[
  {
    "path": ".gitignore",
    "content": "*.pyc\n*.swp\n*.ipynb_checkpoints\n*.json\n*.pth.tar\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": "# Improving Visual-Semantic Embeddings with Hard Negatives\n\nCode for the image-caption retrieval methods from\n**[VSE++: Improving Visual-Semantic Embeddings with Hard Negatives](https://arxiv.org/abs/1707.05612)**\n*, F. Faghri, D. J. Fleet, J. R. Kiros, S. Fidler, Proceedings of the British Machine Vision Conference (BMVC),  2018. (BMVC Spotlight)*\n\n## Dependencies\nWe recommended to use Anaconda for the following packages.\n\n* Python 2.7 (Checkout branch `python3`)\n* [PyTorch](http://pytorch.org/) (>0.2) (Checkout branch `pytorch4.1`)\n* [NumPy](http://www.numpy.org/) (>1.12.1)\n* [TensorBoard](https://github.com/TeamHG-Memex/tensorboard_logger)\n* [pycocotools](https://github.com/cocodataset/cocoapi)\n* [torchvision]()\n* [matplotlib]()\n\n\n* Punkt Sentence Tokenizer:\n```python\nimport nltk\nnltk.download()\n> d punkt\n```\n\n## Download data\n\nDownload the dataset files and pre-trained models. We use splits produced by [Andrej Karpathy](http://cs.stanford.edu/people/karpathy/deepimagesent/). The precomputed image features are from [here](https://github.com/ryankiros/visual-semantic-embedding/) and [here](https://github.com/ivendrov/order-embedding). To use full image encoders, download the images from their original sources [here](http://nlp.cs.illinois.edu/HockenmaierGroup/Framing_Image_Description/KCCA.html), [here](http://shannon.cs.illinois.edu/DenotationGraph/) and [here](http://mscoco.org/).\n\n```bash\nwget http://www.cs.toronto.edu/~faghri/vsepp/vocab.tar\nwget http://www.cs.toronto.edu/~faghri/vsepp/data.tar\nwget http://www.cs.toronto.edu/~faghri/vsepp/runs.tar\n```\n\nWe refer to the path of extracted files for `data.tar` as `$DATA_PATH` and \nfiles for `models.tar` as `$RUN_PATH`. Extract `vocab.tar` to `./vocab` \ndirectory.\n\n*Update: The vocabulary was originally built using all sets (including test set \ncaptions). Please see issue #29 for details. Please consider not using test set \ncaptions if building up on this project.*\n\n## Evaluate pre-trained models\n\n```python\npython -c \"\\\nfrom vocab import Vocabulary\nimport evaluation\nevaluation.evalrank('$RUN_PATH/coco_vse++/model_best.pth.tar', data_path='$DATA_PATH', split='test')\"\n```\n\nTo do cross-validation on MSCOCO, pass `fold5=True` with a model trained using \n`--data_name coco`.\n\n## Training new models\nRun `train.py`:\n\n```bash\npython train.py --data_path \"$DATA_PATH\" --data_name coco_precomp --logger_name \nruns/coco_vse++ --max_violation\n```\n\nArguments used to train pre-trained models:\n\n| Method    | Arguments |\n| :-------: | :-------: |\n| VSE0      | `--no_imgnorm` |\n| VSE++     | `--max_violation` |\n| Order0    | `--measure order --use_abs --margin .05 --learning_rate .001` |\n| Order++   | `--measure order --max_violation` |\n\n\n## Reference\n\nIf you found this code useful, please cite the following paper:\n\n    @article{faghri2018vse++,\n      title={VSE++: Improving Visual-Semantic Embeddings with Hard Negatives},\n      author={Faghri, Fartash and Fleet, David J and Kiros, Jamie Ryan and Fidler, Sanja},\n      booktitle = {Proceedings of the British Machine Vision Conference ({BMVC})},\n      url = {https://github.com/fartashf/vsepp},\n      year={2018}\n    }\n\n## License\n\n[Apache License 2.0](http://www.apache.org/licenses/LICENSE-2.0)\n"
  },
  {
    "path": "data.py",
    "content": "import torch\nimport torch.utils.data as data\nimport torchvision.transforms as transforms\nimport os\nimport nltk\nfrom PIL import Image\nfrom pycocotools.coco import COCO\nimport numpy as np\nimport json as jsonmod\n\n\ndef get_paths(path, name='coco', use_restval=False):\n    \"\"\"\n    Returns paths to images and annotations for the given datasets. For MSCOCO\n    indices are also returned to control the data split being used.\n    The indices are extracted from the Karpathy et al. splits using this\n    snippet:\n\n    >>> import json\n    >>> dataset=json.load(open('dataset_coco.json','r'))\n    >>> A=[]\n    >>> for i in range(len(D['images'])):\n    ...   if D['images'][i]['split'] == 'val':\n    ...     A+=D['images'][i]['sentids'][:5]\n    ...\n\n    :param name: Dataset names\n    :param use_restval: If True, the the `restval` data is included in train.\n    \"\"\"\n    roots = {}\n    ids = {}\n    if 'coco' == name:\n        imgdir = os.path.join(path, 'images')\n        capdir = os.path.join(path, 'annotations')\n        roots['train'] = {\n            'img': os.path.join(imgdir, 'train2014'),\n            'cap': os.path.join(capdir, 'captions_train2014.json')\n        }\n        roots['val'] = {\n            'img': os.path.join(imgdir, 'val2014'),\n            'cap': os.path.join(capdir, 'captions_val2014.json')\n        }\n        roots['test'] = {\n            'img': os.path.join(imgdir, 'val2014'),\n            'cap': os.path.join(capdir, 'captions_val2014.json')\n        }\n        roots['trainrestval'] = {\n            'img': (roots['train']['img'], roots['val']['img']),\n            'cap': (roots['train']['cap'], roots['val']['cap'])\n        }\n        ids['train'] = np.load(os.path.join(capdir, 'coco_train_ids.npy'))\n        ids['val'] = np.load(os.path.join(capdir, 'coco_dev_ids.npy'))[:5000]\n        ids['test'] = np.load(os.path.join(capdir, 'coco_test_ids.npy'))\n        ids['trainrestval'] = (\n            ids['train'],\n            np.load(os.path.join(capdir, 'coco_restval_ids.npy')))\n        if use_restval:\n            roots['train'] = roots['trainrestval']\n            ids['train'] = ids['trainrestval']\n    elif 'f8k' == name:\n        imgdir = os.path.join(path, 'images')\n        cap = os.path.join(path, 'dataset_flickr8k.json')\n        roots['train'] = {'img': imgdir, 'cap': cap}\n        roots['val'] = {'img': imgdir, 'cap': cap}\n        roots['test'] = {'img': imgdir, 'cap': cap}\n        ids = {'train': None, 'val': None, 'test': None}\n    elif 'f30k' == name:\n        imgdir = os.path.join(path, 'images')\n        cap = os.path.join(path, 'dataset_flickr30k.json')\n        roots['train'] = {'img': imgdir, 'cap': cap}\n        roots['val'] = {'img': imgdir, 'cap': cap}\n        roots['test'] = {'img': imgdir, 'cap': cap}\n        ids = {'train': None, 'val': None, 'test': None}\n\n    return roots, ids\n\n\nclass CocoDataset(data.Dataset):\n    \"\"\"COCO Custom Dataset compatible with torch.utils.data.DataLoader.