[ { "path": ".gitignore", "content": "# 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": "LICENSE", "content": " Apache License\n Version 2.0, January 2004\n http://www.apache.org/licenses/\n\n TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION\n\n 1. Definitions.\n\n \"License\" shall mean the terms and conditions for use, reproduction,\n and distribution as defined by Sections 1 through 9 of this document.\n\n \"Licensor\" shall mean the copyright owner or entity authorized by\n the copyright owner that is granting the License.\n\n \"Legal Entity\" shall mean the union of the acting entity and all\n other entities that control, are controlled by, or are under common\n control with that entity. For the purposes of this definition,\n \"control\" means (i) the power, direct or indirect, to cause the\n direction or management of such entity, whether by contract or\n otherwise, or (ii) ownership of fifty percent (50%) or more of the\n outstanding shares, or (iii) beneficial ownership of such entity.\n\n \"You\" (or \"Your\") shall mean an individual or Legal Entity\n exercising permissions granted by this License.\n\n \"Source\" form shall mean the preferred form for making modifications,\n including but not limited to software source code, documentation\n source, and configuration files.\n\n \"Object\" form shall mean any form resulting from mechanical\n transformation or translation of a Source form, including but\n not limited to compiled object code, generated documentation,\n and conversions to other media types.\n\n \"Work\" shall mean the work of authorship, whether in Source or\n Object form, made available under the License, as indicated by a\n copyright notice that is included in or attached to the work\n (an example is provided in the Appendix below).\n\n \"Derivative Works\" shall mean any work, whether in Source or Object\n form, that is based on (or derived from) the Work and for which the\n editorial revisions, annotations, elaborations, or other modifications\n represent, as a whole, an original work of authorship. For the purposes\n of this License, Derivative Works shall not include works that remain\n separable from, or merely link (or bind by name) to the interfaces of,\n the Work and Derivative Works thereof.\n\n \"Contribution\" shall mean any work of authorship, including\n the original version of the Work and any modifications or additions\n to that Work or Derivative Works thereof, that is intentionally\n submitted to Licensor for inclusion in the Work by the copyright owner\n or by an individual or Legal Entity authorized to submit on behalf of\n the copyright owner. For the purposes of this definition, \"submitted\"\n means any form of electronic, verbal, or written communication sent\n to the Licensor or its representatives, including but not limited to\n communication on electronic mailing lists, source code control systems,\n and issue tracking systems that are managed by, or on behalf of, the\n Licensor for the purpose of discussing and improving the Work, but\n excluding communication that is conspicuously marked or otherwise\n designated in writing by the copyright owner as \"Not a Contribution.\"\n\n \"Contributor\" shall mean Licensor and any individual or Legal Entity\n on behalf of whom a Contribution has been received by Licensor and\n subsequently incorporated within the Work.\n\n 2. Grant of Copyright License. Subject to the terms and conditions of\n this License, each Contributor hereby grants to You a perpetual,\n worldwide, non-exclusive, no-charge, royalty-free, irrevocable\n copyright license to reproduce, prepare Derivative Works of,\n publicly display, publicly perform, sublicense, and distribute the\n Work and such Derivative Works in Source or Object form.\n\n 3. Grant of Patent License. Subject to the terms and conditions of\n this License, each Contributor hereby grants to You a perpetual,\n worldwide, non-exclusive, no-charge, royalty-free, irrevocable\n (except as stated in this section) patent license to make, have made,\n use, offer to sell, sell, import, and otherwise transfer the Work,\n where such license applies only to those patent claims licensable\n by such Contributor that are necessarily infringed by their\n Contribution(s) alone or by combination of their Contribution(s)\n with the Work to which such Contribution(s) was submitted. If You\n institute patent litigation against any entity (including a\n cross-claim or counterclaim in a lawsuit) alleging that the Work\n or a Contribution incorporated within the Work constitutes direct\n or contributory patent infringement, then any patent licenses\n granted to You under this License for that Work shall terminate\n as of the date such litigation is filed.\n\n 4. Redistribution. You may reproduce and distribute copies of the\n Work or Derivative Works thereof in any medium, with or without\n modifications, and in Source or Object form, provided that You\n meet the following conditions:\n\n (a) You must give any other recipients of the Work or\n Derivative Works a copy of this License; and\n\n (b) You must cause any modified files to carry prominent notices\n stating that You changed the files; and\n\n (c) You must retain, in the Source form of any Derivative Works\n that You distribute, all copyright, patent, trademark, and\n attribution notices from the Source form of the Work,\n excluding those notices that do not pertain to any part of\n the Derivative Works; and\n\n (d) If the Work includes a \"NOTICE\" text file as part of its\n distribution, then any Derivative Works that You distribute must\n include a readable copy of the attribution notices contained\n within such NOTICE file, excluding those notices that do not\n pertain to any part of the Derivative Works, in at least one\n of the following places: within a NOTICE text file distributed\n as part of the Derivative Works; within the Source form or\n documentation, if provided along with the Derivative Works; or,\n within a display generated by the Derivative Works, if and\n wherever such third-party notices normally appear. The contents\n of the NOTICE file are for informational purposes only and\n do not modify the License. You may add Your own attribution\n notices within Derivative Works that You distribute, alongside\n or as an addendum to the NOTICE text from the Work, provided\n that such additional attribution notices cannot be construed\n as modifying the License.\n\n You may add Your own copyright statement to Your modifications and\n may provide additional or different license terms and conditions\n for use, reproduction, or distribution of Your modifications, or\n for any such Derivative Works as a whole, provided Your use,\n reproduction, and distribution of the Work otherwise complies with\n the conditions stated in this License.\n\n 5. Submission of Contributions. Unless You explicitly state otherwise,\n any Contribution intentionally submitted for inclusion in the Work\n by You to the Licensor shall be under the terms and conditions of\n this License, without any additional terms or conditions.\n Notwithstanding the above, nothing herein shall supersede or modify\n the terms of any separate license agreement you may have executed\n with Licensor regarding such Contributions.\n\n 6. Trademarks. This License does not grant permission to use the trade\n names, trademarks, service marks, or product names of the Licensor,\n except as required for reasonable and customary use in describing the\n origin of the Work and reproducing the content of the NOTICE file.\n\n 7. Disclaimer of Warranty. Unless required by applicable law or\n agreed to in writing, Licensor provides the Work (and each\n Contributor provides its Contributions) on an \"AS IS\" BASIS,\n WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or\n implied, including, without limitation, any warranties or conditions\n of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A\n PARTICULAR PURPOSE. You are solely responsible for determining the\n appropriateness of using or redistributing the Work and assume any\n risks associated with Your exercise of permissions under this License.