n sr = self.model_TA(lr_blur, torch.cat([lr_blur, hr2], dim=1))\n # sr = self.model_TA(lr_blur, torch.cat([hr_blur,hr],dim=1))\n\n timer_test.hold()\n\n sr = utility2.quantize(sr, self.args.rgb_range)\n hr = utility2.quantize(hr, self.args.rgb_range)\n\n # metrics\n eval_psnr += utility2.calc_psnr(\n sr, hr, scale, self.args.rgb_range,\n benchmark=d.dataset.benchmark\n )\n # eval_psnr += utility3.calc_psnr(\n # sr, hr, scale\n # )\n eval_ssim += utility2.calc_ssim(\n (sr*255).round().clamp(0,255), (hr*255).round().clamp(0,255),scale,\n benchmark=d.dataset.benchmark\n )\n\n # save results\n if self.args.save_results:\n save_list = [sr]\n filename = filename[0]\n self.ckp.save_results(filename, save_list, scale)\n\n if len(self.test_res_psnr)>10:\n self.test_res_psnr.pop(0)\n self.test_res_ssim.pop(0)\n self.test_res_psnr.append(eval_psnr / len(d))\n self.test_res_ssim.append(eval_ssim / len(d))\n\n self.ckp.log[-1, idx_scale] = eval_psnr / len(d)\n self.ckp.write_log(\n '[Epoch {}---{} x{}]\\tPSNR: {:.3f} SSIM: {:.4f} mean_PSNR: {:.3f} mean_SSIM: {:.4f}'.format(\n self.args.resume,\n self.args.data_test,\n scale,\n eval_psnr / len(d),\n eval_ssim / len(d),\n np.mean(self.test_res_psnr),\n np.mean(self.test_res_ssim)\n ))\n\n def crop_border(self, img_hr, scale):\n b, c, h, w = img_hr.size()\n\n img_hr = img_hr[:, :, :int(h//scale*scale), :int(w//scale*scale)]\n\n return img_hr\n\n def get_patch(self, img, patch_size=48, scale=4):\n tb, tc, th, tw = img.shape ## HR image\n tp = round(scale * patch_size)\n tx = random.randrange(0, (tw - tp))\n ty = random.randrange(0, (th - tp))\n\n return img[:,:,ty:ty + tp, tx:tx + tp]\n\n def crop(self, img_hr):\n # b, c, h, w = img_hr.size()\n tp_hr = []\n for i in range(self.task_batch_size):\n tp_hr.append(self.get_patch(img_hr,self.args.patch_size,self.scale[0]))\n tp_hr = torch.cat(tp_hr,dim=0)\n return tp_hr\n\n def terminate(self):\n if self.args.test_only:\n self.test()\n return True\n else:\n epoch = self.scheduler.last_epoch + 1\n return epoch >= self.args.epochs_sr"
},
{
"path": "KDSR-classic/trainer_anisonoise_stage4.py",
"content": "import os\nimport utility\nimport torch\nfrom decimal import Decimal\nimport torch.nn.functional as F\nfrom utils import util2\nimport numpy as np\nimport torch.nn as nn\n\n\nclass Trainer():\n def __init__(self, args, loader, model_ST, model_TA, my_loss, ckp):\n self.is_first =True\n self.args = args\n self.scale = args.scale\n self.test_res_psnr = []\n self.test_res_ssim = []\n self.ckp = ckp\n self.loader_train = loader.loader_train\n self.loader_test = loader.loader_test\n self.model_ST = model_ST\n self.model_TA = model_TA\n self.model_Est = torch.nn.DataParallel(self.model_ST.get_model().E_st, range(self.args.n_GPUs))\n self.model_Eta = torch.nn.DataParallel(self.model_TA.get_model().E, range(self.args.n_GPUs))\n self.loss1 = nn.L1Loss()\n self.loss = my_loss\n self.optimizer = utility.make_optimizer(args, self.model_ST)\n self.scheduler = utility.make_scheduler(args, self.optimizer)\n self.temperature = args.temperature\n self.pixel_unshuffle = nn.PixelUnshuffle(self.scale[0])\n if self.args.load != '.':\n self.optimizer.load_state_dict(\n torch.load(os.path.join(ckp.dir, 'optimizer.pt'))\n )\n for _ in range(len(ckp.log)): self.scheduler.step()\n\n def train(self):\n self.scheduler.step()\n self.loss.step()\n epoch = self.scheduler.last_epoch + 1\n\n if epoch <= self.args.epochs_encoder:\n lr = self.args.lr_encoder * (self.args.gamma_encoder ** (epoch // self.args.lr_decay_encoder))\n for param_group in self.optimizer.param_groups:\n param_group['lr'] = lr\n else:\n lr = self.args.lr_sr * (self.args.gamma_sr ** ((epoch - self.args.epochs_encoder) // self.args.lr_decay_sr))\n for param_group in self.optimizer.param_groups:\n param_group['lr'] = lr\n\n self.ckp.write_log('[Epoch {}]\\tLearning rate: {:.2e}'.format(epoch, Decimal(lr)))\n self.loss.start_log()\n self.model_ST.train()\n\n degrade = util2.SRMDPreprocessing(\n self.scale[0],\n kernel_size=self.args.blur_kernel,\n blur_type=self.args.blur_type,\n sig_min=self.args.sig_min,\n sig_max=self.args.sig_max,\n lambda_min=self.args.lambda_min,\n lambda_max=self.args.lambda_max,\n noise=self.args.noise\n )\n\n timer = utility.timer()\n losses_sr, losses_distill_distribution, losses_distill_abs = utility.AverageMeter(),utility.AverageMeter(),utility.AverageMeter()\n\n\n for batch, (hr, _,) in enumerate(self.loader_train):\n hr = hr.cuda() # b, c, h, w\n timer.tic()\n lr_blur, hr_blur = degrade(hr) # b, c, h, w\n hr2 = self.pixel_unshuffle(hr)\n _, T_fea = self.model_Eta(torch.cat([lr_blur,hr2],dim=1))\n\n loss_distill_dis = 0\n loss_distill_abs = 0\n\n if epoch <= self.args.epochs_encoder:\n _, S_fea = self.model_Est(lr_blur)\n for i in range(len(T_fea)):\n student_distance = F.log_softmax(S_fea[i] / self.temperature, dim=1)\n teacher_distance = F.softmax(T_fea[i].detach()/ self.temperature, dim=1)\n loss_distill_dis += F.kl_div(\n student_distance, teacher_distance, reduction='batchmean')\n loss_distill_abs += nn.L1Loss()(S_fea[i], T_fea[i].detach())\n losses_distill_distribution.update(loss_distill_dis.item())\n #losses_distill_distribution.update(loss_distill_dis)\n losses_distill_abs.update(loss_distill_abs.item())\n loss = loss_distill_dis #+ 0.1* loss_distill_abs\n else:\n sr, S_fea = self.model_ST(lr_blur)\n loss_SR = self.loss(sr, hr)\n for i in range(len(T_fea)):\n student_distance = F.log_softmax(S_fea[i] / self.temperature, dim=1)\n teacher_distance = F.softmax(T_fea[i].detach() / self.temperature, dim=1)\n loss_distill_dis += F.kl_div(\n student_distance, teacher_distance, reduction='batchmean')\n loss_distill_abs += nn.L1Loss()(S_fea[i], T_fea[i].detach())\n losses_distill_distribution.update(loss_distill_dis.item())\n #losses_distill_distribution.update(loss_distill_dis)\n losses_distill_abs.update(loss_distill_abs.item())\n loss = loss_SR + loss_distill_dis #+ 0.1* loss_distill_abs\n losses_sr.update(loss_SR.item())\n\n # backward\n self.optimizer.zero_grad()\n loss.backward()\n self.optimizer.step()\n timer.hold()\n\n if epoch <= self.args.epochs_encoder:\n if (batch + 1) % self.args.print_every == 0:\n self.ckp.write_log(\n 'Epoch: [{:04d}][{:04d}/{:04d}]\\t'\n 'Loss [ distill_dis loss:{:.3f}, distill_abs loss:{:.3f}]\\t'\n 'Time [{:.1f}s]'.format(\n epoch, (batch + 1) * self.args.batch_size, len(self.loader_train.dataset),\n losses_distill_distribution.avg, losses_distill_abs.avg,\n timer.release(),\n ))\n else:\n if (batch + 1) % self.args.print_every == 0:\n self.ckp.write_log(\n 'Epoch: [{:04d}][{:04d}/{:04d}]\\t'\n 'Loss [SR loss:{:.3f}, distill_dis loss:{:.3f}, distill_abs loss:{:.3f}]\\t'\n 'Time [{:.1f}s]'.format(\n epoch, (batch + 1) * self.args.batch_size, len(self.loader_train.dataset),\n losses_sr.avg, losses_distill_distribution.avg, losses_distill_abs.avg,\n timer.release(),\n ))\n\n self.loss.end_log(len(self.loader_train))\n\n # save model\n if epoch > self.args.st_save_epoch or epoch%30==0:\n target = self.model_ST.get_model()\n model_dict = target.state_dict()\n torch.save(\n model_dict,\n os.path.join(self.ckp.dir, 'model', 'model_ST_{}.pt'.format(epoch))\n )\n\n target = self.model_ST.get_model()\n model_dict = target.state_dict()\n torch.save(\n model_dict,\n os.path.join(self.ckp.dir, 'model', 'model_ST_last.pt')\n )\n\n def test(self):\n self.ckp.write_log('\\nEvaluation:')\n self.ckp.add_log(torch.zeros(1, len(self.scale)))\n self.model_ST.eval()\n\n timer_test = utility.timer()\n\n with torch.no_grad():\n degrade = util2.SRMDPreprocessing(\n self.scale[0],\n kernel_size=self.args.blur_kernel,\n blur_type=self.args.blur_type,\n sig=self.args.sig,\n lambda_1=self.args.lambda_1,\n lambda_2=self.args.lambda_2,\n theta=self.args.theta,\n noise=10\n )\n for idx_data, d in enumerate(self.loader_test):\n for idx_scale, scale in enumerate(self.scale):\n d.dataset.set_scale(idx_scale)\n eval_psnr = 0\n eval_ssim = 0\n for idx, (hr, filename) in enumerate(d):\n hr = hr.cuda() # b, c, h, w\n hr = self.crop_border(hr, scale)\n lr_blur, hr_blur = degrade(hr, random=False) # b, c, h, w\n\n # inference\n timer_test.tic()\n sr = self.model_ST(lr_blur)\n timer_test.hold()\n\n sr = utility.quantize(sr, self.args.rgb_range)\n hr = utility.quantize(hr, self.args.rgb_range)\n\n # metrics\n eval_psnr += utility.calc_psnr(\n sr, hr, scale, self.args.rgb_range,\n benchmark=d.dataset.benchmark\n )\n eval_ssim += utility.calc_ssim(\n (sr * 255).round().clamp(0, 255), (hr * 255).round().clamp(0, 255), scale,\n benchmark=d.dataset.benchmark\n )\n\n # save results\n if self.args.save_results:\n save_list = [sr]\n filename = filename[0]\n self.ckp.save_results(filename, save_list, scale)\n\n if len(self.test_res_psnr) > 10:\n self.test_res_psnr.pop(0)\n self.test_res_ssim.pop(0)\n self.test_res_psnr.append(eval_psnr / len(d))\n self.test_res_ssim.append(eval_ssim / len(d))\n\n self.ckp.log[-1, idx_scale] = eval_psnr / len(d)\n self.ckp.write_log(\n '[Epoch {}---{} x{}]\\tPSNR: {:.3f} SSIM: {:.4f} mean_PSNR: {:.3f} mean_SSIM: {:.4f}'.format(\n self.args.resume,\n self.args.data_test,\n scale,\n eval_psnr / len(d),\n eval_ssim / len(d),\n np.mean(self.test_res_psnr),\n np.mean(self.test_res_ssim)\n ))\n\n def crop_border(self, img_hr, scale):\n b, c, h, w = img_hr.size()\n\n img_hr = img_hr[:, :, :int(h//scale*scale), :int(w//scale*scale)]\n\n return img_hr\n\n def terminate(self):\n if self.args.test_only:\n self.test()\n return True\n else:\n epoch = self.scheduler.last_epoch + 1\n return epoch >= self.args.epochs_sr\n\n"