\"\"\"\n\n    def __init__(self, root, json, vocab, transform=None, ids=None):\n        \"\"\"\n        Args:\n            root: image directory.\n            json: coco annotation file path.\n            vocab: vocabulary wrapper.\n            transform: transformer for image.\n        \"\"\"\n        self.root = root\n        # when using `restval`, two json files are needed\n        if isinstance(json, tuple):\n            self.coco = (COCO(json[0]), COCO(json[1]))\n        else:\n            self.coco = (COCO(json),)\n            self.root = (root,)\n        # if ids provided by get_paths, use split-specific ids\n        if ids is None:\n            self.ids = list(self.coco.anns.keys())\n        else:\n            self.ids = ids\n\n        # if `restval` data is to be used, record the break point for ids\n        if isinstance(self.ids, tuple):\n            self.bp = len(self.ids[0])\n            self.ids = list(self.ids[0]) + list(self.ids[1])\n        else:\n            self.bp = len(self.ids)\n        self.vocab = vocab\n        self.transform = transform\n\n    def __getitem__(self, index):\n        \"\"\"This function returns a tuple that is further passed to collate_fn\n        \"\"\"\n        vocab = self.vocab\n        root, caption, img_id, path, image = self.get_raw_item(index)\n\n        if self.transform is not None:\n            image = self.transform(image)\n\n        # Convert caption (string) to word ids.\n        tokens = nltk.tokenize.word_tokenize(\n            str(caption).lower().decode('utf-8'))\n        caption = []\n        caption.append(vocab('<start>'))\n        caption.extend([vocab(token) for token in tokens])\n        caption.append(vocab('<end>'))\n        target = torch.Tensor(caption)\n        return image, target, index, img_id\n\n    def get_raw_item(self, index):\n        if index < self.bp:\n            coco = self.coco[0]\n            root = self.root[0]\n        else:\n            coco = self.coco[1]\n            root = self.root[1]\n        ann_id = self.ids[index]\n        caption = coco.anns[ann_id]['caption']\n        img_id = coco.anns[ann_id]['image_id']\n        path = coco.loadImgs(img_id)[0]['file_name']\n        image = Image.open(os.path.join(root, path)).convert('RGB')\n\n        return root, caption, img_id, path, image\n\n    def __len__(self):\n        return len(self.ids)\n\n\nclass FlickrDataset(data.Dataset):\n    \"\"\"\n    Dataset loader for Flickr30k and Flickr8k full datasets.\n    \"\"\"\n\n    def __init__(self, root, json, split, vocab, transform=None):\n        self.root = root\n        self.vocab = vocab\n        self.split = split\n        self.transform = transform\n        self.dataset = jsonmod.load(open(json, 'r'))['images']\n        self.ids = []\n        for i, d in enumerate(self.dataset):\n            if d['split'] == split:\n                self.ids += [(i, x) for x in range(len(d['sentences']))]\n\n    def __getitem__(self, index):\n        \"\"\"This function returns a tuple that is further passed to collate_fn\n        \"\"\"\n        vocab = self.vocab\n        root = self.root\n        ann_id = self.ids[index]\n        img_id = ann_id[0]\n        caption = self.dataset[img_id]['sentences'][ann_id[1]]['raw']\n        path = self.dataset[img_id]['filename']\n\n        image = Image.open(os.path.join(root, path)).convert('RGB')\n        if self.transform is not None:\n            image = self.transform(image)\n\n        # Convert caption (string) to word ids.\n        tokens = nltk.tokenize.word_tokenize(\n            str(caption).lower().decode('utf-8'))\n        caption = []\n        caption.append(vocab('<start>'))\n        caption.extend([vocab(token) for token in tokens])\n        caption.append(vocab('<end>'))\n        target = torch.Tensor(caption)\n        return image, target, index, img_id\n\n    def __len__(self):\n        return len(self.ids)\n\n\nclass PrecompDataset(data.Dataset):\n    \"\"\"\n    Load precomputed captions and image features\n    Possible options: f8k, f30k, coco, 10crop\n    \"\"\"\n\n    def __init__(self, data_path, data_split, vocab):\n        self.vocab = vocab\n        loc = data_path + '/'\n\n        # Captions\n        self.captions = []\n        with open(loc+'%s_caps.txt' % data_split, 'rb') as f:\n            for line in f:\n                self.captions.append(line.strip())\n\n        # Image features\n        self.images = np.load(loc+'%s_ims.npy' % data_split)\n        self.length = len(self.captions)\n        # rkiros data has redundancy in images, we divide by 5, 10crop doesn't\n        if self.images.shape[0] != self.length:\n            self.im_div = 5\n        else:\n            self.im_div = 1\n        # the development set for coco is large and so validation would be slow\n        if data_split == 'dev':\n            self.length = 5000\n\n    def __getitem__(self, index):\n        # handle the image redundancy\n        img_id = index/self.im_div\n        image = torch.Tensor(self.images[img_id])\n        caption = self.captions[index]\n        vocab = self.vocab\n\n        # Convert caption (string) to word ids.