\n\n 8. Limitation of Liability. In no event and under no legal theory,\n whether in tort (including negligence), contract, or otherwise,\n unless required by applicable law (such as deliberate and grossly\n negligent acts) or agreed to in writing, shall any Contributor be\n liable to You for damages, including any direct, indirect, special,\n incidental, or consequential damages of any character arising as a\n result of this License or out of the use or inability to use the\n Work (including but not limited to damages for loss of goodwill,\n work stoppage, computer failure or malfunction, or any and all\n other commercial damages or losses), even if such Contributor\n has been advised of the possibility of such damages.\n\n 9. Accepting Warranty or Additional Liability. While redistributing\n the Work or Derivative Works thereof, You may choose to offer,\n and charge a fee for, acceptance of support, warranty, indemnity,\n or other liability obligations and/or rights consistent with this\n License. However, in accepting such obligations, You may act only\n on Your own behalf and on Your sole responsibility, not on behalf\n of any other Contributor, and only if You agree to indemnify,\n defend, and hold each Contributor harmless for any liability\n incurred by, or claims asserted against, such Contributor by reason\n of your accepting any such warranty or additional liability.\n\n END OF TERMS AND CONDITIONS\n\n APPENDIX: How to apply the Apache License to your work.\n\n To apply the Apache License to your work, attach the following\n boilerplate notice, with the fields enclosed by brackets \"[]\"\n replaced with your own identifying information. (Don't include\n the brackets!) The text should be enclosed in the appropriate\n comment syntax for the file format. We also recommend that a\n file or class name and description of purpose be included on the\n same \"printed page\" as the copyright notice for easier\n identification within third-party archives.\n\n Copyright [yyyy] [name of copyright owner]\n\n Licensed under the Apache License, Version 2.0 (the \"License\");\n you may not use this file except in compliance with the License.\n You may obtain a copy of the License at\n\n http://www.apache.org/licenses/LICENSE-2.0\n\n Unless required by applicable law or agreed to in writing, software\n distributed under the License is distributed on an \"AS IS\" BASIS,\n WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n See the License for the specific language governing permissions and\n limitations under the License.\n" }, { "path": "README.md", "content": "### pytorch 图像分类竞赛框架\n\n### 1. 更新日志\n- (2020年5月2日) 基础版本上线\n\n### 2. 依赖库\n- pretrainedmodels\n- progress\n- efficientnet-pytorch\n- apex\n\n### 3. 支持功能\n\n- [x] pytorch官网模型\n- [x] [pretrained-models.pytorch](https://github.com/Cadene/pretrained-models.pytorch) 复现的部分模型\n- [x] [EfficientNet-PyTorch](https://github.com/lukemelas/EfficientNet-PyTorch) \n- [x] fp16混合精度训练\n- [x] TTA\n- [x] 固定验证集/随机划分验证集\n- [x] 多种优化器：adam、radam、novograd、sgd、ranger、ralamb、over9000、lookahead、lamb\n- [x] OneCycle训练策略\n- [x] LabelSmoothLoss\n- [x] Focal Loss\n- [ ] AotuAgument\n \n### 4. 使用方法\n更改`config.py`中的参数，训练执行 `python main.py`，预测执行`python test.py`\n\n### 5. submit_example.csv \n每一行：filename,label\n样例：\n```\n0001.jpg,dog\n0002.jpg,dog\n0003.jpg,dog\n```\n注：预测图像可能没有label，所以label可以随意给个临时的，但一些比赛平台对都会给个提交样例，随意给个label\n### 6.TODO\n\n- [ ] 优化模型融合策略\n- [ ] 优化online数据增强\n- [ ] 优化pytorch官方模型调用接口\n- [ ] 增加模型全连接层初始化\n- [ ] 增加更多学习率衰减策略\n- [ ] 增加find lr\n- [ ] 增加dali\n- [ ] 增加wsl模型\n- [ ] 增加tensorboardX\n- [ ] 优化文件夹创建\n" }, { "path": "config.py", "content": "class DefaultConfigs(object):\n # set default configs, if you don't understand, don't modify\n seed = 666 # set random seed\n workers = 4 # set number of data loading workers (default: 4)\n beta1 = 0.9 # adam parameters beta1\n beta2 = 0.999 # adam parameters beta2\n mom = 0.9 # momentum parameters\n wd = 1e-4 # weight-decay\n resume = None # path to latest checkpoint (default: none),should endswith \".pth\" or \".tar\" if used\n evaluate = False # just do evaluate\n start_epoch = 0 # deault start epoch is zero,if use resume change it\n split_online = False # split dataset to train and val online or offline\n\n # set changeable configs, you can change one during your experiment\n dataset = \"/dataset/df/cloud/data/dataset/\" # dataset folder with train and val\n test_folder = \"/dataset/df/cloud/data/test/\" # test images' folder\n submit_example = \"/dataset/df/cloud/data/submit_example.csv\" # submit example file\n checkpoints = \"./checkpoints/\" # path to save checkpoints\n log_dir = \"./logs/\" # path to save log files\n submits = \"./submits/\" # path to save submission files\n bs = 32 # batch size\n lr = 2e-3 # learning rate\n epochs = 40 # train epochs\n input_size = 512 # model input size or image resied\n num_classes = 9 # num of classes\n gpu_id = \"0\" # default gpu id\n model_name = \"se_resnext50_32x4d-model-sgd-512\" # model name to use\n optim = \"sgd\" # \"adam\",\"radam\",\"novograd\",sgd\",\"ranger\",\"ralamb\",\"over9000\",\"lookahead\",\"lamb\"\n fp16 = True # use float16 to train the model\n opt_level = \"O1\" # if use fp16, \"O0\" means fp32，\"O1\" means mixed，\"O2\" means except BN，\"O3\" means only fp16\n keep_batchnorm_fp32 = False # if use fp16,keep BN layer as fp32\n loss_func = \"CrossEntropy\" # \"CrossEntropy\"、\"FocalLoss\"、\"LabelSmoothCE\"\n lr_scheduler = \"step\" # lr scheduler method,\"adjust\",\"on_loss\",\"on_acc\",\"step\"\n\n \nconfigs = DefaultConfigs()\n" }, { "path": "ensemble.py", "content": "import pandas as pd \nimport numpy as np \nimport os\nfrom IPython import embed\n\nfile1 = pd.read_csv(\"./csvs/efficientnet-b3-model_512-_adam_aug_confidence.csv\",header=None)\nfile2 = pd.read_csv(\"./csvs/efficientnet-b5-model_456_ranger_aug_confidence.csv\",header=None)\nfile3 = pd.read_csv(\"./csvs/efficientnet-b4-model_380_ranger_aug_confidence.csv\",header=None)\n\nfilenames,labels = [],[]\n# embed()\n# for (filename1,label1),(filename2,label2),(filename3,label3),(filename4,label4),(filename5,label5) in zip(file1.values,file2.values,file3.values,file4.values,file5.values):\nfor (filename1,label1) ,(filename2,label2),(filename3,label3) in zip(file1.values,file2.values,file3.values):\n filename = filename1\n filenames.append(filename)\n #embed()\n label1 = np.array(list(map(float,label1.split(\"-\"))))\n label2 = np.array(list(map(float,label2.split(\"-\"))))\n label3 = np.array(list(map(float,label3.split(\"-\"))))\n # label4 = np.array(list(map(float,label4.split(\"[\")[1].split(\"]\")[0].split(\",\"))))\n # label5 = np.array(list(map(float,label5.split(\"[\")[1].split(\"]\")[0].split(\",\"))))\n label = np.argmax((label1 + label2 + label3) / 3.0) + 1\n labels.append(label)\n\nsubmission = pd.DataFrame({'FileName': filenames, 'type': labels})\nsubmission.to_csv(\"./ensemble_efficientnets.csv\", header=None, index=False)\n\n" }, { "path": "main.py", "content": "import random\nimport time\nimport warnings\n\nimport torch.nn as nn\nimport torch.nn.parallel\nimport torch.backends.cudnn as cudnn\nimport torch.optim\nimport torch.utils.data\nimport torch.utils.data.distributed\nimport torchvision.transforms as transforms\nimport torchvision.datasets as datasets\nimport numpy as np\nfrom PIL import ImageFile\nfrom config import configs\nfrom models.model import get_model\nfrom sklearn.model_selection import train_test_split\nfrom utils.misc import *\nfrom utils.logger import *\nfrom utils.losses import *\nfrom progress.bar import Bar\nfrom utils.reader import WeatherDataset\n\n# for train fp16\nif configs.fp16:\n try:\n import apex\n from apex.parallel import DistributedDataParallel as DDP\n from apex.fp16_utils import *\n from apex import amp, optimizers\n from apex.multi_tensor_apply import multi_tensor_applier\n except ImportError:\n raise ImportError(\"Please install apex from https://www.github.com/nvidia/apex to run this example.