},
{
"path": "KDSR-classic/trainer_iso_stage3.py",
"content": "import os\nimport utility2\nimport utility\nimport torch\nfrom decimal import Decimal\nimport torch.nn.functional as F\nfrom utils import util\nfrom utils import util2\nfrom collections import OrderedDict\nimport random\nimport numpy as np\nimport torch.nn as nn\nimport utility3\n\n\nclass Trainer():\n def __init__(self, args, loader, model_TA, my_loss, ckp):\n self.test_res_psnr = []\n self.test_res_ssim = []\n self.args = args\n self.scale = args.scale\n self.loss1= nn.L1Loss()\n self.ckp = ckp\n self.loader_train = loader.loader_train\n self.loader_test = loader.loader_test\n self.model_TA = model_TA\n self.loss = my_loss\n self.optimizer = utility2.make_optimizer(args, self.model_TA)\n self.scheduler = utility2.make_scheduler(args, self.optimizer)\n self.pixel_unshuffle = nn.PixelUnshuffle(self.scale[0])\n if self.args.load != '.':\n self.optimizer.load_state_dict(\n torch.load(os.path.join(ckp.dir, 'optimizer.pt'))\n )\n for _ in range(len(ckp.log)): self.scheduler.step()\n\n def train(self):\n self.scheduler.step()\n self.loss.step()\n epoch = self.scheduler.last_epoch + 1\n\n lr = self.args.lr_sr * (self.args.gamma_sr ** ((epoch - self.args.epochs_encoder) // self.args.lr_decay_sr))\n for param_group in self.optimizer.param_groups:\n param_group['lr'] = lr\n\n self.ckp.write_log('[Epoch {}]\\tLearning rate: {:.2e}'.format(epoch, Decimal(lr)))\n self.loss.start_log()\n self.model_TA.train()\n\n degrade = util2.SRMDPreprocessing(\n self.scale[0],\n kernel_size=self.args.blur_kernel,\n blur_type=self.args.blur_type,\n sig_min=self.args.sig_min,\n sig_max=self.args.sig_max,\n lambda_min=self.args.lambda_min,\n lambda_max=self.args.lambda_max,\n noise=self.args.noise\n )\n\n timer = utility2.timer()\n\n for batch, ( hr, _,) in enumerate(self.loader_train):\n hr = hr.cuda() # b, c, h, w\n timer.tic()\n loss_all = 0\n lr_blur, hr_blur = degrade(hr) # b, c, h, w\n hr2 = self.pixel_unshuffle(hr)\n sr, _ = self.model_TA(lr_blur, torch.cat([lr_blur,hr2], dim=1))\n # sr,_ = self.model_TA(lr_blur, torch.cat([hr_blur,hr],dim=1))\n loss_all += self.loss1(sr,hr)\n self.optimizer.zero_grad()\n loss_all.backward()\n self.optimizer.step()\n\n # Remove the hooks before next training phase\n timer.hold()\n\n if (batch + 1) % self.args.print_every == 0:\n self.ckp.write_log(\n 'Epoch: [{:04d}][{:04d}/{:04d}]\\t'\n 'Loss [SR loss:{:.3f}]\\t'\n 'Time [{:.1f}s]'.format(\n epoch, (batch + 1) * self.args.batch_size, len(self.loader_train.dataset),\n loss_all.item(),\n timer.release(),\n ))\n\n self.loss.end_log(len(self.loader_train))\n\n # save model\n if epoch > self.args.st_save_epoch or (epoch %20 ==0):\n target = self.model_TA.get_model()\n model_dict = target.state_dict()\n torch.save(\n model_dict,\n os.path.join(self.ckp.dir, 'model', 'model_TA_{}.pt'.format(epoch))\n )\n\n optimzer_dict = self.optimizer.state_dict()\n torch.save(\n optimzer_dict,\n os.path.join(self.ckp.dir, 'optimzer', 'optimzer_TA_{}.pt'.format(epoch))\n )\n\n target = self.model_TA.get_model()\n model_dict = target.state_dict()\n torch.save(\n model_dict,\n os.path.join(self.ckp.dir, 'model', 'model_TA_last.pt')\n )\n optimzer_dict = self.optimizer.state_dict()\n torch.save(\n optimzer_dict,\n os.path.join(self.ckp.dir, 'optimzer', 'optimzer_TA_last.pt')\n )\n\n\n def test(self):\n self.ckp.write_log('\\nEvaluation:')\n self.ckp.add_log(torch.zeros(1, len(self.scale)))\n self.model_TA.eval()\n if self.scale[0]==4:\n sig_list=[1.8,2.0,2.2,2.4,2.6,2.8,3.0,3.2]\n elif self.scale[0]==2:\n sig_list=[0.9,1,1.1,1.2,1.3,1.4,1.5,1.6]\n timer_test = utility.timer()\n\n with torch.no_grad():\n\n for idx_data, d in enumerate(self.loader_test):\n for idx_scale, scale in enumerate(self.scale):\n d.dataset.set_scale(idx_scale)\n eval_psnr = 0\n eval_ssim = 0\n for idx, (hr0, filename) in enumerate(d):\n for sig in sig_list:\n degrade = util2.SRMDPreprocessing(\n self.scale[0],\n kernel_size=self.args.blur_kernel,\n blur_type=self.args.blur_type,\n sig=sig,\n lambda_1=self.args.lambda_1,\n lambda_2=self.args.lambda_2,\n theta=self.args.theta,\n noise=self.args.noise\n )\n\n hr = hr0.cuda() # b, c, h, w\n hr = self.crop_border(hr, scale)\n\n # inference\n timer_test.tic()\n lr_blur, hr_blur = degrade(hr, random=False)\n hr2 = self.pixel_unshuffle(hr)\n sr = self.model_TA(lr_blur, torch.cat([lr_blur, hr2], dim=1))\n timer_test.hold()\n\n sr = utility.quantize(sr, self.args.rgb_range)\n hr = utility.quantize(hr, self.args.rgb_range)\n\n # metrics\n # eval_psnr += utility.calc_psnr(\n # sr, hr, scale, self.args.rgb_range,\n # benchmark=d.dataset.benchmark\n # )\n eval_psnr += utility3.calc_psnr(\n sr, hr, scale\n )\n eval_ssim += utility.calc_ssim(\n (sr * 255).round().clamp(0, 255), (hr * 255).round().clamp(0, 255), scale,\n benchmark=d.dataset.benchmark\n )\n\n # save results\n if self.args.save_results:\n save_list = [sr]\n filename = filename[0]\n self.ckp.save_results(filename, save_list, scale)\n\n # if len(self.test_res_psnr) > 10:\n # self.test_res_psnr.pop(0)\n # self.test_res_ssim.pop(0)\n # self.test_res_psnr.append(eval_psnr / len(d) /len(sig_list))\n # self.test_res_ssim.append(eval_ssim / len(d)/len(sig_list))\n\n self.ckp.log[-1, idx_scale] = eval_psnr / len(d)/len(sig_list)\n self.ckp.write_log(\n '[Epoch {}---{} x{}]\\tPSNR: {:.3f} SSIM: {:.4f}'.format(\n self.args.resume,\n self.args.data_test,\n scale,\n eval_psnr / len(d) /len(sig_list),\n eval_ssim / len(d) /len(sig_list)\n ))\n \n def crop_border(self, img_hr, scale):\n b, c, h, w = img_hr.size()\n\n img_hr = img_hr[:, :, :int(h//scale*scale), :int(w//scale*scale)]\n\n return img_hr\n\n def terminate(self):\n if self.args.test_only:\n self.test()\n return True\n else:\n epoch = self.scheduler.last_epoch + 1\n return epoch >= self.args.epochs_sr"
},
{
"path": "KDSR-classic/trainer_iso_stage4.py",
"content": "import os\nimport utility\nimport torch\nfrom decimal import Decimal\nimport torch.nn.functional as F\nfrom utils import util2\nimport numpy as np\nimport torch.nn as nn\n\n\nclass Trainer():\n def __init__(self, args, loader, model_ST, model_TA, my_loss, ckp):\n self.is_first =True\n self.args = args\n self.scale = args.scale\n self.test_res_psnr = []\n self.test_res_ssim = []\n self.ckp = ckp\n self.loader_train = loader.loader_train\n self.loader_test = loader.loader_test\n self.model_ST = model_ST\n self.model_TA = model_TA\n self.model_Est = torch.nn.DataParallel(self.model_ST.get_model().E_st, range(self.args.n_GPUs))\n self.model_Eta = torch.nn.DataParallel(self.model_TA.get_model().E, range(self.args.n_GPUs))\n self.loss1 = nn.L1Loss()\n self.loss = my_loss\n self.optimizer = utility.make_optimizer(args, self.model_ST)\n self.scheduler = utility.make_scheduler(args, self.optimizer)\n self.temperature = args.temperature\n self.pixel_unshuffle = nn.PixelUnshuffle(self.scale[0])\n\n if self.args.load != '.':\n self.optimizer.load_state_dict(\n torch.load(os.path.join(ckp.dir, 'optimizer.pt'))\n )\n for _ in range(len(ckp.log)): self.scheduler.step()\n\n def train(self):\n self.scheduler.step()\n self.loss.step()\n epoch = self.scheduler.last_epoch + 1\n\n if epoch <= self.args.epochs_encoder:\n lr = self.args.lr_encoder * (self.args.gamma_encoder ** (epoch // self.args.lr_decay_encoder))\n for param_group in self.optimizer.param_groups:\n param_group['lr'] = lr\n else:\n lr = self.args.lr_sr * (self.args.gamma_sr ** ((epoch - self.args.epochs_encoder) // self.args.lr_decay_sr))\n for param_group in self.optimizer.param_groups:\n param_group['lr'] = lr\n\n self.ckp.write_log('[Epoch {}]\\tLearning rate: {:.2e}'.format(epoch, Decimal(lr)))\n self.loss.start_log()\n self.model_ST.train()\n\n degrade = util2.SRMDPreprocessing(\n self.scale[0],\n kernel_size=self.args.blur_kernel,\n blur_type=self.args.blur_type,\n sig_min=self.args.sig_min,\n sig_max=self.args.sig_max,\n lambda_min=self.args.lambda_min,\n lambda_max=self.args.lambda_max,\n noise=self.args.noise\n )\n\n timer = utility.timer()\n losses_sr, losses_distill_distribution, losses_distill_abs = utility.AverageMeter(),utility.AverageMeter(),utility.AverageMeter()\n\n\n for batch, (hr, _,) in enumerate(self.loader_train):\n hr = hr.cuda() # b, c, h, w\n timer.tic()\n lr_blur, hr_blur = degrade(hr) # b, c, h, w\n hr2 = self.pixel_unshuffle(hr)\n _, T_fea = self.model_Eta(torch.cat([lr_blur,hr2],dim=1))\n\n loss_distill_dis = 0\n loss_distill_abs = 0\n\n if epoch <= self.args.epochs_encoder:\n _, S_fea = self.model_Est(lr_blur)\n for i in range(len(T_fea)):\n student_distance = F.log_softmax(S_fea[i] / self.temperature, dim=1)\n teacher_distance = F.softmax(T_fea[i].detach()/ self.temperature, dim=1)\n loss_distill_dis += F.kl_div(\n student_distance, teacher_distance, reduction='batchmean')\n loss_distill_abs += nn.L1Loss()(S_fea[i], T_fea[i].detach())\n losses_distill_distribution.update(loss_distill_dis.item())\n #losses_distill_distribution.update(loss_distill_dis)\n losses_distill_abs.update(loss_distill_abs.item())\n loss = loss_distill_dis #+ 0.1* loss_distill_abs\n else:\n sr, S_fea = self.model_ST(lr_blur)\n loss_SR = self.loss(sr, hr)\n for i in range(len(T_fea)):\n student_distance = F.log_softmax(S_fea[i] / self.temperature, dim=1)\n teacher_distance = F.softmax(T_fea[i].detach() / self.temperature, dim=1)\n loss_distill_dis += F.kl_div(\n student_distance, teacher_distance, reduction='batchmean')\n loss_distill_abs += nn.L1Loss()(S_fea[i], T_fea[i].detach())\n losses_distill_distribution.update(loss_distill_dis.item())\n #losses_distill_distribution.update(loss_distill_dis)\n losses_distill_abs.update(loss_distill_abs.item())\n loss = loss_SR + loss_distill_dis #+ 0.1* loss_distill_abs\n losses_sr.update(loss_SR.item())\n\n # backward\n self.optimizer.zero_grad()\n loss.backward()\n self.optimizer.step()\n timer.hold()\n\n if epoch <= self.args.epochs_encoder:\n if (batch + 1) % self.args.print_every == 0:\n self.ckp.write_log(\n 'Epoch: [{:04d}][{:04d}/{:04d}]\\t'\n 'Loss [ distill_dis loss:{:.3f}, distill_abs loss:{:.3f}]\\t'\n 'Time [{:.1f}s]'.format(\n epoch, (batch + 1) * self.args.batch_size, len(self.loader_train.dataset),\n losses_distill_distribution.avg, losses_distill_abs.avg,\n timer.release(),\n ))\n else:\n if (batch + 1) % self.args.print_every == 0:\n self.ckp.write_log(\n 'Epoch: [{:04d}][{:04d}/{:04d}]\\t'\n 'Loss [SR loss:{:.3f}, distill_dis loss:{:.3f}, distill_abs loss:{:.3f}]\\t'\n 'Time [{:.1f}s]'.format(\n epoch, (batch + 1) * self.args.batch_size, len(self.loader_train.dataset),\n losses_sr.avg, losses_distill_distribution.avg, losses_distill_abs.avg,\n timer.release(),\n ))\n\n self.loss.end_log(len(self.loader_train))\n\n # save model\n if epoch > self.args.st_save_epoch or epoch%30==0:\n target = self.model_ST.get_model()\n model_dict = target.state_dict()\n torch.save(\n model_dict,\n os.path.join(self.ckp.dir, 'model', 'model_ST_{}.pt'.format(epoch))\n )\n\n target = self.model_ST.get_model()\n model_dict = target.state_dict()\n torch.save(\n model_dict,\n os.path.join(self.ckp.dir, 'model', 'model_ST_last.pt')\n )\n\n def test(self):\n self.ckp.write_log('\\nEvaluation:')\n self.ckp.add_log(torch.zeros(1, len(self.scale)))\n self.model_ST.eval()\n if self.scale[0]==4:\n sig_list=[1.8,2.0,2.2,2.4,2.6,2.8,3.0,3.2]\n elif self.scale[0]==2:\n sig_list=[0.9,1,1.1,1.2,1.3,1.4,1.5,1.6]\n timer_test = utility.timer()\n # print(sig_list)\n with torch.no_grad():\n for idx_data, d in enumerate(self.loader_test):\n for idx_scale, scale in enumerate(self.scale):\n d.dataset.set_scale(idx_scale)\n eval_psnr = 0\n eval_ssim = 0\n for idx, (hr0, filename) in enumerate(d):\n for sig in sig_list:\n degrade = util2.SRMDPreprocessing(\n self.scale[0],\n kernel_size=self.args.blur_kernel,\n blur_type=self.args.blur_type,\n sig=sig,\n lambda_1=self.args.lambda_1,\n lambda_2=self.args.lambda_2,\n theta=self.args.theta,\n noise=self.args.noise\n )\n\n hr = hr0.cuda() # b, c, h, w\n hr = self.crop_border(hr, scale)\n\n lr_blur, hr_blur = degrade(hr, random=False) # b, c, h, w\n\n # inference\n timer_test.tic()\n sr = self.model_ST(lr_blur)\n timer_test.hold()\n\n sr = utility.quantize(sr, self.args.rgb_range)\n hr = utility.quantize(hr, self.args.rgb_range)\n\n # metrics\n eval_psnr += utility.calc_psnr(\n sr, hr, scale, self.args.rgb_range,\n benchmark=d.dataset.benchmark\n )\n # eval_psnr += utility3.calc_psnr(\n # sr, hr, scale\n # )\n eval_ssim += utility.calc_ssim(\n (sr * 255).round().clamp(0, 255), (hr * 255).round().clamp(0, 255), scale,\n benchmark=d.dataset.benchmark\n )\n\n # save results\n if self.args.save_results:\n save_list = [sr]\n filename = filename[0]\n self.ckp.save_results(filename, save_list, scale)\n\n # if len(self.test_res_psnr) > 10:\n # self.test_res_psnr.pop(0)\n # self.test_res_ssim.pop(0)\n # self.test_res_psnr.append(eval_psnr / len(d) /len(sig_list))\n # self.test_res_ssim.append(eval_ssim / len(d)/len(sig_list))\n\n self.ckp.log[-1, idx_scale] = eval_psnr / len(d)/len(sig_list)\n self.ckp.write_log(\n '[Epoch {}---{} x{}]\\tPSNR: {:.3f} SSIM: {:.4f} '.format(\n self.args.resume,\n self.args.data_test,\n scale,\n eval_psnr / len(d) /len(sig_list),\n eval_ssim / len(d) /len(sig_list)\n ))\n\n def crop_border(self, img_hr, scale):\n b, c, h, w = img_hr.size()\n\n img_hr = img_hr[:, :, :int(h//scale*scale), :int(w//scale*scale)]\n\n return img_hr\n\n def terminate(self):\n if self.args.test_only:\n self.test()\n return True\n else:\n epoch = self.scheduler.last_epoch + 1\n return epoch >= self.args.epochs_sr\n\n"
},
{
"path": "KDSR-classic/utility.py",
"content": "import os\nimport math\nimport time\nimport datetime\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport scipy.misc as misc\nimport cv2\nimport torch\nimport torch.optim as optim\nimport torch.optim.lr_scheduler as lrs\nimport imageio\n\n\nclass AverageMeter(object):\n \"\"\"Computes and stores the average and current value\"\"\"\n def __init__(self):\n self.reset()\n\n def reset(self):\n self.val = 0\n self.avg = 0\n self.sum = 0\n self.count = 0\n\n def update(self, val, n=1):\n self.val = val\n self.sum += val * n\n self.count += n\n self.avg = self.sum / self.count\n\n\nclass timer():\n def __init__(self):\n self.acc = 0\n self.tic()\n\n def tic(self):\n self.t0 = time.time()\n\n def toc(self):\n return time.time() - self.t0\n\n def hold(self):\n self.acc += self.toc()\n\n def release(self):\n ret = self.acc\n self.acc = 0\n\n return ret\n\n def reset(self):\n self.acc = 0\n\n\nclass checkpoint():\n def __init__(self, args):\n self.args = args\n self.ok = True\n self.log = torch.Tensor()\n now = datetime.datetime.now().strftime('%Y-%m-%d-%H:%M:%S')\n\n if args.blur_type == 'iso_gaussian':\n self.dir = './experiment/' + args.save + '_x' + str(int(args.scale[0])) + '_' + args.mode + '_iso'\n elif args.blur_type == 'aniso_gaussian':\n self.dir = './experiment/' + args.save + '_x' + str(int(args.scale[0])) + '_' + args.mode + '_aniso'\n\n def _make_dir(path):\n if not os.path.exists(path): os.makedirs(path)\n\n _make_dir(self.dir)\n _make_dir(self.dir + '/model')\n _make_dir(self.dir + '/optimzer')\n _make_dir(self.dir + '/results')\n\n open_type = 'a' if os.path.exists(self.dir + '/log.txt') else 'w'\n self.log_file = open(self.dir + '/log.txt', open_type)\n with open(self.dir + '/config.txt', open_type) as f:\n f.write(now + '\\n\\n')\n for arg in vars(args):\n f.write('{}: {}\\n'.format(arg, getattr(args, arg)))\n f.write('\\n')\n\n def save(self, trainer, epoch, is_best=False):\n trainer.model.save(self.dir, epoch, is_best=is_best)\n trainer.loss.save(self.dir)\n trainer.loss.plot_loss(self.dir, epoch)\n\n self.plot_psnr(epoch)\n torch.save(self.log, os.path.join(self.dir, 'psnr_log.pt'))\n torch.save(\n trainer.optimizer.state_dict(),\n os.path.join(self.dir, 'optimizer.pt')\n )\n\n def add_log(self, log):\n self.log = torch.cat([self.log, log])\n\n