\n        tokens = nltk.tokenize.word_tokenize(\n            str(caption).lower().decode('utf-8'))\n        caption = []\n        caption.append(vocab('<start>'))\n        caption.extend([vocab(token) for token in tokens])\n        caption.append(vocab('<end>'))\n        target = torch.Tensor(caption)\n        return image, target, index, img_id\n\n    def __len__(self):\n        return self.length\n\n\ndef collate_fn(data):\n    \"\"\"Build mini-batch tensors from a list of (image, caption) tuples.\n    Args:\n        data: list of (image, caption) tuple.\n            - image: torch tensor of shape (3, 256, 256).\n            - caption: torch tensor of shape (?); variable length.\n\n    Returns:\n        images: torch tensor of shape (batch_size, 3, 256, 256).\n        targets: torch tensor of shape (batch_size, padded_length).\n        lengths: list; valid length for each padded caption.\n    \"\"\"\n    # Sort a data list by caption length\n    data.sort(key=lambda x: len(x[1]), reverse=True)\n    images, captions, ids, img_ids = zip(*data)\n\n    # Merge images (convert tuple of 3D tensor to 4D tensor)\n    images = torch.stack(images, 0)\n\n    # Merget captions (convert tuple of 1D tensor to 2D tensor)\n    lengths = [len(cap) for cap in captions]\n    targets = torch.zeros(len(captions), max(lengths)).long()\n    for i, cap in enumerate(captions):\n        end = lengths[i]\n        targets[i, :end] = cap[:end]\n\n    return images, targets, lengths, ids\n\n\ndef get_loader_single(data_name, split, root, json, vocab, transform,\n                      batch_size=100, shuffle=True,\n                      num_workers=2, ids=None, collate_fn=collate_fn):\n    \"\"\"Returns torch.utils.data.DataLoader for custom coco dataset.\"\"\"\n    if 'coco' in data_name:\n        # COCO custom dataset\n        dataset = CocoDataset(root=root,\n                              json=json,\n                              vocab=vocab,\n                              transform=transform, ids=ids)\n    elif 'f8k' in data_name or 'f30k' in data_name:\n        dataset = FlickrDataset(root=root,\n                                split=split,\n                                json=json,\n                                vocab=vocab,\n                                transform=transform)\n\n    # Data loader\n    data_loader = torch.utils.data.DataLoader(dataset=dataset,\n                                              batch_size=batch_size,\n                                              shuffle=shuffle,\n                                              pin_memory=True,\n                                              num_workers=num_workers,\n                                              collate_fn=collate_fn)\n    return data_loader\n\n\ndef get_precomp_loader(data_path, data_split, vocab, opt, batch_size=100,\n                       shuffle=True, num_workers=2):\n    \"\"\"Returns torch.utils.data.DataLoader for custom coco dataset.\"\"\"\n    dset = PrecompDataset(data_path, data_split, vocab)\n\n    data_loader = torch.utils.data.DataLoader(dataset=dset,\n                                              batch_size=batch_size,\n                                              shuffle=shuffle,\n                                              pin_memory=True,\n                                              collate_fn=collate_fn)\n    return data_loader\n\n\ndef get_transform(data_name, split_name, opt):\n    normalizer = transforms.Normalize(mean=[0.485, 0.456, 0.406],\n                                      std=[0.229, 0.224, 0.225])\n    t_list = []\n    if split_name == 'train':\n        t_list = [transforms.RandomResizedCrop(opt.crop_size),\n                  transforms.RandomHorizontalFlip()]\n    elif split_name == 'val':\n        t_list = [transforms.Resize(256), transforms.CenterCrop(224)]\n    elif split_name == 'test':\n        t_list = [transforms.Resize(256), transforms.CenterCrop(224)]\n\n    t_end = [transforms.ToTensor(), normalizer]\n    transform = transforms.Compose(t_list + t_end)\n    return transform\n\n\ndef get_loaders(data_name, vocab, crop_size, batch_size, workers, opt):\n    dpath = os.path.join(opt.data_path, data_name)\n    if opt.data_name.endswith('_precomp'):\n        train_loader = get_precomp_loader(dpath, 'train', vocab, opt,\n                                          batch_size, True, workers)\n        val_loader = get_precomp_loader(dpath, 'dev', vocab, opt,\n                                        batch_size, False, workers)\n    else:\n        # Build Dataset Loader\n        roots, ids = get_paths(dpath, data_name, opt.use_restval)\n\n        transform = get_transform(data_name, 'train', opt)\n        train_loader = get_loader_single(opt.data_name, 'train',\n                                         roots['train']['img'],\n                                         roots['train']['cap'],\n                                         vocab, transform, ids=ids['train'],\n                                         batch_size=batch_size, shuffle=True,\n                                         num_workers=workers,\n                                         collate_fn=collate_fn)\n\n        transform = get_transform(data_name, 'val', opt)\n        val_loader = get_loader_single(opt.data_name, 'val',\n                                       roots['val']['img'],\n                                       roots['val']['cap'],\n                                       vocab, transform, ids=ids['val'],\n                                       batch_size=batch_size, shuffle=False,\n                                       num_workers=workers,\n                                       collate_fn=collate_fn)\n\n    return train_loader, val_loader\n\n\ndef get_test_loader(split_name, data_name, vocab, crop_size, batch_size,\n                    