\")\n\nImageFile.LOAD_TRUNCATED_IMAGES = True\nwarnings.filterwarnings(\"ignore\")\nos.environ['CUDA_VISIBLE_DEVICES'] = configs.gpu_id\n\n# set random seed\ndef seed_everything(seed):\n random.seed(seed)\n os.environ['PYTHONHASHSEED'] = str(seed)\n np.random.seed(seed)\n torch.manual_seed(seed)\n torch.cuda.manual_seed(seed)\n torch.backends.cudnn.deterministic = True\nseed_everything(configs.seed)\n\n# make dir for use\ndef makdir():\n if not os.path.exists(configs.checkpoints):\n os.makedirs(configs.checkpoints)\n if not os.path.exists(configs.log_dir):\n os.makedirs(configs.log_dir)\n if not os.path.exists(configs.submits):\n os.makedirs(configs.submits)\nmakdir()\n\nbest_acc = 0 # best test accuracy\nbest_loss = 999 # lower loss\n\ndef main():\n global best_acc\n global best_loss\n start_epoch = configs.start_epoch\n # set normalize configs for imagenet\n normalize_imgnet = transforms.Normalize(mean=[0.485, 0.456, 0.406],\n std=[0.229, 0.224, 0.225])\n \n transform_train = transforms.Compose([\n transforms.RandomResizedCrop(configs.input_size),\n transforms.RandomHorizontalFlip(p=0.5),\n transforms.RandomVerticalFlip(p=0.5),\n transforms.ToTensor(),\n normalize_imgnet\n ])\n \n transform_val = transforms.Compose([\n transforms.Resize(int(configs.input_size * 1.2)),\n transforms.CenterCrop(configs.input_size),\n transforms.ToTensor(),\n normalize_imgnet\n ])\n\n # Data loading code\n if configs.split_online:\n # use online random split dataset method\n total_files = get_files(configs.dataset,\"train\")\n train_files,val_files = train_test_split(total_files,test_size = 0.1,stratify=total_files[\"label\"])\n train_dataset = WeatherDataset(train_files,transform_train)\n val_dataset = WeatherDataset(val_files,transform_val)\n else:\n # use offline split dataset\n train_files = get_files(configs.dataset+\"/train/\",\"train\")\n val_files = get_files(configs.dataset+\"/val/\",\"train\")\n train_dataset = WeatherDataset(train_files,transform_train)\n val_dataset = WeatherDataset(val_files,transform_val)\n train_loader = torch.utils.data.DataLoader(\n train_dataset, batch_size=configs.bs, shuffle=True,\n num_workers=configs.workers, pin_memory=True,\n )\n val_loader = torch.utils.data.DataLoader(\n val_dataset, batch_size=configs.bs, shuffle=False,\n num_workers=configs.workers, pin_memory=True\n ) \n # get model\n model = get_model()\n model.cuda()\n # choose loss func,default is CE\n if configs.loss_func == \"LabelSmoothCE\":\n criterion = LabelSmoothingLoss(0.1, configs.num_classes).cuda()\n elif configs.loss_func == \"CrossEntropy\":\n criterion = nn.CrossEntropyLoss().cuda()\n elif configs.loss_func == \"FocalLoss\":\n criterion = FocalLoss(gamma=2).cuda()\n else:\n criterion = nn.CrossEntropyLoss().cuda()\n optimizer = get_optimizer(model)\n # set lr scheduler method\n if configs.lr_scheduler == \"step\":\n scheduler = torch.optim.lr_scheduler.StepLR(optimizer,step_size=10,gamma=0.1)\n elif configs.lr_scheduler == \"on_loss\":\n scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.2, patience=5, verbose=False)\n elif configs.lr_scheduler == \"on_acc\":\n scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='max', factor=0.2, patience=5, verbose=False)\n else:\n scheduler = torch.optim.lr_scheduler.StepLR(optimizer,step_size=6,gamma=0.1)\n # for fp16\n if configs.fp16:\n model, optimizer = amp.initialize(model, optimizer,\n opt_level=configs.opt_level,\n keep_batchnorm_fp32= None if configs.opt_level == \"O1\" else configs.keep_batchnorm_fp32\n )\n if configs.resume:\n # Load checkpoint.\n print('==> Resuming from checkpoint..')\n assert os.path.isfile(configs.resume), 'Error: no checkpoint directory found!'\n configs.checkpoint = os.path.dirname(configs.resume)\n checkpoint = torch.load(configs.resume)\n best_acc = checkpoint['best_acc']\n start_epoch = checkpoint['epoch']\n model.module.load_state_dict(checkpoint['state_dict'])\n optimizer.load_state_dict(checkpoint['optimizer'])\n logger = Logger(os.path.join(configs.log_dir, '%s_log.txt'%configs.model_name), title=configs.model_name, resume=True)\n else:\n logger = Logger(os.path.join(configs.log_dir, '%s_log.txt'%configs.model_name), title=configs.model_name)\n logger.set_names(['Learning Rate', 'Train Loss', 'Valid Loss', 'Train Acc.', 'Valid Acc.'])\n if configs.evaluate:\n print('\\nEvaluation only')\n val_loss, val_acc = validate(val_loader, model, criterion, start_epoch)\n print(' Test Loss: %.8f, Test Acc: %.2f' % (val_loss, val_acc))\n return\n\n # Train and val\n for epoch in range(start_epoch, configs.epochs):\n print('\\nEpoch: [%d | %d] LR: %f' % (epoch + 1, configs.epochs, optimizer.param_groups[0]['lr']))\n\n train_loss, train_acc, train_5 = train(train_loader, model, criterion, optimizer, epoch)\n val_loss, val_acc, test_5 = validate(val_loader, model, criterion, epoch)\n # adjust lr\n if configs.lr_scheduler == \"on_loss\":\n scheduler.step(val_loss)\n elif configs.lr_scheduler == \"on_acc\":\n scheduler.step(val_acc)\n elif configs.lr_scheduler == \"step\":\n scheduler.step(epoch)\n elif configs.lr_scheduler == \"adjust\":\n adjust_learning_rate(optimizer,epoch)\n else:\n scheduler.step(epoch)\n # append logger file\n lr_current = get_lr(optimizer)\n logger.append([lr_current,train_loss, val_loss, train_acc, val_acc])\n print('train_loss:%f, val_loss:%f, train_acc:%f, train_5:%f, val_acc:%f, val_5:%f' % (train_loss, val_loss, train_acc, train_5, val_acc, test_5))\n\n # save model\n is_best = val_acc > best_acc\n is_best_loss = val_loss < best_loss\n best_acc = max(val_acc, best_acc)\n best_loss = min(val_loss,best_loss)\n\n save_checkpoint({\n 'fold': 0,\n 'epoch': epoch + 1,\n 'state_dict': model.state_dict(),\n 'train_acc': train_acc,\n 'acc': val_acc,\n 'best_acc': best_acc,\n 'best_loss': best_loss,\n 'optimizer': optimizer.state_dict(),\n }, is_best,is_best_loss)\n\n logger.close()\n print('Best acc:')\n print(best_acc)\ndef train(train_loader, model, criterion, optimizer, epoch):\n # switch to train mode\n model.train()\n\n batch_time = AverageMeter()\n data_time = AverageMeter()\n losses = AverageMeter()\n top1 = AverageMeter()\n top5 = AverageMeter()\n end = time.time()\n\n bar = Bar('Training: ', max=len(train_loader))\n for batch_idx, (inputs, targets) in enumerate(train_loader):\n # measure data loading time\n data_time.update(time.time() - end)\n inputs, targets = inputs.cuda(), targets.cuda()\n inputs, targets = torch.autograd.Variable(inputs), torch.autograd.Variable(targets)\n\n # compute output\n outputs = model(inputs)\n loss = criterion(outputs, targets)\n\n # measure accuracy and record loss\n prec1, prec5 = accuracy(outputs.data, targets.data, topk=(1, 5))\n losses.update(loss.item(), inputs.size(0))\n top1.update(prec1.item(), inputs.size(0))\n top5.update(prec5.item(), inputs.size(0))\n\n # compute gradient and do SGD step\n optimizer.zero_grad()\n if configs.fp16:\n with amp.scale_loss(loss, optimizer) as scaled_loss:\n scaled_loss.backward()\n else:\n loss.backward()\n # clip gradient\n torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=5.0, norm_type=2)\n optimizer.step()\n\n # measure elapsed time\n batch_time.update(time.time() - end)\n end = time.time()\n\n # plot progress\n bar.suffix = '({batch}/{size}) Data: {data:.3f}s | Batch: {bt:.3f}s | Total: {total:} | ETA: {eta:} | Loss: {loss:.4f} | top1: {top1: .4f} | top5: {top5: .4f}'.format(\n batch=batch_idx + 1,\n size=len(train_loader),\n data=data_time.val,\n bt=batch_time.val,\n total=bar.elapsed_td,\n eta=bar.eta_td,\n