def write_log(self, log, refresh=False):\n print(log)\n self.log_file.write(log + '\\n')\n if refresh:\n self.log_file.close()\n self.log_file = open(self.dir + '/log.txt', 'a')\n\n def done(self):\n self.log_file.close()\n\n def plot_psnr(self, epoch):\n axis = np.linspace(1, epoch, epoch)\n label = 'SR on {}'.format(self.args.data_test)\n fig = plt.figure()\n plt.title(label)\n for idx_scale, scale in enumerate(self.args.scale):\n plt.plot(\n axis,\n self.log[:, idx_scale].numpy(),\n label='Scale {}'.format(scale)\n )\n plt.legend()\n plt.xlabel('Epochs')\n plt.ylabel('PSNR')\n plt.grid(True)\n plt.savefig('{}/test_{}.pdf'.format(self.dir, self.args.data_test))\n plt.close(fig)\n\n def save_results(self, filename, save_list, scale):\n filename = '{}/results/{}_x{}_'.format(self.dir, filename, scale)\n\n normalized = save_list[0][0].data.mul(255 / self.args.rgb_range)\n ndarr = normalized.byte().permute(1, 2, 0).cpu().numpy()\n imageio.imsave('{}{}.png'.format(filename, 'SR'), ndarr)\n\n\ndef quantize(img, rgb_range):\n pixel_range = 255 / rgb_range\n return img.mul(pixel_range).clamp(0, 255).round().div(pixel_range)\n\n\ndef calc_psnr(sr, hr, scale, rgb_range, benchmark=False):\n diff = (sr - hr).data.div(rgb_range)\n if benchmark:\n shave = scale\n if diff.size(1) > 1:\n convert = diff.new(1, 3, 1, 1)\n convert[0, 0, 0, 0] = 65.738\n convert[0, 1, 0, 0] = 129.057\n convert[0, 2, 0, 0] = 25.064\n diff.mul_(convert).div_(256)\n diff = diff.sum(dim=1, keepdim=True)\n else:\n shave = scale + 6\n import math\n shave = math.ceil(shave)\n valid = diff[:, :, shave:-shave, shave:-shave]\n mse = valid.pow(2).mean()\n\n return -10 * math.log10(mse)\n\n\ndef calc_ssim(img1, img2, scale=2, benchmark=False):\n '''calculate SSIM\n the same outputs as MATLAB's\n img1, img2: [0, 255]\n '''\n if benchmark:\n border = math.ceil(scale)\n else:\n border = math.ceil(scale) + 6\n\n img1 = img1.data.squeeze().float().clamp(0, 255).round().cpu().numpy()\n img1 = np.transpose(img1, (1, 2, 0))\n img2 = img2.data.squeeze().cpu().numpy()\n img2 = np.transpose(img2, (1, 2, 0))\n\n img1_y = np.dot(img1, [65.738, 129.057, 25.064]) / 255.0 + 16.0\n img2_y = np.dot(img2, [65.738, 129.057, 25.064]) / 255.0 + 16.0\n if not img1.shape == img2.shape:\n raise ValueError('Input images must have the same dimensions.')\n h, w = img1.shape[:2]\n img1_y = img1_y[border:h - border, border:w - border]\n img2_y = img2_y[border:h - border, border:w - border]\n\n if img1_y.ndim == 2:\n return ssim(img1_y, img2_y)\n elif img1.ndim == 3:\n if img1.shape[2] == 3:\n ssims = []\n for i in range(3):\n ssims.append(ssim(img1, img2))\n return np.array(ssims).mean()\n elif img1.shape[2] == 1:\n return ssim(np.squeeze(img1), np.squeeze(img2))\n else:\n raise ValueError('Wrong input image dimensions.')\n\n\ndef ssim(img1, img2):\n C1 = (0.01 * 255) ** 2\n C2 = (0.03 * 255) ** 2\n\n img1 = img1.astype(np.float64)\n img2 = img2.astype(np.float64)\n kernel = cv2.getGaussianKernel(11, 1.5)\n window = np.outer(kernel, kernel.transpose())\n\n mu1 = cv2.filter2D(img1, -1, window)[5:-5, 5:-5] # valid\n mu2 = cv2.filter2D(img2, -1, window)[5:-5, 5:-5]\n mu1_sq = mu1 ** 2\n mu2_sq = mu2 ** 2\n mu1_mu2 = mu1 * mu2\n sigma1_sq = cv2.filter2D(img1 ** 2, -1, window)[5:-5, 5:-5] - mu1_sq\n sigma2_sq = cv2.filter2D(img2 ** 2, -1, window)[5:-5, 5:-5] - mu2_sq\n sigma12 = cv2.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2\n\n ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) *\n (sigma1_sq + sigma2_sq + C2))\n return ssim_map.mean()\n\n\ndef make_optimizer(args, my_model):\n trainable = filter(lambda x: x.requires_grad, my_model.parameters())\n\n if args.optimizer == 'SGD':\n optimizer_function = optim.SGD\n kwargs = {'momentum': args.momentum}\n elif args.optimizer == 'ADAM':\n optimizer_function = optim.Adam\n kwargs = {\n 'betas': (args.beta1, args.beta2),\n 'eps': args.epsilon\n }\n elif args.optimizer == 'RMSprop':\n optimizer_function = optim.RMSprop\n kwargs = {'eps': args.epsilon}\n\n kwargs['weight_decay'] = args.weight_decay\n\n return optimizer_function(trainable, **kwargs)\n\n\ndef make_scheduler(args, my_optimizer):\n if args.decay_type == 'step':\n scheduler = lrs.StepLR(\n my_optimizer,\n step_size=args.lr_decay_sr,\n gamma=args.gamma_sr,\n )\n elif args.decay_type.find('step') >= 0:\n milestones = args.decay_type.split('_')\n milestones.pop(0)\n milestones = list(map(lambda x: int(x), milestones))\n scheduler = lrs.MultiStepLR(\n my_optimizer,\n milestones=milestones,\n gamma=args.gamma\n )\n\n scheduler.step(args.start_epoch - 1)\n\n return scheduler\n\n"
},
{
"path": "KDSR-classic/utility2.py",
"content": "import os\nimport math\nimport time\nimport datetime\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport scipy.misc as misc\nimport cv2\nimport torch\nimport torch.optim as optim\nimport torch.optim.lr_scheduler as lrs\n\n\nclass AverageMeter(object):\n \"\"\"Computes and stores the average and current value\"\"\"\n def __init__(self):\n self.reset()\n\n def reset(self):\n self.val = 0\n self.avg = 0\n self.sum = 0\n self.count = 0\n\n def update(self, val, n=1):\n self.val = val\n self.sum += val * n\n self.count += n\n self.avg = self.sum / self.count\n\n\nclass timer():\n def __init__(self):\n self.acc = 0\n self.tic()\n\n def tic(self):\n self.t0 = time.time()\n\n def toc(self):\n return time.time() - self.t0\n\n def hold(self):\n self.acc += self.toc()\n\n def release(self):\n ret = self.acc\n self.acc = 0\n\n return ret\n\n def reset(self):\n self.acc = 0\n\n\nclass checkpoint():\n def __init__(self, args):\n self.args = args\n self.ok = True\n self.log = torch.Tensor()\n now = datetime.datetime.now().strftime('%Y-%m-%d-%H:%M:%S')\n\n if args.blur_type == 'iso_gaussian':\n self.dir = './experiment/' + args.save + '_x' + str(int(args.scale[0])) + '_' + args.mode + '_iso'\n elif args.blur_type == 'aniso_gaussian':\n self.dir = './experiment/' + args.save + '_x' + str(int(args.scale[0])) + '_' + args.mode + '_aniso'\n\n def _make_dir(path):\n if not os.path.exists(path): os.makedirs(path)\n\n _make_dir(self.dir)\n _make_dir(self.dir + '/model')\n _make_dir(self.dir + '/results')\n\n open_type = 'a' if os.path.exists(self.dir + '/log.txt') else 'w'\n self.log_file = open(self.dir + '/log.txt', open_type)\n with open(self.dir + '/config.txt', open_type) as f:\n f.write(now + '\\n\\n')\n for arg in vars(args):\n f.write('{}: {}\\n'.format(arg, getattr(args, arg)))\n f.write('\\n')\n\n def save(self, trainer, epoch, is_best=False):\n trainer.model.save(self.dir, epoch, is_best=is_best)\n trainer.loss.save(self.dir)\n trainer.loss.plot_loss(self.dir, epoch)\n\n self.plot_psnr(epoch)\n torch.save(self.log, os.path.join(self.dir, 'psnr_log.pt'))\n torch.save(\n trainer.optimizer.state_dict(),\n os.path.join(self.dir, 'optimizer.pt')\n )\n\n def add_log(self, log):\n self.log = torch.cat([self.log, log])\n\n def write_log(self, log, refresh=False):\n print(log)\n self.log_file.write(log + '\\n')\n if refresh:\n self.log_file.close()\n self.log_file = open(self.dir + '/log.txt', 'a')\n\n def done(self):\n self.log_file.close()\n\n def plot_psnr(self, epoch):\n axis = np.linspace(1, epoch, epoch)\n label = 'SR on {}'.format(self.args.data_test)\n fig = plt.figure()\n plt.title(label)\n for idx_scale, scale in enumerate(self.args.scale):\n plt.plot(\n axis,\n self.log[:, idx_scale].numpy(),\n label='Scale {}'.format(scale)\n )\n plt.legend()\n plt.xlabel('Epochs')\n plt.ylabel('PSNR')\n plt.grid(True)\n plt.savefig('{}/test_{}.pdf'.format(self.dir, self.args.data_test))\n plt.close(fig)\n\n def save_results(self, filename, save_list, scale):\n filename = '{}/results/{}_x{}_'.format(self.dir, filename, scale)\n\n normalized = save_list[0][0].data.mul(255 / self.args.rgb_range)\n ndarr = normalized.byte().permute(1, 2, 0).cpu().numpy()\n misc.imsave('{}{}.png'.format(filename, 'SR'), ndarr)\n\n\ndef quantize(img, rgb_range):\n pixel_range = 255 / rgb_range\n return img.mul(pixel_range).clamp(0, 255).round().div(pixel_range)\n\n\ndef calc_psnr(sr, hr, scale, rgb_range, benchmark=False):\n diff = (sr - hr).data.div(rgb_range)\n if benchmark:\n shave = scale\n if diff.size(1) > 1:\n convert = diff.new(1, 3, 1, 1)\n convert[0, 0, 0, 0] = 65.738\n convert[0, 1, 0, 0] = 129.057\n convert[0, 2, 0, 0] = 25.064\n diff.mul_(convert).div_(256)\n diff = diff.sum(dim=1, keepdim=True)\n else:\n shave = scale + 6\n import math\n shave = math.ceil(shave)\n valid = diff[:, :, shave:-shave, shave:-shave]\n mse = valid.pow(2).mean()\n\n return -10 * math.log10(mse)\n\n\ndef calc_ssim(img1, img2, scale=2, benchmark=False):\n '''calculate SSIM\n the same outputs as MATLAB's\n img1, img2: [0, 255]\n '''\n if benchmark:\n border = math.ceil(scale)\n else:\n border = math.ceil(scale) + 6\n\n img1 = img1.data.squeeze().float().clamp(0, 255).round().cpu().numpy()\n img1 = np.transpose(img1, (1, 2, 0))\n img2 = img2.data.squeeze().cpu().numpy()\n img2 = np.transpose(img2, (1, 2, 0))\n\n img1_y = np.dot(img1, [65.738, 129.057, 25.064]) / 255.0 + 16.0\n img2_y = np.dot(img2, [65.738, 129.057, 25.064]) / 255.0 + 16.0\n if not img1.shape == img2.shape:\n raise ValueError('Input images must have the same dimensions.')