workers, opt):\n    dpath = os.path.join(opt.data_path, data_name)\n    if opt.data_name.endswith('_precomp'):\n        test_loader = get_precomp_loader(dpath, split_name, vocab, opt,\n                                         batch_size, False, workers)\n    else:\n        # Build Dataset Loader\n        roots, ids = get_paths(dpath, data_name, opt.use_restval)\n\n        transform = get_transform(data_name, split_name, opt)\n        test_loader = get_loader_single(opt.data_name, split_name,\n                                        roots[split_name]['img'],\n                                        roots[split_name]['cap'],\n                                        vocab, transform, ids=ids[split_name],\n                                        batch_size=batch_size, shuffle=False,\n                                        num_workers=workers,\n                                        collate_fn=collate_fn)\n\n    return test_loader\n"
  },
  {
    "path": "evaluation.py",
    "content": "from __future__ import print_function\nimport os\nimport pickle\n\nimport numpy\nfrom data import get_test_loader\nimport time\nimport numpy as np\nfrom vocab import Vocabulary  # NOQA\nimport torch\nfrom model import VSE, order_sim\nfrom collections import OrderedDict\n\n\nclass AverageMeter(object):\n    \"\"\"Computes and stores the average and current value\"\"\"\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=0):\n        self.val = val\n        self.sum += val * n\n        self.count += n\n        self.avg = self.sum / (.0001 + self.count)\n\n    def __str__(self):\n        \"\"\"String representation for logging\n        \"\"\"\n        # for values that should be recorded exactly e.g. iteration number\n        if self.count == 0:\n            return str(self.val)\n        # for stats\n        return '%.4f (%.4f)' % (self.val, self.avg)\n\n\nclass LogCollector(object):\n    \"\"\"A collection of logging objects that can change from train to val\"\"\"\n\n    def __init__(self):\n        # to keep the order of logged variables deterministic\n        self.meters = OrderedDict()\n\n    def update(self, k, v, n=0):\n        # create a new meter if previously not recorded\n        if k not in self.meters:\n            self.meters[k] = AverageMeter()\n        self.meters[k].update(v, n)\n\n    def __str__(self):\n        \"\"\"Concatenate the meters in one log line\n        \"\"\"\n        s = ''\n        for i, (k, v) in enumerate(self.meters.iteritems()):\n            if i > 0:\n                s += '  '\n            s += k + ' ' + str(v)\n        return s\n\n    def tb_log(self, tb_logger, prefix='', step=None):\n        \"\"\"Log using tensorboard\n        \"\"\"\n        for k, v in self.meters.iteritems():\n            tb_logger.log_value(prefix + k, v.val, step=step)\n\n\ndef encode_data(model, data_loader, log_step=10, logging=print):\n    \"\"\"Encode all images and captions loadable by `data_loader`\n    \"\"\"\n    batch_time = AverageMeter()\n    val_logger = LogCollector()\n\n    # switch to evaluate mode\n    model.val_start()\n\n    end = time.time()\n\n    # numpy array to keep all the embeddings\n    img_embs = None\n    cap_embs = None\n    for i, (images, captions, lengths, ids) in enumerate(data_loader):\n        # make sure val logger is used\n        model.logger = val_logger\n\n        # compute the embeddings\n        img_emb, cap_emb = model.forward_emb(images, captions, lengths,\n                                             volatile=True)\n\n        # initialize the numpy arrays given the size of the embeddings\n        if img_embs is None:\n            img_embs = np.zeros((len(data_loader.dataset), img_emb.size(1)))\n            cap_embs = np.zeros((len(data_loader.dataset), cap_emb.size(1)))\n\n        # preserve the embeddings by copying from gpu and converting to numpy\n        img_embs[ids] = img_emb.data.cpu().numpy().copy()\n        cap_embs[ids] = cap_emb.data.cpu().numpy().copy()\n\n        # measure accuracy and record loss\n        model.forward_loss(img_emb, cap_emb)\n\n        # measure elapsed time\n        batch_time.update(time.time() - end)\n        end = time.time()\n\n        if i % log_step == 0:\n            logging('Test: [{0}/{1}]\\t'\n                    '{e_log}\\t'\n                    'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\\t'\n                    .format(\n                        i, len(data_loader), batch_time=batch_time,\n                        e_log=str(model.logger)))\n        del images, captions\n\n    return img_embs, cap_embs\n\n\ndef evalrank(model_path, data_path=None, split='dev', fold5=False):\n    \"\"\"\n    Evaluate a trained model on either dev or test. If `fold5=True`, 5 fold\n    cross-validation is done (only for MSCOCO). Otherwise, the full data is\n    used for evaluation.\n    \"\"\"\n    # load model and options\n    checkpoint = torch.load(model_path)\n    opt = checkpoint['opt']\n    if data_path is not None:\n        opt.data_path = data_path\n\n    # load vocabulary used by the model\n    with open(os.path.join(opt.vocab_path,\n                           '%s_vocab.pkl' % opt.data_name), 'rb') as f:\n        vocab = pickle.load(f)\n    opt.vocab_size = len(vocab)\n\n    # construct model\n    model = VSE(opt)\n\n    # load model state\n    model.load_state_dict(checkpoint['model'])\n\n    print('Loading dataset')\n    data_loader = get_test_loader(split, opt.data_name, vocab, opt.crop_size,\n                                  opt.batch_size, opt.workers, opt)\n\n    print('Computing results...')