loss=losses.avg,\n top1=top1.avg,\n top5=top5.avg,\n )\n bar.next()\n bar.finish()\n return (losses.avg, top1.avg, top5.avg)\n\ndef validate(val_loader, model, criterion, epoch):\n global best_acc\n\n batch_time = AverageMeter()\n data_time = AverageMeter()\n losses = AverageMeter()\n top1 = AverageMeter()\n top5 = AverageMeter()\n\n # switch to evaluate mode\n model.eval()\n\n end = time.time()\n bar = Bar('Validating: ', max=len(val_loader))\n with torch.no_grad():\n for batch_idx, (inputs, targets) in enumerate(val_loader):\n # measure data loading time\n data_time.update(time.time() - end)\n\n inputs, targets = inputs.cuda(), targets.cuda()\n inputs, targets = torch.autograd.Variable(inputs), torch.autograd.Variable(targets)\n\n # compute output\n outputs = model(inputs)\n loss = criterion(outputs, targets)\n\n # measure accuracy and record loss\n prec1, prec5 = accuracy(outputs.data, targets.data, topk=(1, 5))\n losses.update(loss.item(), inputs.size(0))\n top1.update(prec1.item(), inputs.size(0))\n top5.update(prec5.item(), inputs.size(0))\n\n # measure elapsed time\n batch_time.update(time.time() - end)\n end = time.time()\n\n # plot progress\n bar.suffix = '({batch}/{size}) Data: {data:.3f}s | Batch: {bt:.3f}s | Total: {total:} | ETA: {eta:} | Loss: {loss:.4f} | top1: {top1: .4f} | top5: {top5: .4f}'.format(\n batch=batch_idx + 1,\n size=len(val_loader),\n data=data_time.avg,\n bt=batch_time.avg,\n total=bar.elapsed_td,\n eta=bar.eta_td,\n loss=losses.avg,\n top1=top1.avg,\n top5=top5.avg,\n )\n bar.next()\n bar.finish()\n return (losses.avg, top1.avg, top5.avg)\n\nif __name__ == '__main__':\n main()\n" }, { "path": "models/__init__.py", "content": "from .model import *" }, { "path": "models/model.py", "content": "from pretrainedmodels import models as pm\nimport pretrainedmodels\nfrom torch import nn\nfrom torchvision import models as tm\nfrom config import configs\nfrom efficientnet_pytorch import EfficientNet\nimport torch\nfrom torch.nn.parameter import Parameter\nimport torch.nn.functional as F\nfrom torch.nn.parameter import Parameter\n\nweights = {\n \"efficientnet-b3\":\"/data/dataset/detection/pretrainedmodels/efficientnet-b3-c8376fa2.pth\",\n \"efficientnet-b4\":\"/data/dataset/detection/pretrainedmodels/efficientnet-b4-6ed6700e.pth\",\n \"efficientnet-b5\":\"/data/dataset/detection/pretrainedmodels/efficientnet-b5-b6417697.pth\",\n \"efficientnet-b6\":\"/data/dataset/detection/pretrainedmodels/efficientnet-b6-c76e70fd.pth\",\n }\n\ndef gem(x, p=3, eps=1e-6):\n return F.avg_pool2d(x.clamp(min=eps).pow(p), (x.size(-2), x.size(-1))).pow(1./p)\n\nclass GeM(nn.Module):\n def __init__(self, p=3, eps=1e-6):\n super(GeM,self).__init__()\n self.p = Parameter(torch.ones(1)*p)\n self.eps = eps\n def forward(self, x):\n return gem(x, p=self.p, eps=self.eps) \n def __repr__(self):\n return self.__class__.__name__ + '(' + 'p=' + '{:.4f}'.format(self.p.data.tolist()[0]) + ', ' + 'eps=' + str(self.eps) + ')'\n\ndef get_model():\n if configs.model_name.startswith(\"resnext50_32x4d\"):\n model = tm.resnext50_32x4d(pretrained=True)\n model.avgpool = nn.AdaptiveAvgPool2d(1)\n model.fc = nn.Linear(2048,configs.num_classes)\n model.cuda()\n elif configs.model_name.startswith(\"efficient\"):\n # efficientNet\n model_name = configs.model_name[:15]\n model = EfficientNet.from_name(model_name)\n model.load_state_dict(torch.load(weights[model_name]))\n in_features = model._fc.in_features\n model._fc = nn.Sequential(\n nn.BatchNorm1d(in_features),\n nn.Dropout(0.5),\n nn.Linear(in_features, configs.num_classes),\n )\n model.cuda()\n else:\n pretrained = \"imagenet+5k\" if configs.model_name.startswith(\"dpn\") else \"imagenet\"\n model = pretrainedmodels.__dict__[configs.model_name.split(\"-model\")[0]](num_classes=1000, pretrained=pretrained)\n if configs.model_name.startswith(\"pnasnet\"):\n model.last_linear = nn.Linear(4320, configs.num_classes)\n model.avg_pool = nn.AdaptiveAvgPool2d(1)\n elif configs.model_name.startswith(\"inception\"):\n model.last_linear = nn.Linear(1536, configs.num_classes)\n model.avgpool_1a = nn.AdaptiveAvgPool2d(1) \n else:\n model.last_linear = nn.Linear(2048, configs.num_classes)\n model.avg_pool = nn.AdaptiveAvgPool2d(1) \n \n model.cuda()\n return model" }, { "path": "test.py", "content": "import os\nimport torch\nimport warnings\nimport pandas as pd\nimport numpy as np\nimport torch.backends.cudnn as cudnn\nfrom tqdm import tqdm\nfrom glob import glob\nfrom PIL import Image,ImageFile\nfrom config import configs\nfrom models.model import get_model\nfrom torch.utils.data import Dataset,DataLoader\nfrom torchvision import transforms\nfrom utils.misc import get_files\nfrom IPython import embed\n\nImageFile.LOAD_TRUNCATED_IMAGES = True\nwarnings.filterwarnings(\"ignore\")\nos.environ['CUDA_VISIBLE_DEVICES'] = configs.gpu_id\nlen_data = 0\n\nclass WeatherTTADataset(Dataset):\n def __init__(self,labels_file,aug):\n imgs = []\n for index, row in labels_file.iterrows():\n imgs.append((row[\"FileName\"],row[\"type\"]))\n self.imgs = imgs\n self.length = len(imgs)\n global len_data\n len_data = self.length\n self.aug = aug\n self.Hflip = transforms.RandomHorizontalFlip(p=1)\n self.Vflip = transforms.RandomVerticalFlip(p=1)\n self.Rotate = transforms.functional.rotate\n self.resize = transforms.Resize((configs.input_size,configs.input_size))\n self.randomCrop = transforms.Compose([transforms.Resize(int(configs.input_size * 1.2)),\n transforms.CenterCrop(configs.input_size),\n ])\n def __getitem__(self,index):\n filename,label_tmp = self.imgs[index]\n img = Image.open(configs.test_folder + os.sep + filename).convert('RGB')\n img = self.transform_(img,self.aug)\n return img,filename\n\n def __len__(self):\n return self.length\n def transform_(self,data_torch,aug):\n if aug == 'Ori':\n data_torch = data_torch\n data_torch = self.resize(data_torch)\n if aug == 'Ori_Hflip':\n data_torch = self.Hflip(data_torch)\n data_torch = self.resize(data_torch)\n if aug == 'Ori_Vflip':\n data_torch = self.Vflip(data_torch)\n data_torch = self.resize(data_torch)\n if aug == 'Ori_Rotate_90':\n data_torch = self.Rotate(data_torch, 90)\n data_torch = self.resize(data_torch)\n if aug == 'Ori_Rotate_180':\n data_torch = self.Rotate(data_torch, 180)\n data_torch = self.resize(data_torch)\n if aug == 'Ori_Rotate_270':\n data_torch = self.Rotate(data_torch, 270)\n data_torch = self.resize(data_torch)\n if aug == 'Crop':\n # print(data_torch.size)\n data_torch = self.randomCrop(data_torch)\n data_torch = data_torch\n if aug == 'Crop_Hflip':\n data_torch = self.randomCrop(data_torch)\n data_torch = self.Hflip(data_torch)\n if aug == 'Crop_Vflip':\n data_torch = self.randomCrop(data_torch)\n data_torch = self.Vflip(data_torch)\n if aug == 'Crop_Rotate_90':\n data_torch = self.randomCrop(data_torch)\n data_torch = self.Rotate(data_torch, 90)\n if aug == 'Crop_Rotate_180':\n data_torch = self.randomCrop(data_torch)\n data_torch = self.Rotate(data_torch, 180)\n if aug == 'Crop_Rotate_270':\n data_torch = self.randomCrop(data_torch)\n data_torch = self.Rotate(data_torch, 270)\n data_torch = transforms.ToTensor()(data_torch)\n data_torch = transforms.Normalize(mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225])(data_torch)\n return data_torch\n\n#aug = ['Ori','Ori_Hflip','Ori_Vflip','Ori_Rotate_90','Ori_Rotate_180','Ori_Rotate_270',\n # 'Crop','Crop_Hflip','Crop_Vflip','Crop_Rotate_90','Crop_Rotate_180','Crop_Rotate_270']\naug = ['Ori_Hflip']\n\ncpk_filename = configs.checkpoints + os.sep + configs.model_name + \"-checkpoint.pth.tar\"\nbest_cpk = cpk_filename.replace(\"-checkpoint.pth.tar\",\"-best_model.pth.tar\")\ncheckpoint = torch.load(best_cpk)\ncudnn.benchmark = True\nmodel = get_model()\nmodel.load_state_dict(checkpoint['state_dict'])\nmodel.eval()\ntest_files = pd.read_csv(configs.submit_example)\n\nwith torch.no_grad():\n y_pred_prob = torch.FloatTensor([])\n for a in tqdm(aug):\n print(a)\n test_set = WeatherTTADataset(test_files, a)\n test_loader = DataLoader(dataset=test_set, batch_size=configs.bs, shuffle=False,\n num_workers=4, pin_memory=True, sampler=None)\n total = 0\n correct = 0\n for inputs, labels in tqdm(test_loader):\n inputs = inputs.cuda()\n outputs = model(inputs)\n outputs = torch.nn.functional.softmax(outputs, dim=1)\n # print(outputs.shape)\n y_pred_prob = torch.cat([y_pred_prob, outputs.to(\"cpu\")], dim=0)\n #embed()\n y_pred_prob = y_pred_prob.reshape((len(aug), len_data, configs.num_classes))\n y_pred_prob = torch.sum(y_pred_prob, 0) / (len(aug) * 1.0)\n _, predicted_all = torch.max(y_pred_prob, 1)\n predicted = predicted_all + 1 # If the category starts with 1 ,else delet 1\n test_files.type = predicted.data.cpu().numpy().tolist()\n test_files.to_csv('./submits/%s_baseline.csv' % configs.model_name, index=False)\n" }, { "path": "utils/__init__.py", "content": "from .optimizers import *\nfrom .logger import *\nfrom .losses import *" }, { "path": "utils/logger.py", "content": "# A simple torch style logger\n# (C) Wei YANG 2017\nfrom __future__ import absolute_import\nimport matplotlib.pyplot as plt\nimport os\nimport sys\nimport numpy as np\n\n__all__ = ['Logger', 'LoggerMonitor', 'savefig']\n\ndef savefig(fname, dpi=None):\n dpi = 150 if dpi == None else dpi\n plt.savefig(fname, dpi=dpi)\n \ndef plot_overlap(logger, names=None):\n names = logger.names if names == None else names\n numbers = logger.numbers\n for _, name in enumerate(names):\n x = np.arange(len(numbers[name]))\n plt.plot(x, np.asarray(numbers[name]))\n return [logger.title + '(' + name + ')' for name in names]\n\nclass Logger(object):\n '''Save training process to log file with simple plot function.'''\n def __init__(self, fpath, title=None, resume=False): \n self.file = None\n self.resume = resume\n self.title = '' if title == None else title\n if fpath is not None:\n if resume: \n self.file = open(fpath, 'r') \n name = self.file.readline()\n self.names = name.rstrip().split('\\t')\n self.numbers = {}\n for _, name in enumerate(self.names):\n self.numbers[name] = []\n\n for numbers in self.file:\n numbers = numbers.rstrip().split('\\t')\n for i in range(0, len(numbers)):\n self.numbers[self.names[i]].append(numbers[i])\n self.file.close()\n self.file = open(fpath, 'a') \n else:\n self.file = open(fpath, 'w')\n\n def set_names(self, names):\n if self.resume: \n pass\n # initialize numbers as empty list\n self.numbers = {}\n self.names = names\n for _, name in enumerate(self.names):\n self.file.write(name)\n self.file.write('\\t')\n self.numbers[name] = []\n self.file.write('\\n')\n self.file.flush()\n\n\n def append(self, numbers):\n assert len(self.names) == len(numbers), 'Numbers do not match names'\n for index, num in enumerate(numbers):\n self.file.write(\"{0:.6f}\".format(num))\n self.file.write('\\t')\n self.numbers[self.names[index]].append(num)\n self.file.write('\\n')\n self.file.flush()\n\n def plot(self, names=None): \n names = self.names if names == None else names\n numbers = self.numbers\n for _, name in enumerate(names):\n x = np.arange(len(numbers[name]))\n plt.plot(x, np.asarray(numbers[name]))\n plt.legend([self.title + '(' + name + ')' for name in names])\n plt.grid(True)\n\n def close(self):\n if self.file is not None:\n self.file.close()\n\nclass LoggerMonitor(object):\n '''Load and visualize multiple logs.'''\n def __init__ (self, paths):\n '''paths is a distionary with {name:filepath} pair'''\n self.loggers = []\n for title, path in paths.items():\n logger = Logger(path, title=title, resume=True)\n self.loggers.append(logger)\n\n def plot(self, names=None):\n plt.figure()\n plt.subplot(121)\n legend_text = []\n for logger in self.loggers:\n legend_text += plot_overlap(logger, names)\n plt.legend(legend_text, bbox_to_anchor=(1.05, 1), loc=2, borderaxespad=0.)\n plt.grid(True)\n \nif __name__ == '__main__':\n # Example: logger monitor\n paths = {\n 'temp':'./logs/efficientnet-b3/log.txt', \n }\n\n field = ['Valid Acc.']\n\n monitor = LoggerMonitor(paths)\n monitor.plot(names=field)\n savefig('test.eps')" }, { "path": "utils/losses/__init__.py", "content": "from .label_smoothing import LabelSmoothingLoss\nfrom .focalloss import FocalLoss" }, { "path": "utils/losses/focalloss.py", "content": "import torch\nfrom torch import nn\n\nclass FocalLoss(nn.Module):\n def __init__(self, gamma=2., reduction='mean'):\n super().__init__()\n self.gamma = gamma\n self.reduction = reduction\n\n def forward(self, inputs, targets):\n CE_loss = nn.CrossEntropyLoss(reduction='none')(inputs, targets)\n pt = torch.exp(-CE_loss)\n F_loss = ((1 - pt)**self.gamma) * CE_loss\n if self.reduction == 'sum':\n return F_loss.sum()\n elif self.reduction == 'mean':\n return F_loss.mean()" }, { "path": "utils/losses/label_smoothing.py", "content": "import torch\n\nfrom torch import nn\n\nimport torch.nn.functional as F\n\n\nclass LabelSmoothingLoss(nn.Module):\n def __init__(self, label_smoothing, class_nums, ignore_index=-100):\n assert 0.0 < label_smoothing <= 1.0\n self.ignore_index = ignore_index\n super(LabelSmoothingLoss, self).__init__()\n\n smoothing_value = label_smoothing / (class_nums - 1)\n one_hot = torch.full((class_nums,), smoothing_value)\n if self.ignore_index >= 0:\n one_hot[self.ignore_index] = 0\n self.register_buffer('one_hot', one_hot.unsqueeze(0))\n\n self.confidence = 1.0 - label_smoothing\n\n def forward(self, output, target):\n \"\"\"\n output (FloatTensor): batch_size x n_classes\n target (LongTensor): batch_size\n \"\"\"\n\n log_output = F.log_softmax(output, dim=1)\n model_prob = self.one_hot.repeat(target.size(0), 1)\n model_prob.scatter_(1, target.unsqueeze(1), self.confidence)\n if self.ignore_index >= 0:\n model_prob.masked_fill_((target == self.ignore_index).unsqueeze(1), 0)\n # print(\"model_prob:{}\".format(model_prob))\n # print(\"log_output:{}\".format(log_output))\n\n return -torch.sum(model_prob * log_output) / target.size(0)" }, { "path": "utils/misc.py", "content": "import os\nimport torch\nimport shutil\nimport pandas as pd\nfrom .optimizers import *\nfrom config import configs\nfrom torch import optim as optim_t\nfrom tqdm import tqdm\nfrom glob import glob\nfrom itertools import chain\n\ndef get_optimizer(model):\n if configs.optim == \"adam\":\n return optim_t.Adam(model.parameters(),\n configs.lr,\n betas=(configs.beta1,configs.beta2),\n weight_decay=configs.wd)\n elif configs.optim == \"radam\":\n return RAdam(model.parameters(),\n configs.lr,\n betas=(configs.beta1,configs.beta2),\n weight_decay=configs.wd)\n elif configs.optim == \"ranger\":\n return Ranger(model.parameters(),\n lr = configs.lr,\n betas=(configs.beta1,configs.beta2),\n weight_decay=configs.wd)\n elif configs.optim == \"over9000\":\n return Over9000(model.parameters(),\n lr = configs.lr,\n betas=(configs.beta1,configs.beta2),\n weight_decay=configs.wd)\n elif configs.optim == \"ralamb\":\n return Ralamb(model.parameters(),\n lr = configs.lr,\n betas=(configs.beta1,configs.beta2),\n weight_decay=configs.wd)\n elif configs.optim == \"sgd\":\n return optim_t.SGD(model.parameters(),\n lr = configs.lr,\n momentum=configs.mom,\n weight_decay=configs.wd)\n else:\n print(\"%s optimizer will be add later\"%configs.optim)\n\ndef save_checkpoint(state,is_best,is_best_loss):\n filename = configs.checkpoints + os.sep + configs.model_name + \"-checkpoint.pth.tar\"\n torch.save(state, filename)\n if is_best:\n message = filename.replace(\"-checkpoint.pth.tar\",\"-best_model.pth.tar\")\n shutil.copyfile(filename, message)\n if is_best_loss:\n message = filename.replace(\"-checkpoint.pth.tar\",\"-best_loss.pth.tar\")\n shutil.copyfile(filename, message)\n\ndef get_lr(optimizer):\n for param_group in optimizer.param_groups:\n return param_group['lr']\n\nclass AverageMeter(object):\n \"\"\"Computes and stores the