\n h, w = img1.shape[:2]\n img1_y = img1_y[border:h - border, border:w - border]\n img2_y = img2_y[border:h - border, border:w - border]\n\n if img1_y.ndim == 2:\n return ssim(img1_y, img2_y)\n elif img1.ndim == 3:\n if img1.shape[2] == 3:\n ssims = []\n for i in range(3):\n ssims.append(ssim(img1, img2))\n return np.array(ssims).mean()\n elif img1.shape[2] == 1:\n return ssim(np.squeeze(img1), np.squeeze(img2))\n else:\n raise ValueError('Wrong input image dimensions.')\n\n\ndef ssim(img1, img2):\n C1 = (0.01 * 255) ** 2\n C2 = (0.03 * 255) ** 2\n\n img1 = img1.astype(np.float64)\n img2 = img2.astype(np.float64)\n kernel = cv2.getGaussianKernel(11, 1.5)\n window = np.outer(kernel, kernel.transpose())\n\n mu1 = cv2.filter2D(img1, -1, window)[5:-5, 5:-5] # valid\n mu2 = cv2.filter2D(img2, -1, window)[5:-5, 5:-5]\n mu1_sq = mu1 ** 2\n mu2_sq = mu2 ** 2\n mu1_mu2 = mu1 * mu2\n sigma1_sq = cv2.filter2D(img1 ** 2, -1, window)[5:-5, 5:-5] - mu1_sq\n sigma2_sq = cv2.filter2D(img2 ** 2, -1, window)[5:-5, 5:-5] - mu2_sq\n sigma12 = cv2.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2\n\n ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) *\n (sigma1_sq + sigma2_sq + C2))\n return ssim_map.mean()\n\n\ndef make_optimizer(args, my_model):\n trainable = filter(lambda x: x.requires_grad, my_model.parameters())\n\n if args.optimizer == 'SGD':\n optimizer_function = optim.SGD\n kwargs = {'momentum': args.momentum}\n elif args.optimizer == 'ADAM':\n optimizer_function = optim.Adam\n kwargs = {\n 'betas': (args.beta1, args.beta2),\n 'eps': args.epsilon\n }\n elif args.optimizer == 'RMSprop':\n optimizer_function = optim.RMSprop\n kwargs = {'eps': args.epsilon}\n\n kwargs['weight_decay'] = args.weight_decay\n\n return optimizer_function(trainable, **kwargs)\n\n\ndef make_scheduler(args, my_optimizer):\n if args.decay_type == 'step':\n scheduler = lrs.StepLR(\n my_optimizer,\n step_size=args.lr_decay_sr,\n gamma=args.gamma_sr,\n )\n elif args.decay_type.find('step') >= 0:\n milestones = args.decay_type.split('_')\n milestones.pop(0)\n milestones = list(map(lambda x: int(x), milestones))\n scheduler = lrs.MultiStepLR(\n my_optimizer,\n milestones=milestones,\n gamma=args.gamma\n )\n\n scheduler.step(args.start_epoch - 1)\n\n return scheduler\n\n"
},
{
"path": "KDSR-classic/utility3.py",
"content": "import os\nimport math\nimport time\nimport datetime\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport scipy.misc as misc\nimport cv2\nimport torch\nimport torch.optim as optim\nimport torch.optim.lr_scheduler as lrs\n\ndef tensor2img(tensor, out_type=np.float32, min_max=(0, 1)):\n \"\"\"\n Converts a torch Tensor into an image Numpy array\n Input: 4D(B,(3/1),H,W), 3D(C,H,W), or 2D(H,W), any range, RGB channel order\n Output: 3D(H,W,C) or 2D(H,W), [0,255], np.uint8 (default)\n \"\"\"\n tensor = tensor.squeeze().float().cpu().clamp_(*min_max) # clamp\n tensor = (tensor - min_max[0]) / (min_max[1] - min_max[0]) # to range [0,1]\n n_dim = tensor.dim()\n if n_dim == 4:\n n_img = len(tensor)\n img_np = make_grid(tensor, nrow=int(math.sqrt(n_img)), normalize=False).numpy()\n img_np = np.transpose(img_np[[2, 1, 0], :, :], (1, 2, 0)) # HWC, BGR\n elif n_dim == 3:\n img_np = tensor.numpy()\n img_np = np.transpose(img_np[[2, 1, 0], :, :], (1, 2, 0)) # HWC, BGR\n elif n_dim == 2:\n img_np = tensor.numpy()\n else:\n raise TypeError(\n \"Only support 4D, 3D and 2D tensor. But received with dimension: {:d}\".format(\n n_dim\n )\n )\n if out_type == np.uint8:\n img_np = (img_np * 255.0).round()\n # Important. Unlike matlab, numpy.unit8() WILL NOT round by default.\n return img_np.astype(out_type)\n\ndef bgr2ycbcr(img, only_y=True):\n '''bgr version of rgb2ycbcr\n only_y: only return Y channel\n Input:\n uint8, [0, 255]\n float, [0, 1]\n '''\n img = tensor2img(img.squeeze())\n in_img_type = img.dtype\n img.astype(np.float32)\n if in_img_type != np.uint8:\n img *= 255.\n # convert\n if only_y:\n rlt = np.dot(img, [24.966, 128.553, 65.481]) / 255.0 + 16.0\n else:\n rlt = np.matmul(img, [[24.966, 112.0, -18.214], [128.553, -74.203, -93.786],\n [65.481, -37.797, 112.0]]) / 255.0 + [16, 128, 128]\n if in_img_type == np.uint8:\n rlt = rlt.round()\n else:\n rlt /= 255.\n return rlt.astype(in_img_type)\n\ndef calc_psnr(img1, img2,scale):\n # img1 and img2 have range [0, 255]\n img1 = img1.data\n img2 = img2.data\n img1 = img1[:, :, scale:-scale, scale:-scale]\n img2 = img2[:, :, scale:-scale, scale:-scale]\n img1 = bgr2ycbcr(img1, only_y=True) * 255\n img2 = bgr2ycbcr(img2, only_y=True) * 255\n img1 = img1.astype(np.float64)\n img2 = img2.astype(np.float64)\n mse = np.mean((img1 - img2) ** 2)\n if mse == 0:\n return float(\"inf\")\n return 20 * math.log10(255.0 / math.sqrt(mse))\n\n\ndef ssim(img1, img2):\n C1 = (0.01 * 255) ** 2\n C2 = (0.03 * 255) ** 2\n\n img1 = img1.astype(np.float64)\n img2 = img2.astype(np.float64)\n kernel = cv2.getGaussianKernel(11, 1.5)\n window = np.outer(kernel, kernel.transpose())\n\n mu1 = cv2.filter2D(img1, -1, window)[5:-5, 5:-5] # valid\n mu2 = cv2.filter2D(img2, -1, window)[5:-5, 5:-5]\n mu1_sq = mu1 ** 2\n mu2_sq = mu2 ** 2\n mu1_mu2 = mu1 * mu2\n sigma1_sq = cv2.filter2D(img1 ** 2, -1, window)[5:-5, 5:-5] - mu1_sq\n sigma2_sq = cv2.filter2D(img2 ** 2, -1, window)[5:-5, 5:-5] - mu2_sq\n sigma12 = cv2.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2\n\n ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / (\n (mu1_sq + mu2_sq + C1) * (sigma1_sq + sigma2_sq + C2)\n )\n return ssim_map.mean()\n\n\ndef calculate_ssim(img1, img2):\n \"\"\"calculate SSIM\n the same outputs as MATLAB's\n img1, img2: [0, 255]\n \"\"\"\n if not img1.shape == img2.shape:\n raise ValueError(\"Input images must have the same dimensions.\")\n if img1.ndim == 2:\n return ssim(img1, img2)\n elif img1.ndim == 3:\n if img1.shape[2] == 3:\n ssims = []\n for i in range(3):\n ssims.append(ssim(img1, img2))\n return np.array(ssims).mean()\n elif img1.shape[2] == 1:\n return ssim(np.squeeze(img1), np.squeeze(img2))\n else:\n raise ValueError(\"Wrong input image dimensions.\")"
},
{
"path": "KDSR-classic/utils/__init__.py",
"content": ""
},
{
"path": "KDSR-classic/utils/util.py",
"content": "import math\nimport numpy as np\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport utility\n\n\ndef cal_sigma(sig_x, sig_y, radians):\n sig_x = sig_x.view(-1, 1, 1)\n sig_y = sig_y.view(-1, 1, 1)\n radians = radians.view(-1, 1, 1)\n\n D = torch.cat([F.pad(sig_x ** 2, [0, 1, 0, 0]), F.pad(sig_y ** 2, [1, 0, 0, 0])], 1)\n U = torch.cat([torch.cat([radians.cos(), -radians.sin()], 2),\n torch.cat([radians.sin(), radians.cos()], 2)], 1)\n sigma = torch.bmm(U, torch.bmm(D, U.transpose(1, 2)))\n\n return sigma\n\n\ndef anisotropic_gaussian_kernel(batch, kernel_size, covar):\n ax = torch.arange(kernel_size).float().cuda() - kernel_size // 2\n\n xx = ax.repeat(kernel_size).view(1, kernel_size, kernel_size).expand(batch, -1, -1)\n yy = ax.repeat_interleave(kernel_size).view(1, kernel_size, kernel_size).expand(batch, -1, -1)\n xy = torch.stack([xx, yy], -1).view(batch, -1, 2)\n\n inverse_sigma = torch.inverse(covar)\n kernel = torch.exp(- 0.5 * (torch.bmm(xy, inverse_sigma) * xy).sum(2)).view(batch, kernel_size, kernel_size)\n\n return kernel / kernel.sum([1, 2], keepdim=True)\n\n\ndef isotropic_gaussian_kernel(batch, kernel_size, sigma):\n ax = torch.arange(kernel_size).float().cuda() - kernel_size//2\n xx = ax.repeat(kernel_size).view(1, kernel_size, kernel_size).expand(batch, -1, -1)\n yy = ax.repeat_interleave(kernel_size).view(1, kernel_size, kernel_size).expand(batch, -1, -1)\n kernel = torch.exp(-(xx ** 2 + yy ** 2) / (2. * sigma.view(-1, 1, 1) ** 2))\n\n return kernel / kernel.sum([1,2], keepdim=True)\n\n\ndef