\n    img_embs, cap_embs = encode_data(model, data_loader)\n    print('Images: %d, Captions: %d' %\n          (img_embs.shape[0] / 5, cap_embs.shape[0]))\n\n    if not fold5:\n        # no cross-validation, full evaluation\n        r, rt = i2t(img_embs, cap_embs, measure=opt.measure, return_ranks=True)\n        ri, rti = t2i(img_embs, cap_embs,\n                      measure=opt.measure, return_ranks=True)\n        ar = (r[0] + r[1] + r[2]) / 3\n        ari = (ri[0] + ri[1] + ri[2]) / 3\n        rsum = r[0] + r[1] + r[2] + ri[0] + ri[1] + ri[2]\n        print(\"rsum: %.1f\" % rsum)\n        print(\"Average i2t Recall: %.1f\" % ar)\n        print(\"Image to text: %.1f %.1f %.1f %.1f %.1f\" % r)\n        print(\"Average t2i Recall: %.1f\" % ari)\n        print(\"Text to image: %.1f %.1f %.1f %.1f %.1f\" % ri)\n    else:\n        # 5fold cross-validation, only for MSCOCO\n        results = []\n        for i in range(5):\n            r, rt0 = i2t(img_embs[i * 5000:(i + 1) * 5000],\n                         cap_embs[i * 5000:(i + 1) *\n                                  5000], measure=opt.measure,\n                         return_ranks=True)\n            print(\"Image to text: %.1f, %.1f, %.1f, %.1f, %.1f\" % r)\n            ri, rti0 = t2i(img_embs[i * 5000:(i + 1) * 5000],\n                           cap_embs[i * 5000:(i + 1) *\n                                    5000], measure=opt.measure,\n                           return_ranks=True)\n            if i == 0:\n                rt, rti = rt0, rti0\n            print(\"Text to image: %.1f, %.1f, %.1f, %.1f, %.1f\" % ri)\n            ar = (r[0] + r[1] + r[2]) / 3\n            ari = (ri[0] + ri[1] + ri[2]) / 3\n            rsum = r[0] + r[1] + r[2] + ri[0] + ri[1] + ri[2]\n            print(\"rsum: %.1f ar: %.1f ari: %.1f\" % (rsum, ar, ari))\n            results += [list(r) + list(ri) + [rsum, ar, ari]]\n\n        print(\"-----------------------------------\")\n        print(\"Mean metrics: \")\n        mean_metrics = tuple(np.array(results).mean(axis=0).flatten())\n        print(\"rsum: %.1f\" % (mean_metrics[10] * 6))\n        print(\"Average i2t Recall: %.1f\" % mean_metrics[11])\n        print(\"Image to text: %.1f %.1f %.1f %.1f %.1f\" %\n              mean_metrics[:5])\n        print(\"Average t2i Recall: %.1f\" % mean_metrics[12])\n        print(\"Text to image: %.1f %.1f %.1f %.1f %.1f\" %\n              mean_metrics[5:10])\n\n    torch.save({'rt': rt, 'rti': rti}, 'ranks.pth.tar')\n\n\ndef i2t(images, captions, npts=None, measure='cosine', return_ranks=False):\n    \"\"\"\n    Images->Text (Image Annotation)\n    Images: (5N, K) matrix of images\n    Captions: (5N, K) matrix of captions\n    \"\"\"\n    if npts is None:\n        npts = images.shape[0] / 5\n    index_list = []\n\n    ranks = numpy.zeros(npts)\n    top1 = numpy.zeros(npts)\n    for index in range(npts):\n\n        # Get query image\n        im = images[5 * index].reshape(1, images.shape[1])\n\n        # Compute scores\n        if measure == 'order':\n            bs = 100\n            if index % bs == 0:\n                mx = min(images.shape[0], 5 * (index + bs))\n                im2 = images[5 * index:mx:5]\n                d2 = order_sim(torch.Tensor(im2).cuda(),\n                               torch.Tensor(captions).cuda())\n                d2 = d2.cpu().numpy()\n            d = d2[index % bs]\n        else:\n            d = numpy.dot(im, captions.T).flatten()\n        inds = numpy.argsort(d)[::-1]\n        index_list.append(inds[0])\n\n        # Score\n        rank = 1e20\n        for i in range(5 * index, 5 * index + 5, 1):\n            tmp = numpy.where(inds == i)[0][0]\n            if tmp < rank:\n                rank = tmp\n        ranks[index] = rank\n        top1[index] = inds[0]\n\n    # Compute metrics\n    r1 = 100.0 * len(numpy.where(ranks < 1)[0]) / len(ranks)\n    r5 = 100.0 * len(numpy.where(ranks < 5)[0]) / len(ranks)\n    r10 = 100.0 * len(numpy.where(ranks < 10)[0]) / len(ranks)\n    medr = numpy.floor(numpy.median(ranks)) + 1\n    meanr = ranks.mean() + 1\n    if return_ranks:\n        return (r1, r5, r10, medr, meanr), (ranks, top1)\n    else:\n        return (r1, r5, r10, medr, meanr)\n\n\ndef t2i(images, captions, npts=None, measure='cosine', return_ranks=False):\n    \"\"\"\n    Text->Images (Image Search)\n    Images: (5N, K) matrix of images\n    Captions: (5N, K) matrix of captions\n    \"\"\"\n    if npts is None:\n        npts = images.shape[0] / 5\n    ims = numpy.array([images[i] for i in range(0, len(images), 5)])\n\n    ranks = numpy.zeros(5 * npts)\n    top1 = numpy.zeros(5 * npts)\n    for index in range(npts):\n\n        # Get query captions\n        queries = captions[5 * index:5 * index + 5]\n\n        # Compute scores\n        if measure == 'order':\n            bs = 100\n            if 5 * index % bs == 0:\n                mx = min(captions.shape[0], 5 * index + bs)\n                q2 = captions[5 * index:mx]\n                d2 = order_sim(torch.Tensor(ims).cuda(),\n                               torch.Tensor(q2).cuda())\n                d2 = d2.cpu().numpy()\n\n            d = d2[:, (5 * index) % bs:(5 * index) % bs + 5].T\n        else:\n            d = numpy.dot(queries, ims.T)\n        inds = numpy.zeros(d.shape)\n        for i in range(len(inds)):\n            inds[i] = numpy.argsort(d[i])[::-1]\n            ranks[5 * index + i] = numpy.where(inds[i] == index)[0][0]\n            top1[5 * index + i] = inds[i][0]\n\n    # Compute metrics\n    r1 = 100.0 * len(numpy.where(ranks < 1)[0]) / len(ranks)\n    r5 = 100.0 * len(numpy.where(ranks < 5)[0]) / len(ranks)\n    r10 = 100.0 * len(numpy.where(ranks < 10)[0]) / len(ranks)\n    medr = numpy.floor(numpy.median(ranks)) + 1\n    meanr = ranks.mean() + 1\n    if return_ranks:\n        return (r1, r5, r10, medr, meanr), (ranks, top1)\n    else:\n        return (r1, r5, r10, medr, meanr)\n"
  },
  {
    "path": "model.py",