average and current value\n Imported from https://github.com/pytorch/examples/blob/master/imagenet/main.py#L247-L262\n \"\"\"\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\ndef accuracy(output, target, topk=(1,)):\n \"\"\"Computes the precision@k for the specified values of k\"\"\"\n maxk = max(topk)\n batch_size = target.size(0)\n\n _, pred = output.topk(maxk, 1, True, True)\n pred = pred.t()\n correct = pred.eq(target.view(1, -1).expand_as(pred))\n\n res = []\n for k in topk:\n correct_k = correct[:k].view(-1).float().sum(0)\n res.append(correct_k.mul_(100.0 / batch_size))\n return res\n\ndef get_files(root,mode):\n if mode == \"test\":\n files = []\n for img in os.listdir(root):\n files.append(root + img)\n files = pd.DataFrame({\"filename\":files})\n return files\n else:\n all_data_path, labels = [], []\n image_folders = list(map(lambda x: root + x, os.listdir(root)))\n all_images = list(chain.from_iterable(list(map(lambda x: glob(x + \"/*\"), image_folders))))\n print(\"loading train dataset\")\n for file in tqdm(all_images):\n all_data_path.append(file)\n labels.append(int(file.split(os.sep)[-2]))\n all_files = pd.DataFrame({\"filename\": all_data_path, \"label\": labels})\n return all_files\ndef adjust_learning_rate(optimizer, epoch):\n \"\"\"Sets the learning rate to the initial LR decayed by 10 every 30 epochs\"\"\"\n lrs = [5e-4, 1e-4, 1e-5, 1e-6]\n if epoch<=10:\n lr = lrs[0]\n elif epoch>10 and epoch<=16:\n lr = lrs[1]\n elif epoch>16 and epoch<=22:\n lr = lrs[2]\n else:\n lr = lrs[-1]\n for param_group in optimizer.param_groups:\n param_group['lr'] = lr" }, { "path": "utils/optimizers/__init__.py", "content": "from .lookahead import *\nfrom .novograd import *\nfrom .over9000 import *\nfrom .radam import *\nfrom .ralamb import *\nfrom .ranger import *" }, { "path": "utils/optimizers/lookahead.py", "content": "# Lookahead implementation from https://github.com/rwightman/pytorch-image-models/blob/master/timm/optim/lookahead.py\n\n\"\"\" Lookahead Optimizer Wrapper.\nImplementation modified from: https://github.com/alphadl/lookahead.pytorch\nPaper: `Lookahead Optimizer: k steps forward, 1 step back` - https://arxiv.org/abs/1907.08610\n\"\"\"\nimport torch\nfrom torch.optim.optimizer import Optimizer\nfrom collections import defaultdict\n\nclass Lookahead(Optimizer):\n def __init__(self, base_optimizer, alpha=0.5, k=6):\n if not 0.0 <= alpha <= 1.0:\n raise ValueError(f'Invalid slow update rate: {alpha}')\n if not 1 <= k:\n raise ValueError(f'Invalid lookahead steps: {k}')\n defaults = dict(lookahead_alpha=alpha, lookahead_k=k, lookahead_step=0)\n self.base_optimizer = base_optimizer\n self.param_groups = self.base_optimizer.param_groups\n self.defaults = base_optimizer.defaults\n self.defaults.update(defaults)\n self.state = defaultdict(dict)\n # manually add our defaults to the param groups\n for name, default in defaults.items():\n for group in self.param_groups:\n group.setdefault(name, default)\n\n def update_slow(self, group):\n for fast_p in group[\"params\"]:\n if fast_p.grad is None:\n continue\n param_state = self.state[fast_p]\n if 'slow_buffer' not in param_state:\n param_state['slow_buffer'] = torch.empty_like(fast_p.data)\n param_state['slow_buffer'].copy_(fast_p.data)\n slow = param_state['slow_buffer']\n slow.add_(group['lookahead_alpha'], fast_p.data - slow)\n fast_p.data.copy_(slow)\n\n def sync_lookahead(self):\n for group in self.param_groups:\n self.update_slow(group)\n\n def step(self, closure=None):\n # print(self.k)\n #assert id(self.param_groups) == id(self.base_optimizer.param_groups)\n loss = self.base_optimizer.step(closure)\n for group in self.param_groups:\n group['lookahead_step'] += 1\n if group['lookahead_step'] % group['lookahead_k'] == 0:\n self.update_slow(group)\n return loss\n\n def state_dict(self):\n fast_state_dict = self.base_optimizer.state_dict()\n slow_state = {\n (id(k) if isinstance(k, torch.Tensor) else k): v\n for k, v in self.state.items()\n }\n fast_state = fast_state_dict['state']\n param_groups = fast_state_dict['param_groups']\n return {\n 'state': fast_state,\n 'slow_state': slow_state,\n 'param_groups': param_groups,\n }\n\n def load_state_dict(self, state_dict):\n fast_state_dict = {\n 'state': state_dict['state'],\n 'param_groups': state_dict['param_groups'],\n }\n self.base_optimizer.load_state_dict(fast_state_dict)\n\n # We want to restore the slow state, but share param_groups reference\n # with base_optimizer. This is a bit redundant but least code\n slow_state_new = False\n if 'slow_state' not in state_dict:\n print('Loading state_dict from optimizer without Lookahead applied.')\n state_dict['slow_state'] = defaultdict(dict)\n slow_state_new = True\n slow_state_dict = {\n 'state': state_dict['slow_state'],\n 'param_groups': state_dict['param_groups'], # this is pointless but saves code\n }\n super(Lookahead, self).load_state_dict(slow_state_dict)\n self.param_groups = self.base_optimizer.param_groups # make both ref same container\n if slow_state_new:\n # reapply defaults to catch missing lookahead specific ones\n for name, default in self.defaults.items():\n for group in self.param_groups:\n group.setdefault(name, default)\n\ndef LookaheadAdam(params, alpha=0.5, k=6, *args, **kwargs):\n adam = Adam(params, *args, **kwargs)\n return Lookahead(adam, alpha, k)\n" }, { "path": "utils/optimizers/novograd.py", "content": "# Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n\nimport torch\nfrom torch.optim import Optimizer\nimport math\n\nclass AdamW(Optimizer):\n \"\"\"Implements AdamW algorithm.\n \n It has been proposed in `Adam: A Method for Stochastic Optimization`_.\n \n Arguments:\n params (iterable): iterable of parameters to optimize or dicts defining\n parameter groups\n lr (float, optional): learning rate (default: 1e-3)\n betas (Tuple[float, float], optional): coefficients used for computing\n running averages of gradient and its square (default: (0.9, 0.999))\n eps (float, optional): term added to the denominator to improve\n numerical stability (default: 1e-8)\n weight_decay (float, optional): weight decay (L2 penalty) (default: 0)\n amsgrad (boolean, optional): whether to use the AMSGrad variant of this\n algorithm from the paper `On the Convergence of Adam and Beyond`_\n \n Adam: A Method for Stochastic Optimization:\n https://arxiv.org/abs/1412.6980\n On the Convergence of Adam and Beyond:\n https://openreview.net/forum?id=ryQu7f-RZ\n \"\"\"\n \n def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8,\n weight_decay=0, amsgrad=False):\n if not 0.0 <= lr:\n raise ValueError(\"Invalid learning rate: {}\".format(lr))\n if not 0.0 <= eps:\n raise ValueError(\"Invalid epsilon value: {}\".format(eps))\n if not 0.0 <= betas[0] < 1.0:\n raise ValueError(\"Invalid beta parameter at index 0: {}\".format(betas[0]))\n if not 0.0 <= betas[1] < 1.0:\n raise ValueError(\"Invalid beta parameter at index 1: {}\".format(betas[1]))\n defaults = dict(lr=lr, betas=betas, eps=eps,\n weight_decay=weight_decay, amsgrad=amsgrad)\n super(AdamW, self).__init__(params, defaults)\n \n def __setstate__(self, state):\n super(AdamW, self).__setstate__(state)\n for group in self.param_groups:\n group.setdefault('amsgrad', False)\n \n def step(self, closure=None):\n \"\"\"Performs a single optimization step.\n \n Arguments:\n closure (callable, optional): A closure that reevaluates the model\n and returns the loss.