random_anisotropic_gaussian_kernel(batch=1, kernel_size=21, lambda_min=0.2, lambda_max=4.0):\n theta = torch.rand(batch).cuda() * math.pi\n lambda_1 = torch.rand(batch).cuda() * (lambda_max - lambda_min) + lambda_min\n lambda_2 = torch.rand(batch).cuda() * (lambda_max - lambda_min) + lambda_min\n\n covar = cal_sigma(lambda_1, lambda_2, theta)\n kernel = anisotropic_gaussian_kernel(batch, kernel_size, covar)\n return kernel\n\n\ndef stable_anisotropic_gaussian_kernel(kernel_size=21, theta=0, lambda_1=0.2, lambda_2=4.0):\n theta = torch.ones(1).cuda() * theta / 180 * math.pi\n lambda_1 = torch.ones(1).cuda() * lambda_1\n lambda_2 = torch.ones(1).cuda() * lambda_2\n\n covar = cal_sigma(lambda_1, lambda_2, theta)\n kernel = anisotropic_gaussian_kernel(1, kernel_size, covar)\n return kernel\n\n\ndef random_isotropic_gaussian_kernel(batch=1, kernel_size=21, sig_min=0.2, sig_max=4.0):\n x = torch.rand(batch).cuda() * (sig_max - sig_min) + sig_min\n k = isotropic_gaussian_kernel(batch, kernel_size, x)\n return k\n\n\ndef stable_isotropic_gaussian_kernel(kernel_size=21, sig=4.0):\n x = torch.ones(1).cuda() * sig\n k = isotropic_gaussian_kernel(1, kernel_size, x)\n return k\n\n\ndef random_gaussian_kernel(batch, kernel_size=21, blur_type='iso_gaussian', sig_min=0.2, sig_max=4.0, lambda_min=0.2, lambda_max=4.0):\n if blur_type == 'iso_gaussian':\n return random_isotropic_gaussian_kernel(batch=batch, kernel_size=kernel_size, sig_min=sig_min, sig_max=sig_max)\n elif blur_type == 'aniso_gaussian':\n return random_anisotropic_gaussian_kernel(batch=batch, kernel_size=kernel_size, lambda_min=lambda_min, lambda_max=lambda_max)\n\n\ndef stable_gaussian_kernel(kernel_size=21, blur_type='iso_gaussian', sig=2.6, lambda_1=0.2, lambda_2=4.0, theta=0):\n if blur_type == 'iso_gaussian':\n return stable_isotropic_gaussian_kernel(kernel_size=kernel_size, sig=sig)\n elif blur_type == 'aniso_gaussian':\n return stable_anisotropic_gaussian_kernel(kernel_size=kernel_size, lambda_1=lambda_1, lambda_2=lambda_2, theta=theta)\n\n\n# implementation of matlab bicubic interpolation in pytorch\nclass bicubic(nn.Module):\n def __init__(self):\n super(bicubic, self).__init__()\n\n def cubic(self, x):\n absx = torch.abs(x)\n absx2 = torch.abs(x) * torch.abs(x)\n absx3 = torch.abs(x) * torch.abs(x) * torch.abs(x)\n\n condition1 = (absx <= 1).to(torch.float32)\n condition2 = ((1 < absx) & (absx <= 2)).to(torch.float32)\n\n f = (1.5 * absx3 - 2.5 * absx2 + 1) * condition1 + (-0.5 * absx3 + 2.5 * absx2 - 4 * absx + 2) * condition2\n return f\n\n def contribute(self, in_size, out_size, scale):\n kernel_width = 4\n if scale < 1:\n kernel_width = 4 / scale\n x0 = torch.arange(start=1, end=out_size[0] + 1).to(torch.float32).cuda()\n x1 = torch.arange(start=1, end=out_size[1] + 1).to(torch.float32).cuda()\n\n u0 = x0 / scale + 0.5 * (1 - 1 / scale)\n u1 = x1 / scale + 0.5 * (1 - 1 / scale)\n\n left0 = torch.floor(u0 - kernel_width / 2)\n left1 = torch.floor(u1 - kernel_width / 2)\n\n P = np.ceil(kernel_width) + 2\n\n indice0 = left0.unsqueeze(1) + torch.arange(start=0, end=P).to(torch.float32).unsqueeze(0).cuda()\n indice1 = left1.unsqueeze(1) + torch.arange(start=0, end=P).to(torch.float32).unsqueeze(0).cuda()\n\n mid0 = u0.unsqueeze(1) - indice0.unsqueeze(0)\n mid1 = u1.unsqueeze(1) - indice1.unsqueeze(0)\n\n if scale < 1:\n weight0 = scale * self.cubic(mid0 * scale)\n weight1 = scale * self.cubic(mid1 * scale)\n else:\n weight0 = self.cubic(mid0)\n weight1 = self.cubic(mid1)\n\n weight0 = weight0 / (torch.sum(weight0, 2).unsqueeze(2))\n weight1 = weight1 / (torch.sum(weight1, 2).unsqueeze(2))\n\n indice0 = torch.min(torch.max(torch.FloatTensor([1]).cuda(), indice0), torch.FloatTensor([in_size[0]]).cuda()).unsqueeze(0)\n indice1 = torch.min(torch.max(torch.FloatTensor([1]).cuda(), indice1), torch.FloatTensor([in_size[1]]).cuda()).unsqueeze(0)\n\n kill0 = torch.eq(weight0, 0)[0][0]\n kill1 = torch.eq(weight1, 0)[0][0]\n\n weight0 = weight0[:, :, kill0 == 0]\n weight1 = weight1[:, :, kill1 == 0]\n\n indice0 = indice0[:, :, kill0 == 0]\n indice1 = indice1[:, :, kill1 == 0]\n\n return weight0, weight1, indice0, indice1\n\n def forward(self, input, scale=1/4):\n b, c, h, w = input.shape\n\n weight0, weight1, indice0, indice1 = self.contribute([h, w], [int(h * scale), int(w * scale)], scale)\n weight0 = weight0[0]\n weight1 = weight1[0]\n\n indice0 = indice0[0].long()\n indice1 = indice1[0].long()\n\n out = input[:, :, (indice0 - 1), :] * (weight0.unsqueeze(0).unsqueeze(1).unsqueeze(4))\n out = (torch.sum(out, dim=3))\n A = out.permute(0, 1, 3, 2)\n\n out = A[:, :, (indice1 - 1), :] * (weight1.unsqueeze(0).unsqueeze(1).unsqueeze(4))\n out = out.sum(3).permute(0, 1, 3, 2)\n\n return out\n\n\nclass Gaussin_Kernel(object):\n def __init__(self, kernel_size=21, blur_type='iso_gaussian',\n sig=2.6, sig_min=0.2, sig_max=4.0,\n lambda_1=0.2, lambda_2=4.0, theta=0, lambda_min=0.2, lambda_max=4.0):\n self.kernel_size = kernel_size\n self.blur_type = blur_type\n\n self.sig = sig\n self.sig_min = sig_min\n self.sig_max = sig_max\n\n self.lambda_1 = lambda_1\n self.lambda_2 = lambda_2\n self.theta = theta\n self.lambda_min = lambda_min\n self.lambda_max = lambda_max\n\n def __call__(self, batch, random):\n # random kernel\n if random == True:\n return random_gaussian_kernel(batch, kernel_size=self.kernel_size, blur_type=self.blur_type,\n sig_min=self.sig_min, sig_max=self.sig_max,\n lambda_min=self.lambda_min, lambda_max=self.lambda_max)\n\n # stable kernel\n else:\n return stable_gaussian_kernel(kernel_size=self.kernel_size, blur_type=self.blur_type,\n sig=self.sig,\n lambda_1=self.lambda_1, lambda_2=self.lambda_2, theta=self.theta)\n\nclass BatchBlur(nn.Module):\n def __init__(self, kernel_size=21):\n super(BatchBlur, self).__init__()\n self.kernel_size = kernel_size\n if kernel_size % 2 == 1:\n self.pad = nn.ReflectionPad2d(kernel_size//2)\n else:\n self.pad = nn.ReflectionPad2d((kernel_size//2, kernel_size//2-1, kernel_size//2, kernel_size//2-1))\n\n def forward(self, input, kernel):\n B, C, H, W = input.size()\n input_pad = self.pad(input)\n H_p, W_p = input_pad.size()[-2:]\n\n if len(kernel.size()) == 2:\n input_CBHW = input_pad.view((C * B, 1, H_p, W_p))\n kernel = kernel.contiguous().view((1, 1, self.kernel_size, self.kernel_size))\n\n return F.conv2d(input_CBHW, kernel, padding=0).view((B, C, H, W))\n else:\n input_CBHW = input_pad.view((1, C * B, H_p, W_p))\n kernel = kernel.contiguous().view((B, 1, self.kernel_size, self.kernel_size))\n kernel = kernel.repeat(1, C, 1, 1).view((B * C, 1, self.kernel_size, self.kernel_size))\n\n return F.conv2d(input_CBHW, kernel, groups=B*C).view((B, C, H, W))\n\n\nclass SRMDPreprocessing(object):\n def __init__(self,\n scale,\n mode='bicubic',\n kernel_size=21,\n blur_type='iso_gaussian',\n sig=2.6,\n sig_min=0.2,\n sig_max=4.0,\n lambda_1=0.2,\n lambda_2=4.0,\n theta=0,\n lambda_min=0.2,\n lambda_max=4.0,\n noise=0.0,\n rgb_range=1\n ):\n '''\n # sig, sig_min and sig_max are used for isotropic Gaussian blurs\n During training phase (random=True):\n the width of the blur kernel is randomly selected from [sig_min, sig_max]\n During test phase (random=False):\n the width of the blur kernel is set to sig\n\n # lambda_1, lambda_2, theta, lambda_min and lambda_max are used for anisotropic Gaussian blurs\n During training phase (random=True):\n the eigenvalues of the covariance is randomly selected from [lambda_min, lambda_max]\n the angle value is randomly selected from [0, pi]\n During test phase (random=False):\n the eigenvalues of the covariance are set to lambda_1 and lambda_2\n the angle value is set to theta\n '''\n self.kernel_size = kernel_size\n self.scale = scale\n self.mode = mode\n self.noise = noise / (255/rgb_range)\n self.rgb_range=rgb_range\n\n self.gen_kernel = Gaussin_Kernel(\n kernel_size=kernel_size, blur_type=blur_type,\n sig=sig, sig_min=sig_min, sig_max=sig_max,\n lambda_1=lambda_1, lambda_2=lambda_2, theta=theta, lambda_min=lambda_min, lambda_max=lambda_max\n )\n self.blur = BatchBlur(kernel_size=kernel_size)\n self.bicubic = bicubic()\n\n def __call__(self, hr_tensor, random=True):\n with torch.no_grad():\n # only downsampling\n if self.gen_kernel.blur_type == 'iso_gaussian' and self.gen_kernel.sig == 0:\n B, C, H, W = hr_tensor.size()\n hr_blured = hr_tensor.view(-1, C, H, W)\n b_kernels = None\n\n # gaussian blur + downsampling\n else:\n B, C, H, W = hr_tensor.size()\n b_kernels = self.gen_kernel(B, random) # B degradations\n\n # blur\n hr_blured = self.blur(hr_tensor.view(B, -1, H, W), b_kernels)\n hr_blured = hr_blured.view(-1, C, H, W) # B, C, H, W\n\n # downsampling\n if self.mode == 'bicubic':\n lr_blured = self.bicubic(hr_blured, scale=1/self.scale)\n elif self.mode == 's-fold':\n lr_blured = hr_blured.view(-1, C, H//self.scale, self.scale, W//self.scale, self.scale)[:, :, :, 0, :, 0]\n\n\n # add noise\n noise_level = None\n if self.noise > 0:\n _, C, H_lr, W_lr = lr_blured.size()\n noise_level = torch.rand(B, 1, 1, 1).to(lr_blured.device) * self.noise if random else self.noise\n noise = torch.randn_like(lr_blured).view(-1, C, H_lr, W_lr).mul_(noise_level).view(-1, C, H_lr, W_lr)\n lr_blured.add_(noise)\n\n lr_blured = utility.quantize(lr_blured, self.rgb_range)\n\n if isinstance(noise_level, float):\n noise_level = torch.Tensor([noise_level]).view(1,1,1,1).to(lr_blured.device)\n return lr_blured.view(B, C, H//int(self.scale), W//int(self.scale)), [b_kernels,noise_level]\n\n\nclass BicubicPreprocessing(object):\n def __init__(self,\n scale,\n rgb_range=1\n ):\n '''\n # sig, sig_min and sig_max are used for isotropic Gaussian blurs\n During training phase (random=True):\n the width of the blur kernel is randomly selected from [sig_min, sig_max]\n During test phase (random=False):\n the width of the blur kernel is set to sig\n\n # lambda_1, lambda_2, theta, lambda_min and lambda_max are used for anisotropic Gaussian blurs\n During training phase (random=True):\n the eigenvalues of the covariance is randomly selected from [lambda_min, lambda_max]\n the angle value is randomly selected from [0, pi]\n During test phase (random=False):\n the eigenvalues of the covariance are set to lambda_1 and lambda_2\n the angle value is set to theta\n '''\n self.scale = scale\n self.rgb_range=rgb_range\n self.bicubic = bicubic()\n\n def __call__(self, hr_tensor, random=True):\n with torch.no_grad():\n B, C, H, W = hr_tensor.size()\n\n lr = self.bicubic(hr_tensor, scale=1/self.scale)\n lr_bic = self.bicubic(lr, scale=self.scale)\n\n\n lr = utility.quantize(lr, self.rgb_range)\n lr_bic = utility.quantize(lr_bic, self.rgb_range)\n\n return lr.view(B, C, H//int(self.scale), W//int(self.scale)),lr_bic.view(B, C, H, W)\n"
},
{
"path": "KDSR-classic/utils/util2.py",
"content": "import math\nimport numpy as np\nimport utility\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\n\ndef cal_sigma(sig_x, sig_y, radians):\n sig_x = sig_x.view(-1, 1, 1)\n sig_y = sig_y.view(-1, 1, 1)\n radians = radians.view(-1, 1, 1)\n\n D = torch.cat([F.pad(sig_x ** 2, [0, 1, 0, 0]), F.pad(sig_y ** 2, [1, 0, 0, 0])], 1)\n U = torch.cat([torch.cat([radians.cos(), -radians.sin()], 2),\n torch.cat([radians.sin(), radians.cos()], 2)], 1)\n sigma = torch.bmm(U, torch.bmm(D, U.transpose(1, 2)))\n\n return sigma\n\n\ndef anisotropic_gaussian_kernel(batch, kernel_size, covar):\n ax = torch.arange(kernel_size).float().cuda() - kernel_size // 2\n\n xx = ax.repeat(kernel_size).view(1, kernel_size, kernel_size).expand(batch, -1, -1)\n yy = ax.repeat_interleave(kernel_size).view(1, kernel_size, kernel_size).expand(batch, -1, -1)\n xy = torch.stack([xx, yy], -1).view(batch, -1, 2)\n\n inverse_sigma = torch.inverse(covar)\n kernel = torch.exp(- 0.5 * (torch.bmm(xy, inverse_sigma) * xy).sum(2)).view(batch, kernel_size, kernel_size)\n\n return kernel / kernel.sum([1, 2], keepdim=True)\n\n\ndef isotropic_gaussian_kernel(batch, kernel_size, sigma):\n ax = torch.arange(kernel_size).float().cuda() - kernel_size//2\n xx = ax.repeat(kernel_size).view(1, kernel_size, kernel_size).expand(batch, -1, -1)\n yy = ax.repeat_interleave(kernel_size).view(1, kernel_size, kernel_size).expand(batch, -1, -1)\n kernel = torch.exp(-(xx ** 2 + yy ** 2) / (2. * sigma.view(-1, 1, 1) ** 2))\n\n return kernel / kernel.sum([1,2], keepdim=True)\n\n\ndef random_anisotropic_gaussian_kernel(batch=1, kernel_size=21, lambda_min=0.2, lambda_max=4.0):\n theta = torch.rand(batch).cuda() * math.pi\n lambda_1 = torch.rand(batch).cuda() * (lambda_max - lambda_min) + lambda_min\n lambda_2 = torch.rand(batch).cuda() * (lambda_max - lambda_min) + lambda_min\n\n covar = cal_sigma(lambda_1, lambda_2, theta)\n kernel = anisotropic_gaussian_kernel(batch, kernel_size, covar)\n return kernel\n\n\ndef stable_anisotropic_gaussian_kernel(kernel_size=21, theta=0, lambda_1=0.2, lambda_2=4.0):\n theta = torch.ones(1).cuda() * theta / 180 * math.pi\n lambda_1 = torch.ones(1).cuda() * lambda_1\n lambda_2 = torch.ones(1).cuda() * lambda_2\n\n covar = cal_sigma(lambda_1, lambda_2, theta)\n kernel = anisotropic_gaussian_kernel(1, kernel_size, covar)\n return kernel\n\n\ndef random_isotropic_gaussian_kernel(batch=1, kernel_size=21, sig_min=0.2, sig_max=4.0):\n x = torch.rand(batch).cuda() * (sig_max - sig_min) + sig_min\n k = isotropic_gaussian_kernel(batch, kernel_size, x)\n return k\n\n\ndef stable_isotropic_gaussian_kernel(kernel_size=21, sig=4.0):\n x = torch.ones(1).cuda() * sig\n k = isotropic_gaussian_kernel(1, kernel_size, x)\n return k\n\n\ndef random_gaussian_kernel(batch, kernel_size=21, blur_type='iso_gaussian', sig_min=0.2, sig_max=4.0, lambda_min=0.2, lambda_max=4.0):\n if blur_type == 'iso_gaussian':\n return random_isotropic_gaussian_kernel(batch=batch, kernel_size=kernel_size, sig_min=sig_min, sig_max=sig_max)\n elif blur_type == 'aniso_gaussian':\n return random_anisotropic_gaussian_kernel(batch=batch, kernel_size=kernel_size, lambda_min=lambda_min, lambda_max=lambda_max)\n\n\ndef stable_gaussian_kernel(kernel_size=21, blur_type='iso_gaussian', sig=2.6, lambda_1=0.2, lambda_2=4.0, theta=0):\n if blur_type == 'iso_gaussian':\n return stable_isotropic_gaussian_kernel(kernel_size=kernel_size, sig=sig)\n elif blur_type == 'aniso_gaussian':\n return stable_anisotropic_gaussian_kernel(kernel_size=kernel_size, lambda_1=lambda_1, lambda_2=lambda_2, theta=theta)\n\n\n# implementation of matlab bicubic interpolation in pytorch\nclass bicubic(nn.Module):\n def __init__(self):\n super(bicubic, self).__init__()\n\n def cubic(self, x):\n absx = torch.abs(x)\n absx2 = torch.abs(x) * torch.abs(x)\n absx3 = torch.abs(x) * torch.abs(x) * torch.abs(x)\n\n condition1 = (absx <= 1).to(torch.float32)\n condition2 = ((1 < absx) & (absx <= 2)).to(torch.float32)\n\n f = (1.5 * absx3 - 2.5 * absx2 + 1) * condition1 + (-0.5 * absx3 + 2.5 * absx2 - 4 * absx + 2) * condition2\n return f\n\n def contribute(self, in_size, out_size, scale):\n kernel_width = 4\n if scale < 1:\n kernel_width = 4 / scale\n x0 = torch.arange(start=1, end=out_size[0] + 1).to(torch.float32).cuda()\n x1 = torch.arange(start=1, end=out_size[1] + 1).to(torch.float32).cuda()\n\n u0 = x0 / scale + 0.5 * (1 - 1 / scale)\n u1 = x1 / scale + 0.5 * (1 - 1 / scale)\n\n left0 = torch.floor(u0 - kernel_width / 2)\n left1 = torch.floor(u1 - kernel_width / 2)\n\n P = np.ceil(kernel_width) + 2\n\n indice0 = left0.unsqueeze(1) + torch.arange(start=0, end=P).to(torch.float32).unsqueeze(0).cuda()\n indice1 = left1.unsqueeze(1) + torch.arange(start=0, end=P).to(torch.float32).unsqueeze(0).cuda()\n\n mid0 = u0.unsqueeze(1) - indice0.unsqueeze(0)\n mid1 = u1.unsqueeze(1) - indice1.unsqueeze(0)\n\n if scale < 1:\n weight0 = scale * self.cubic(mid0 * scale)\n weight1 = scale * self.cubic(mid1 * scale)\n else:\n weight0 = self.cubic(mid0)\n weight1 = self.cubic(mid1)\n\n weight0 = weight0 / (torch.sum(weight0, 2).unsqueeze(2))\n weight1 = weight1 / (torch.sum(weight1, 2).unsqueeze(2))\n\n indice0 = torch.min(torch.max(torch.FloatTensor([1]).cuda(), indice0), torch.FloatTensor([in_size[0]]).cuda()).unsqueeze(0)\n indice1 = torch.min(torch.max(torch.FloatTensor([1]).cuda(), indice1), torch.FloatTensor([in_size[1]]).cuda()).unsqueeze(0)\n\n kill0 = torch.eq(weight0, 0)[0][0]\n kill1 = torch.eq(weight1, 0)[0][0]\n\n weight0 = weight0[:, :, kill0 == 0]\n weight1 = weight1[:, :, kill1 == 0]\n\n indice0 = indice0[:, :, kill0 == 0]\n indice1 = indice1[:, :, kill1 == 0]\n\n return weight0, weight1, indice0, indice1\n\n def forward(self, input, scale=1/4):\n b, c, h, w = input.shape\n\n weight0, weight1, indice0, indice1 = self.contribute([h, w], [int(h * scale), int(w * scale)], scale)\n weight0 = weight0[0]\n weight1 = weight1[0]\n\n indice0 = indice0[0].long()\n indice1 = indice1[0].long()\n\n out = input[:, :, (indice0 - 1), :] * (weight0.unsqueeze(0).unsqueeze(1).unsqueeze(4))\n out = (torch.sum(out, dim=3))\n A = out.permute(0, 1, 3, 2)\n\n out = A[:, :, (indice1 - 1), :] * (weight1.unsqueeze(0).unsqueeze(1).unsqueeze(4))\n out = out.sum(3).permute(0, 1, 3, 2)\n\n return out\n\n\nclass Gaussin_Kernel(object):\n def __init__(self, kernel_size=21, blur_type='iso_gaussian',\n sig=2.6, sig_min=0.2, sig_max=4.0,\n lambda_1=0.2, lambda_2=4.0, theta=0, lambda_min=0.2, lambda_max=4.0):\n self.kernel_size = kernel_size\n self.blur_type = blur_type\n\n self.sig = sig\n self.sig_min = sig_min\n self.sig_max = sig_max\n\n self.lambda_1 = lambda_1\n self.lambda_2 = lambda_2\n self.theta = theta\n self.lambda_min = lambda_min\n self.lambda_max = lambda_max\n\n def __call__(self, batch, random):\n # random kernel\n if random == True:\n return random_gaussian_kernel(batch, kernel_size=self.kernel_size, blur_type=self.blur_type,\n sig_min=self.sig_min, sig_max=self.sig_max,\n lambda_min=self.lambda_min, lambda_max=self.lambda_max)\n\n # stable kernel\n else:\n return stable_gaussian_kernel(kernel_size=self.kernel_size, blur_type=self.blur_type,\n sig=self.sig,\n lambda_1=self.lambda_1, lambda_2=self.lambda_2, theta=self.theta)\n\nclass BatchBlur(nn.Module):\n def __init__(self, kernel_size=21):\n super(BatchBlur, self).