    "content": "import torch\nimport torch.nn as nn\nimport torch.nn.init\nimport torchvision.models as models\nfrom torch.autograd import Variable\nfrom torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence\nimport torch.backends.cudnn as cudnn\nfrom torch.nn.utils.clip_grad import clip_grad_norm\nimport numpy as np\nfrom collections import OrderedDict\n\n\ndef l2norm(X):\n    \"\"\"L2-normalize columns of X\n    \"\"\"\n    norm = torch.pow(X, 2).sum(dim=1, keepdim=True).sqrt()\n    X = torch.div(X, norm)\n    return X\n\n\ndef EncoderImage(data_name, img_dim, embed_size, finetune=False,\n                 cnn_type='vgg19', use_abs=False, no_imgnorm=False):\n    \"\"\"A wrapper to image encoders. Chooses between an encoder that uses\n    precomputed image features, `EncoderImagePrecomp`, or an encoder that\n    computes image features on the fly `EncoderImageFull`.\n    \"\"\"\n    if data_name.endswith('_precomp'):\n        img_enc = EncoderImagePrecomp(\n            img_dim, embed_size, use_abs, no_imgnorm)\n    else:\n        img_enc = EncoderImageFull(\n            embed_size, finetune, cnn_type, use_abs, no_imgnorm)\n\n    return img_enc\n\n\n# tutorials/09 - Image Captioning\nclass EncoderImageFull(nn.Module):\n\n    def __init__(self, embed_size, finetune=False, cnn_type='vgg19',\n                 use_abs=False, no_imgnorm=False):\n        \"\"\"Load pretrained VGG19 and replace top fc layer.\"\"\"\n        super(EncoderImageFull, self).__init__()\n        self.embed_size = embed_size\n        self.no_imgnorm = no_imgnorm\n        self.use_abs = use_abs\n\n        # Load a pre-trained model\n        self.cnn = self.get_cnn(cnn_type, True)\n\n        # For efficient memory usage.\n        for param in self.cnn.parameters():\n            param.requires_grad = finetune\n\n        # Replace the last fully connected layer of CNN with a new one\n        if cnn_type.startswith('vgg'):\n            self.fc = nn.Linear(self.cnn.classifier._modules['6'].in_features,\n                                embed_size)\n            self.cnn.classifier = nn.Sequential(\n                *list(self.cnn.classifier.children())[:-1])\n        elif cnn_type.startswith('resnet'):\n            self.fc = nn.Linear(self.cnn.module.fc.in_features, embed_size)\n            self.cnn.module.fc = nn.Sequential()\n\n        self.init_weights()\n\n    def get_cnn(self, arch, pretrained):\n        \"\"\"Load a pretrained CNN and parallelize over GPUs\n        \"\"\"\n        if pretrained:\n            print(\"=> using pre-trained model '{}'\".format(arch))\n            model = models.__dict__[arch](pretrained=True)\n        else:\n            print(\"=> creating model '{}'\".format(arch))\n            model = models.__dict__[arch]()\n\n        if arch.startswith('alexnet') or arch.startswith('vgg'):\n            model.features = nn.DataParallel(model.features)\n            model.cuda()\n        else:\n            model = nn.DataParallel(model).cuda()\n\n        return model\n\n    def load_state_dict(self, state_dict):\n        \"\"\"\n        Handle the models saved before commit pytorch/vision@989d52a\n        \"\"\"\n        if 'cnn.classifier.1.weight' in state_dict:\n            state_dict['cnn.classifier.0.weight'] = state_dict[\n                'cnn.classifier.1.weight']\n            del state_dict['cnn.classifier.1.weight']\n            state_dict['cnn.classifier.0.bias'] = state_dict[\n                'cnn.classifier.1.bias']\n            del state_dict['cnn.classifier.1.bias']\n            state_dict['cnn.classifier.3.weight'] = state_dict[\n                'cnn.classifier.4.weight']\n            del state_dict['cnn.classifier.4.weight']\n            state_dict['cnn.classifier.3.bias'] = state_dict[\n                'cnn.classifier.4.bias']\n            del state_dict['cnn.classifier.4.bias']\n\n        super(EncoderImageFull, self).load_state_dict(state_dict)\n\n    def init_weights(self):\n        \"\"\"Xavier initialization for the fully connected layer\n        \"\"\"\n        r = np.sqrt(6.) / np.sqrt(self.fc.in_features +\n                                  self.fc.out_features)\n        self.fc.weight.data.uniform_(-r, r)\n        self.fc.bias.data.fill_(0)\n\n    def forward(self, images):\n        \"\"\"Extract image feature vectors.\"\"\"\n        features = self.cnn(images)\n\n        # normalization in the image embedding space\n        features = l2norm(features)\n\n        # linear projection to the joint embedding space\n        features = self.fc(features)\n\n        # normalization in the joint embedding space\n        if not self.no_imgnorm:\n            features = l2norm(features)\n\n        # take the absolute value of the embedding (used in order embeddings)\n        if self.use_abs:\n            features = torch.abs(features)\n\n        return features\n\n\nclass EncoderImagePrecomp(nn.Module):\n\n    def __init__(self, img_dim, embed_size, use_abs=False, no_imgnorm=False):\n        super(EncoderImagePrecomp, self).__init__()\n        self.embed_size = embed_size\n        self.no_imgnorm = no_imgnorm\n        self.use_abs = use_abs\n\n        self.fc = nn.Linear(img_dim, embed_size)\n\n        self.init_weights()\n\n    def init_weights(self):\n        \"\"\"Xavier initialization for the fully connected layer\n        \"\"\"\n        r = np.sqrt(6.) / np.sqrt(self.fc.in_features +\n                                  self.fc.out_features)\n        self.fc.weight.data.uniform_(-r, r)\n        self.fc.bias.data.fill_(0)\n\n    def forward(self, images):\n        \"\"\"Extract image feature vectors.