\n \"\"\"\n loss = None\n if closure is not None:\n loss = closure()\n \n for group in self.param_groups:\n for p in group['params']:\n if p.grad is None:\n continue\n grad = p.grad.data\n if grad.is_sparse:\n raise RuntimeError('Adam does not support sparse gradients, please consider SparseAdam instead')\n amsgrad = group['amsgrad']\n \n state = self.state[p]\n \n # State initialization\n if len(state) == 0:\n state['step'] = 0\n # Exponential moving average of gradient values\n state['exp_avg'] = torch.zeros_like(p.data)\n # Exponential moving average of squared gradient values\n state['exp_avg_sq'] = torch.zeros_like(p.data)\n if amsgrad:\n # Maintains max of all exp. moving avg. of sq. grad. values\n state['max_exp_avg_sq'] = torch.zeros_like(p.data)\n \n exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']\n if amsgrad:\n max_exp_avg_sq = state['max_exp_avg_sq']\n beta1, beta2 = group['betas']\n \n state['step'] += 1\n # Decay the first and second moment running average coefficient\n exp_avg.mul_(beta1).add_(1 - beta1, grad)\n exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)\n if amsgrad:\n # Maintains the maximum of all 2nd moment running avg. till now\n torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq)\n # Use the max. for normalizing running avg. of gradient\n denom = max_exp_avg_sq.sqrt().add_(group['eps'])\n else:\n denom = exp_avg_sq.sqrt().add_(group['eps'])\n \n bias_correction1 = 1 - beta1 ** state['step']\n bias_correction2 = 1 - beta2 ** state['step']\n step_size = group['lr'] * math.sqrt(bias_correction2) / bias_correction1\n p.data.add_(-step_size, torch.mul(p.data, group['weight_decay']).addcdiv_(1, exp_avg, denom) )\n \n return loss\n \nclass Novograd(Optimizer):\n \"\"\"\n Implements Novograd algorithm.\n\n Args:\n params (iterable): iterable of parameters to optimize or dicts defining\n parameter groups\n lr (float, optional): learning rate (default: 1e-3)\n betas (Tuple[float, float], optional): coefficients used for computing\n running averages of gradient and its square (default: (0.95, 0))\n eps (float, optional): term added to the denominator to improve\n numerical stability (default: 1e-8)\n weight_decay (float, optional): weight decay (L2 penalty) (default: 0)\n grad_averaging: gradient averaging\n amsgrad (boolean, optional): whether to use the AMSGrad variant of this\n algorithm from the paper `On the Convergence of Adam and Beyond`_\n (default: False)\n \"\"\"\n\n def __init__(self, params, lr=1e-3, betas=(0.95, 0), eps=1e-8,\n weight_decay=0, grad_averaging=False, amsgrad=False):\n if not 0.0 <= lr:\n raise ValueError(\"Invalid learning rate: {}\".format(lr))\n if not 0.0 <= eps:\n raise ValueError(\"Invalid epsilon value: {}\".format(eps))\n if not 0.0 <= betas[0] < 1.0:\n raise ValueError(\"Invalid beta parameter at index 0: {}\".format(betas[0]))\n if not 0.0 <= betas[1] < 1.0:\n raise ValueError(\"Invalid beta parameter at index 1: {}\".format(betas[1]))\n defaults = dict(lr=lr, betas=betas, eps=eps,\n weight_decay=weight_decay,\n grad_averaging=grad_averaging,\n amsgrad=amsgrad)\n\n super(Novograd, self).__init__(params, defaults)\n\n def __setstate__(self, state):\n super(Novograd, self).__setstate__(state)\n for group in self.param_groups:\n group.setdefault('amsgrad', False)\n\n def step(self, closure=None):\n \"\"\"Performs a single optimization step.\n\n Arguments:\n closure (callable, optional): A closure that reevaluates the model\n and returns the loss.\n \"\"\"\n loss = None\n if closure is not None:\n loss = closure()\n\n for group in self.param_groups:\n for p in group['params']:\n if p.grad is None:\n continue\n grad = p.grad.data\n if grad.is_sparse:\n raise RuntimeError('Sparse gradients are not supported.')\n amsgrad = group['amsgrad']\n\n state = self.state[p]\n\n # State initialization\n if len(state) == 0:\n state['step'] = 0\n # Exponential moving average of gradient values\n state['exp_avg'] = torch.zeros_like(p.data)\n # Exponential moving average of squared gradient values\n state['exp_avg_sq'] = torch.zeros([]).to(state['exp_avg'].device)\n if amsgrad:\n # Maintains max of all exp. moving avg. of sq. grad. values\n state['max_exp_avg_sq'] = torch.zeros([]).to(state['exp_avg'].device)\n\n exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']\n if amsgrad:\n max_exp_avg_sq = state['max_exp_avg_sq']\n beta1, beta2 = group['betas']\n\n state['step'] += 1\n\n norm = torch.sum(torch.pow(grad, 2))\n\n if exp_avg_sq == 0:\n exp_avg_sq.copy_(norm)\n else:\n exp_avg_sq.mul_(beta2).add_(1 - beta2, norm)\n\n if amsgrad:\n # Maintains the maximum of all 2nd moment running avg. till now\n torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq)\n # Use the max. for normalizing running avg. of gradient\n denom = max_exp_avg_sq.sqrt().add_(group['eps'])\n else:\n denom = exp_avg_sq.sqrt().add_(group['eps'])\n\n grad.div_(denom)\n if group['weight_decay'] != 0:\n grad.add_(group['weight_decay'], p.data)\n if group['grad_averaging']:\n grad.mul_(1 - beta1)\n exp_avg.mul_(beta1).add_(grad)\n\n p.data.add_(-group['lr'], exp_avg)\n \n return loss" }, { "path": "utils/optimizers/over9000.py", "content": "import torch, math\nfrom torch.optim.optimizer import Optimizer\nimport itertools as it\nfrom .lookahead import *\nfrom .ralamb import * \n\n# RAdam + LARS + LookAHead\n\n# Lookahead implementation from https://github.com/lonePatient/lookahead_pytorch/blob/master/optimizer.py\n# RAdam + LARS implementation from https://gist.github.com/redknightlois/c4023d393eb8f92bb44b2ab582d7ec20\n\ndef Over9000(params, alpha=0.5, k=6, *args, **kwargs):\n ralamb = Ralamb(params, *args, **kwargs)\n return Lookahead(ralamb, alpha, k)\n\nRangerLars = Over9000\n" }, { "path": "utils/optimizers/radam.py", "content": "# from https://github.com/LiyuanLucasLiu/RAdam/blob/master/radam.py\n\nimport math\nimport torch\nfrom torch.optim.optimizer import Optimizer, required\n\nclass RAdam(Optimizer):\n\n def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0):\n defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay)\n self.buffer = [[None, None, None] for ind in range(10)]\n super(RAdam, self).__init__(params, defaults)\n\n def __setstate__(self, state):\n super(RAdam, self).__setstate__(state)\n\n def step(self, closure=None):\n\n loss = None\n if closure is not None:\n loss = closure()\n\n for group in self.param_groups:\n\n for p in group['params']:\n if p.grad is None:\n continue\n grad = p.grad.data.float()\n if grad.is_sparse:\n raise RuntimeError('RAdam does not support sparse gradients')\n\n p_data_fp32 = p.data.float()\n\n state = self.state[p]\n\n if len(state) == 0:\n state['step'] = 0\n state['exp_avg'] = torch.zeros_like(p_data_fp32)\n state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)\n else:\n state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32)\n state['exp_avg_sq'] = state['exp_avg_sq'].type_as(p_data_fp32)\n\n exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']\n beta1, beta2 = group['betas']\n\n exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)\n exp_avg.mul_(beta1).add_(1 - beta1, grad)\n\n state['step'] += 1\n buffered = self.buffer[int(state['step'] % 10)]\n if state['step'] == buffered[0]:\n N_sma, step_size = buffered[1], buffered[2]\n else:\n buffered[0] = state['step']\n beta2_t = beta2 ** state['step']\n N_sma_max = 2 / (1 - beta2) - 1\n N_sma = N_sma_max - 2 * state['step'] * beta2_t / (1 - beta2_t)\n buffered[1] = N_sma\n\n # more conservative since it's an approximated value\n if N_sma >= 5:\n step_size = math.sqrt((1 - beta2_t) * (N_sma - 4) / (N_sma_max - 4) * (N_sma - 2) / N_sma * N_sma_max / (N_sma_max - 2)) / (1 - beta1 ** state['step'])\n else:\n step_size = 1.0 / (1 - beta1 ** state['step'])\n buffered[2] = step_size\n\n if group['weight_decay'] != 0:\n p_data_fp32.add_(-group['weight_decay'] * group['lr'], p_data_fp32)\n\n # more conservative since it's an approximated value\n if N_sma >= 5: \n denom = exp_avg_sq.sqrt().add_(group['eps'])\n p_data_fp32.addcdiv_(-step_size * group['lr'], exp_avg, denom)\n else:\n p_data_fp32.add_(-step_size * group['lr'], exp_avg)\n\n p.data.copy_(p_data_fp32)\n\n return loss\n\nclass PlainRAdam(Optimizer):\n\n def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0):\n defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay)\n\n super(PlainRAdam, self).