__init__()\n self.kernel_size = kernel_size\n if kernel_size % 2 == 1:\n self.pad = nn.ReflectionPad2d(kernel_size//2)\n else:\n self.pad = nn.ReflectionPad2d((kernel_size//2, kernel_size//2-1, kernel_size//2, kernel_size//2-1))\n\n def forward(self, input, kernel):\n B, C, H, W = input.size()\n input_pad = self.pad(input)\n H_p, W_p = input_pad.size()[-2:]\n\n if len(kernel.size()) == 2:\n input_CBHW = input_pad.view((C * B, 1, H_p, W_p))\n kernel = kernel.contiguous().view((1, 1, self.kernel_size, self.kernel_size))\n\n return F.conv2d(input_CBHW, kernel, padding=0).view((B, C, H, W))\n else:\n input_CBHW = input_pad.view((1, C * B, H_p, W_p))\n kernel = kernel.contiguous().view((B, 1, self.kernel_size, self.kernel_size))\n kernel = kernel.repeat(1, C, 1, 1).view((B * C, 1, self.kernel_size, self.kernel_size))\n\n return F.conv2d(input_CBHW, kernel, groups=B*C).view((B, C, H, W))\n\n\nclass SRMDPreprocessing(object):\n def __init__(self,\n scale,\n mode='bicubic',\n kernel_size=21,\n blur_type='iso_gaussian',\n sig=2.6,\n sig_min=0.2,\n sig_max=4.0,\n lambda_1=0.2,\n lambda_2=4.0,\n theta=0,\n lambda_min=0.2,\n lambda_max=4.0,\n noise=0.0,\n rgb_range=1\n ):\n '''\n # sig, sig_min and sig_max are used for isotropic Gaussian blurs\n During training phase (random=True):\n the width of the blur kernel is randomly selected from [sig_min, sig_max]\n During test phase (random=False):\n the width of the blur kernel is set to sig\n\n # lambda_1, lambda_2, theta, lambda_min and lambda_max are used for anisotropic Gaussian blurs\n During training phase (random=True):\n the eigenvalues of the covariance is randomly selected from [lambda_min, lambda_max]\n the angle value is randomly selected from [0, pi]\n During test phase (random=False):\n the eigenvalues of the covariance are set to lambda_1 and lambda_2\n the angle value is set to theta\n '''\n self.kernel_size = kernel_size\n self.scale = scale\n self.mode = mode\n self.noise = noise / (255 / rgb_range)\n self.rgb_range = rgb_range\n\n self.gen_kernel = Gaussin_Kernel(\n kernel_size=kernel_size, blur_type=blur_type,\n sig=sig, sig_min=sig_min, sig_max=sig_max,\n lambda_1=lambda_1, lambda_2=lambda_2, theta=theta, lambda_min=lambda_min, lambda_max=lambda_max\n )\n self.blur = BatchBlur(kernel_size=kernel_size)\n self.bicubic = bicubic()\n\n def __call__(self, hr_tensor, random=True):\n with torch.no_grad():\n # only downsampling\n if self.gen_kernel.blur_type == 'iso_gaussian' and self.gen_kernel.sig == 0:\n B, C, H, W = hr_tensor.size()\n hr_blured = hr_tensor.view(-1, C, H, W)\n b_kernels = None\n\n # gaussian blur + downsampling\n else:\n B, C, H, W = hr_tensor.size()\n b_kernels = self.gen_kernel(1, random) # B degradations\n b_kernels = b_kernels.expand(B,-1,-1)\n\n # blur\n hr_blured = self.blur(hr_tensor.view(B, -1, H, W), b_kernels)\n hr_blured = hr_blured.view(-1, C, H, W) # B, C, H, W\n\n # downsampling\n if self.mode == 'bicubic':\n lr_blured = self.bicubic(hr_blured, scale=1/self.scale)\n elif self.mode == 's-fold':\n lr_blured = hr_blured.view(-1, C, H//self.scale, self.scale, W//self.scale, self.scale)[:, :, :, 0, :, 0]\n\n\n # add noise\n if self.noise > 0:\n _, C, H_lr, W_lr = lr_blured.size()\n noise_level = torch.rand(1, 1, 1, 1).to(lr_blured.device) * self.noise if random else self.noise\n noise = torch.randn_like(lr_blured).view(-1, C, H_lr, W_lr).mul_(noise_level).view(-1, C, H_lr, W_lr)\n lr_blured.add_(noise)\n\n hr_blured = self.bicubic(lr_blured, scale= self.scale)\n\n\n lr_blured = utility.quantize(lr_blured, self.rgb_range)\n hr_blured = utility.quantize(hr_blured, self.rgb_range)\n\n\n\n return lr_blured.view(B, C, H//int(self.scale), W//int(self.scale)), hr_blured.view(B, C, H, W)\n\n"
},
{
"path": "README.md",
"content": "# Knowledge Distillation based Degradation Estimation for Blind Super-Resolution （ICLR2023）\n\n\n[Paper](https://arxiv.org/pdf/2211.16928.pdf) | [Project Page](https://github.com/Zj-BinXia/KDSR) | [pretrained models](https://drive.google.com/drive/folders/1_LyZDLu5dNIBaCSu7oB9w9d1SPf1pm8c?usp=sharing)\n\n#### News\n\n- **August 28, 2023:** For real-world SR tasks, we released a pretrained model [KDSR-GANV2](https://drive.google.com/file/d/1plvMt7VrOY9YLbWrpchOzi6t1wcqkzBl/view?usp=sharing) and [training files](KDSR-GAN/options/train_kdsrgan_x4STV2.yml) that is more focused on perception rather than distortion.\n \n\n\n- **Jan 28, 2023:** Training&Testing codes and pre-trained models are released!\n\n
\n\n\n> **Abstract:** *Blind image super-resolution (Blind-SR) aims to recover a high-resolution (HR) image from its corresponding low-resolution (LR) input image with unknown degradations. Most of the existing works design an explicit degradation estimator for each degradation to guide SR. However, it is infeasible to provide concrete labels of multiple degradation combinations (\\eg, blur, noise, jpeg compression) to supervise the degradation estimator training. In addition, these special designs for certain degradation, such as blur, impedes the models from being generalized to handle different degradations. To this end, it is necessary to design an implicit degradation estimator that can extract discriminative degradation representation for all degradations without relying on the supervision of degradation ground-truth. In this paper, we propose a Knowledge Distillation based Blind-SR network (KDSR). It consists of a knowledge distillation based implicit degradation estimator network (KD-IDE) and an efficient SR network. To learn the KDSR model, we first train a teacher network: KD-IDE$_{T}$. It takes paired HR and LR patches as inputs and is optimized with the SR network jointly. Then, we further train a student network KD-IDE$_{S}$, which only takes LR images as input and learns to extract the same implicit degradation representation (IDR) as KD-IDE$_{T}$. In addition, to fully use extracted IDR, we design a simple, strong, and efficient IDR based dynamic convolution residual block (IDR-DCRB) to build an SR network. We conduct extensive experiments under classic and real-world degradation settings. The results show that KDSR achieves SOTA performance and can generalize to various degradation processes.* \n\n\n\n \n
\n\n\n\n---\n\n## Dependencies and Installation\n\n- Python >= 3.8 (Recommend to use [Anaconda](https://www.anaconda.com/download/#linux) or [Miniconda](https://docs.conda.io/en/latest/miniconda.html))\n- [PyTorch >= 1.10](https://pytorch.org/)\n\n\n### Installation\n\n1. Clone repo\n\n ```bash\n git clone git@github.com:Zj-BinXia/KDSR.git\n ```\n\n2. If you want to train or test KDSR-GAN (ie, Real-world SR, trained with the same degradation model as Real-ESRGAN)\n\n ```bash\n cd KDSR-GAN\n ```\n \n3. If you want to train or test KDSR-classic (ie, classic degradation models, trained with the isotropic Gaussian Blur or anisotropic Gaussian blur and noises)\n\n ```bash\n cd KDSR-classic\n ```\n\n**More details please see the README in folder of KDSR-GAN and KDSR-classic** \n\n---\n## BibTeX\n\n @InProceedings{xia2022knowledge,\n title={Knowledge Distillation based Degradation Estimation for Blind Super-Resolution},\n author={Xia, Bin and Zhang, Yulun and Wang, Yitong and Tian, Yapeng and Yang, Wenming and Timofte, Radu and Van Gool, Luc},\n journal={ICLR},\n year={2023}\n }\n\n## 📧 Contact\n\nIf you have any question, please email `zjbinxia@gmail.com`.\n\n"
}
]