__init__(params, defaults)\n\n def __setstate__(self, state):\n super(PlainRAdam, self).__setstate__(state)\n\n def step(self, closure=None):\n\n loss = None\n if closure is not None:\n loss = closure()\n\n for group in self.param_groups:\n\n for p in group['params']:\n if p.grad is None:\n continue\n grad = p.grad.data.float()\n if grad.is_sparse:\n raise RuntimeError('RAdam does not support sparse gradients')\n\n p_data_fp32 = p.data.float()\n\n state = self.state[p]\n\n if len(state) == 0:\n state['step'] = 0\n state['exp_avg'] = torch.zeros_like(p_data_fp32)\n state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)\n else:\n state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32)\n state['exp_avg_sq'] = state['exp_avg_sq'].type_as(p_data_fp32)\n\n exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']\n beta1, beta2 = group['betas']\n\n exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)\n exp_avg.mul_(beta1).add_(1 - beta1, grad)\n\n state['step'] += 1\n beta2_t = beta2 ** state['step']\n N_sma_max = 2 / (1 - beta2) - 1\n N_sma = N_sma_max - 2 * state['step'] * beta2_t / (1 - beta2_t)\n\n if group['weight_decay'] != 0:\n p_data_fp32.add_(-group['weight_decay'] * group['lr'], p_data_fp32)\n\n # more conservative since it's an approximated value\n if N_sma >= 5: \n step_size = group['lr'] * math.sqrt((1 - beta2_t) * (N_sma - 4) / (N_sma_max - 4) * (N_sma - 2) / N_sma * N_sma_max / (N_sma_max - 2)) / (1 - beta1 ** state['step'])\n denom = exp_avg_sq.sqrt().add_(group['eps'])\n p_data_fp32.addcdiv_(-step_size, exp_avg, denom)\n else:\n step_size = group['lr'] / (1 - beta1 ** state['step'])\n p_data_fp32.add_(-step_size, exp_avg)\n\n p.data.copy_(p_data_fp32)\n\n return loss\n\n\nclass AdamW(Optimizer):\n\n def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0, warmup = 0):\n defaults = dict(lr=lr, betas=betas, eps=eps,\n weight_decay=weight_decay, warmup = warmup)\n super(AdamW, self).__init__(params, defaults)\n\n def __setstate__(self, state):\n super(AdamW, self).__setstate__(state)\n\n def step(self, closure=None):\n loss = None\n if closure is not None:\n loss = closure()\n\n for group in self.param_groups:\n\n for p in group['params']:\n if p.grad is None:\n continue\n grad = p.grad.data.float()\n if grad.is_sparse:\n raise RuntimeError('Adam does not support sparse gradients, please consider SparseAdam instead')\n\n p_data_fp32 = p.data.float()\n\n state = self.state[p]\n\n if len(state) == 0:\n state['step'] = 0\n state['exp_avg'] = torch.zeros_like(p_data_fp32)\n state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)\n else:\n state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32)\n state['exp_avg_sq'] = state['exp_avg_sq'].type_as(p_data_fp32)\n\n exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']\n beta1, beta2 = group['betas']\n\n state['step'] += 1\n\n exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)\n exp_avg.mul_(beta1).add_(1 - beta1, grad)\n\n denom = exp_avg_sq.sqrt().add_(group['eps'])\n bias_correction1 = 1 - beta1 ** state['step']\n bias_correction2 = 1 - beta2 ** state['step']\n \n if group['warmup'] > state['step']:\n scheduled_lr = 1e-8 + state['step'] * group['lr'] / group['warmup']\n else:\n scheduled_lr = group['lr']\n\n step_size = group['lr'] * math.sqrt(bias_correction2) / bias_correction1\n \n if group['weight_decay'] != 0:\n p_data_fp32.add_(-group['weight_decay'] * scheduled_lr, p_data_fp32)\n\n p_data_fp32.addcdiv_(-step_size, exp_avg, denom)\n\n p.data.copy_(p_data_fp32)\n\n return loss\n" }, { "path": "utils/optimizers/ralamb.py", "content": "import torch, math\nfrom torch.optim.optimizer import Optimizer\n\n# RAdam + LARS\nclass Ralamb(Optimizer):\n\n def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0):\n defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay)\n self.buffer = [[None, None, None] for ind in range(10)]\n super(Ralamb, self).__init__(params, defaults)\n\n def __setstate__(self, state):\n super(Ralamb, self).__setstate__(state)\n\n def step(self, closure=None):\n\n loss = None\n if closure is not None:\n loss = closure()\n\n for group in self.param_groups:\n\n for p in group['params']:\n if p.grad is None:\n continue\n grad = p.grad.data.float()\n if grad.is_sparse:\n raise RuntimeError('Ralamb does not support sparse gradients')\n\n p_data_fp32 = p.data.float()\n\n state = self.state[p]\n\n if len(state) == 0:\n state['step'] = 0\n state['exp_avg'] = torch.zeros_like(p_data_fp32)\n state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)\n else:\n state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32)\n state['exp_avg_sq'] = state['exp_avg_sq'].type_as(p_data_fp32)\n\n exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']\n beta1, beta2 = group['betas']\n\n # Decay the first and second moment running average coefficient\n # m_t\n exp_avg.mul_(beta1).add_(1 - beta1, grad)\n # v_t\n exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)\n\n state['step'] += 1\n buffered = self.buffer[int(state['step'] % 10)]\n\n if state['step'] == buffered[0]:\n N_sma, radam_step_size = buffered[1], buffered[2]\n else:\n buffered[0] = state['step']\n beta2_t = beta2 ** state['step']\n N_sma_max = 2 / (1 - beta2) - 1\n N_sma = N_sma_max - 2 * state['step'] * beta2_t / (1 - beta2_t)\n buffered[1] = N_sma\n\n # more conservative since it's an approximated value\n if N_sma >= 5:\n radam_step_size = math.sqrt((1 - beta2_t) * (N_sma - 4) / (N_sma_max - 4) * (N_sma - 2) / N_sma * N_sma_max / (N_sma_max - 2)) / (1 - beta1 ** state['step'])\n else:\n radam_step_size = 1.0 / (1 - beta1 ** state['step'])\n buffered[2] = radam_step_size\n\n if group['weight_decay'] != 0:\n p_data_fp32.add_(-group['weight_decay'] * group['lr'], p_data_fp32)\n\n # more conservative since it's an approximated value\n radam_step = p_data_fp32.clone()\n if N_sma >= 5:\n denom = exp_avg_sq.sqrt().add_(group['eps'])\n radam_step.addcdiv_(-radam_step_size * group['lr'], exp_avg, denom)\n else:\n radam_step.add_(-radam_step_size * group['lr'], exp_avg)\n\n radam_norm = radam_step.pow(2).sum().sqrt()\n weight_norm = p.data.pow(2).sum().sqrt().clamp(0, 10)\n if weight_norm == 0 or radam_norm == 0:\n trust_ratio = 1\n else:\n trust_ratio = weight_norm / radam_norm\n\n state['weight_norm'] = weight_norm\n state['adam_norm'] = radam_norm\n state['trust_ratio'] = trust_ratio\n\n if N_sma >= 5:\n p_data_fp32.addcdiv_(-radam_step_size * group['lr'] * trust_ratio, exp_avg, denom)\n else:\n p_data_fp32.add_(-radam_step_size * group['lr'] * trust_ratio, exp_avg)\n\n p.data.copy_(p_data_fp32)\n\n return loss\n" }, { "path": "utils/optimizers/ranger.py", "content": "\nimport math\nimport torch\nfrom torch.optim.optimizer import Optimizer, required\nimport itertools as it\nfrom .lookahead import *\nfrom .radam import * \n\ndef Ranger(params, alpha=0.5, k=6, *args, **kwargs):\n radam = RAdam(params, *args, **kwargs)\n return Lookahead(radam, alpha, k)\n\n" }, { "path": "utils/reader.py", "content": "from torch.utils.data import Dataset\nfrom PIL import Image\n\nclass WeatherDataset(Dataset):\n # define dataset\n def __init__(self,label_list,transforms=None,mode=\"train\"):\n super(WeatherDataset,self).__init__()\n self.label_list = label_list\n self.transforms = transforms\n self.mode = mode\n imgs = []\n if self.mode == \"test\":\n for index,row in label_list.iterrows():\n imgs.append((row[\"filename\"]))\n self.imgs = imgs\n else:\n for index,row in label_list.iterrows():\n imgs.append((row[\"filename\"],row[\"label\"]))\n self.imgs = imgs\n def __len__(self):\n return len(self.imgs)\n def __getitem__(self,index):\n if self.mode == \"test\":\n filename = self.imgs[index]\n img = Image.open(filename).convert('RGB')\n img = self.transforms(img)\n return img,filename\n else:\n filename,label = self.imgs[index]\n img = Image.open(filename).convert('RGB')\n img = self.transforms(img)\n return img,label\n\n\n" } ]