\"\"\"\n        # assuming that the precomputed features are already l2-normalized\n\n        features = self.fc(images)\n\n        # normalize in the joint embedding space\n        if not self.no_imgnorm:\n            features = l2norm(features)\n\n        # take the absolute value of embedding (used in order embeddings)\n        if self.use_abs:\n            features = torch.abs(features)\n\n        return features\n\n    def load_state_dict(self, state_dict):\n        \"\"\"Copies parameters. overwritting the default one to\n        accept state_dict from Full model\n        \"\"\"\n        own_state = self.state_dict()\n        new_state = OrderedDict()\n        for name, param in state_dict.items():\n            if name in own_state:\n                new_state[name] = param\n\n        super(EncoderImagePrecomp, self).load_state_dict(new_state)\n\n\n# tutorials/08 - Language Model\n# RNN Based Language Model\nclass EncoderText(nn.Module):\n\n    def __init__(self, vocab_size, word_dim, embed_size, num_layers,\n                 use_abs=False):\n        super(EncoderText, self).__init__()\n        self.use_abs = use_abs\n        self.embed_size = embed_size\n\n        # word embedding\n        self.embed = nn.Embedding(vocab_size, word_dim)\n\n        # caption embedding\n        self.rnn = nn.GRU(word_dim, embed_size, num_layers, batch_first=True)\n\n        self.init_weights()\n\n    def init_weights(self):\n        self.embed.weight.data.uniform_(-0.1, 0.1)\n\n    def forward(self, x, lengths):\n        \"\"\"Handles variable size captions\n        \"\"\"\n        # Embed word ids to vectors\n        x = self.embed(x)\n        packed = pack_padded_sequence(x, lengths, batch_first=True)\n\n        # Forward propagate RNN\n        out, _ = self.rnn(packed)\n\n        # Reshape *final* output to (batch_size, hidden_size)\n        padded = pad_packed_sequence(out, batch_first=True)\n        I = torch.LongTensor(lengths).view(-1, 1, 1)\n        I = Variable(I.expand(x.size(0), 1, self.embed_size)-1).cuda()\n        out = torch.gather(padded[0], 1, I).squeeze(1)\n\n        # normalization in the joint embedding space\n        out = l2norm(out)\n\n        # take absolute value, used by order embeddings\n        if self.use_abs:\n            out = torch.abs(out)\n\n        return out\n\n\ndef cosine_sim(im, s):\n    \"\"\"Cosine similarity between all the image and sentence pairs\n    \"\"\"\n    return im.mm(s.t())\n\n\ndef order_sim(im, s):\n    \"\"\"Order embeddings similarity measure $max(0, s-im)$\n    \"\"\"\n    YmX = (s.unsqueeze(1).expand(s.size(0), im.size(0), s.size(1))\n           - im.unsqueeze(0).expand(s.size(0), im.size(0), s.size(1)))\n    score = -YmX.clamp(min=0).pow(2).sum(2).sqrt().t()\n    return score\n\n\nclass ContrastiveLoss(nn.Module):\n    \"\"\"\n    Compute contrastive loss\n    \"\"\"\n\n    def __init__(self, margin=0, measure=False, max_violation=False):\n        super(ContrastiveLoss, self).__init__()\n        self.margin = margin\n        if measure == 'order':\n            self.sim = order_sim\n        else:\n            self.sim = cosine_sim\n\n        self.max_violation = max_violation\n\n    def forward(self, im, s):\n        # compute image-sentence score matrix\n        scores = self.sim(im, s)\n        diagonal = scores.diag().view(im.size(0), 1)\n        d1 = diagonal.expand_as(scores)\n        d2 = diagonal.t().expand_as(scores)\n\n        # compare every diagonal score to scores in its column\n        # caption retrieval\n        cost_s = (self.margin + scores - d1).clamp(min=0)\n        # compare every diagonal score to scores in its row\n        # image retrieval\n        cost_im = (self.margin + scores - d2).clamp(min=0)\n\n        # clear diagonals\n        mask = torch.eye(scores.size(0)) > .5\n        I = Variable(mask)\n        if torch.cuda.is_available():\n            I = I.cuda()\n        cost_s = cost_s.masked_fill_(I, 0)\n        cost_im = cost_im.masked_fill_(I, 0)\n\n        # keep the maximum violating negative for each query\n        if self.max_violation:\n            cost_s = cost_s.max(1)[0]\n            cost_im = cost_im.max(0)[0]\n\n        return cost_s.sum() + cost_im.sum()\n\n\nclass VSE(object):\n    \"\"\"\n    rkiros/uvs model\n    \"\"\"\n\n    def __init__(self, opt):\n        # tutorials/09 - Image Captioning\n        # Build Models\n        self.grad_clip = opt.grad_clip\n        self.img_enc = EncoderImage(opt.data_name, opt.img_dim, opt.embed_size,\n                                    opt.finetune, opt.cnn_type,\n                                    use_abs=opt.use_abs,\n                                    no_imgnorm=opt.no_imgnorm)\n        self.txt_enc = EncoderText(opt.vocab_size, opt.word_dim,\n                                   opt.embed_size, opt.num_layers,\n                                   use_abs=opt.use_abs)\n        if torch.cuda.is_available():\n            self.img_enc.cuda()\n            self.txt_enc.cuda()\n            cudnn.benchmark = True\n\n        # Loss and Optimizer\n        self.criterion = ContrastiveLoss(margin=opt.margin,\n                                         measure=opt.measure,\n                                         max_violation=opt.max_violation)\n        params = list(self.txt_enc.parameters())\n        params += list(self.img_enc.fc.parameters())\n        if opt.finetune:\n            params += list(self.img_enc.cnn.parameters())\n        self.params = params\n\n        self.optimizer = torch.optim.Adam(params, lr=opt.learning_rate)\n\n        self.Eiters = 0\n\n    def state_dict(self):\n        state_dict = [self.img_enc.state_dict(), self.txt_enc.state_dict()]\n        return state_dict\n\n    def load_state_dict(self, state_dict):\n        self.img_enc.load_state_dict(state_dict[0])\n        self.txt_enc.load_state_dict(state_dict[1])\n\n    def train_start(self):\n        \"\"\"switch to train mode\n        \"\"\"\n        self.img_enc.train()\n        self.txt_enc.train()\n\n    def val_start(self):\n        \"\"\"switch to evaluate mode\n        \"\"\"\n        self.img_enc.eval()\n        self.txt_enc.eval()\n\n    def forward_emb(self, images, captions, lengths, volatile=False):\n        \"\"\"Compute the image and caption embeddings\n        \"\"\"\n        # Set mini-batch dataset\n        images = Variable(images, volatile=volatile)\n        captions = Variable(captions, volatile=volatile)\n        if torch.cuda.is_available():\n            images = images.cuda()\n            captions = captions.cuda()\n\n        # Forward\n        img_emb = self.img_enc(images)\n        cap_emb = self.txt_enc(captions, lengths)\n        return img_emb, cap_emb\n\n    def forward_loss(self, img_emb, cap_emb, **kwargs):\n        \"\"\"Compute the loss given pairs of image and caption embeddings\n        \"\"\"\n        loss = self.criterion(img_emb, cap_emb)\n        self.logger.update('Le', loss.data[0], img_emb.size(0))\n        return loss\n\n    def train_emb(self, images, captions, lengths, ids=None, *args):\n        \"\"\"One training step given images and captions.\n        \"\"\"\n        self.Eiters += 1\n        self.logger.update('Eit', self.Eiters)\n        self.logger.update('lr', self.optimizer.param_groups[0]['lr'])\n\n        # compute the embeddings\n        img_emb, cap_emb = self.forward_emb(images, captions, lengths)\n\n        # measure accuracy and record loss\n        self.optimizer.zero_grad()\n        loss = self.forward_loss(img_emb, cap_emb)\n\n        # compute gradient and do SGD step\n        loss.backward()\n        if self.grad_clip > 0:\n            clip_grad_norm(self.params, self.grad_clip)\n        self.optimizer.step()\n"
  },
  {
    "path": "train.py",
    "content": "import pickle\nimport os\nimport time\nimport shutil\n\nimport torch\n\nimport data\nfrom vocab import Vocabulary  # NOQA\nfrom model import VSE\nfrom evaluation import i2t, t2i, AverageMeter, LogCollector, encode_data\n\nimport logging\nimport tensorboard_logger as tb_logger\n\nimport argparse\n\n\ndef main():\n    # Hyper Parameters\n    parser = argparse.ArgumentParser()\n    parser.add_argument('--data_path', default='/w/31/faghri/vsepp_data/',\n                        help='path to datasets')\n    parser.add_argument('--data_name', default='precomp',\n                        help='{coco,f8k,f30k,10crop}_precomp|coco|f8k|f30k')\n    parser.add_argument('--vocab_path', default='./vocab/',\n                        help='Path to saved vocabulary pickle files.')\n    parser.add_argument('--margin', default=0.2, type=float,\n                        help='Rank loss margin.')\n    parser.add_argument('--num_epochs', default=30, type=int,\n                        help='Number of training epochs.')\n    parser.add_argument('--batch_size', default=128, type=int,\n                        help='Size of a training mini-batch.')\n    parser.add_argument('--word_dim', default=300, type=int,\n                        help='Dimensionality of the word embedding.')\n    parser.add_argument('--embed_size', default=1024, type=int,\n                        help='Dimensionality of the joint embedding.')\n    parser.add_argument('--grad_clip', default=2., type=float,\n                        help='Gradient clipping threshold.')\n    parser.add_argument('--crop_size', default=224, type=int,\n                        help='Size of an image crop as the CNN input.')\n    parser.add_argument('--num_layers', default=1, type=int,\n                        help='Number of GRU layers.')\n    parser.add_argument('--learning_rate', default=.0002, type=float,\n                        help='Initial learning rate.')\n    parser.add_argument('--lr_update', default=15, type=int,\n                        help='Number of epochs to update the learning rate.')\n    parser.add_argument('--workers', default=10, type=int,\n                        help='Number of data loader workers.')\n    parser.add_argument('--log_step', default=10, type=int,\n                        help='Number of steps to print and record the log.')\n    parser.add_argument('--val_step', default=500, type=int,\n                        help='Number of steps to run validation.')\n    parser.add_argument('--logger_name', default='runs/runX',\n                        help='Path to save the model and Tensorboard log.')\n    parser.add_argument('--resume', default='', type=str, metavar='PATH',\n                        help='path to latest checkpoint (default: none)')\n    parser.add_argument('--max_violation', action='store_true',\n                        help='Use max instead of sum in the rank loss.')\n    parser.add_argument('--img_dim', default=4096, type=int,\n                        help='Dimensionality of the image embedding.')\n    parser.add_argument('--finetune', action='store_true',\n                        help='Fine-tune the image encoder.')\n    parser.add_argument('--cnn_type', default='vgg19',\n                        help=\"\"\"The CNN used for image encoder\n                        (e.g. vgg19, resnet152)\"\"\")\n    parser.add_argument('--use_restval', action='store_true',\n                        help='Use the restval data for training on MSCOCO.')\n    parser.add_argument('--measure', default='cosine',\n                        help='Similarity measure used (cosine|order)')\n    parser.add_argument('--use_abs', action='store_true',\n                        help='Take the absolute value of embedding vectors.')\n    parser.add_argument('--no_imgnorm', action='store_true',\n                        help='Do not normalize the image embeddings.')\n    parser.add_argument('--reset_train', action='store_true',\n                        help='Ensure the training is always done in '\n                        'train mode (Not recommended).')\n    opt = parser.parse_args()\n    print(opt)\n\n    logging.basicConfig(format='%(asctime)s %(message)s', level=logging.INFO)\n    tb_logger.configure(opt.logger_name, flush_secs=5)\n\n    # Load Vocabulary Wrapper\n    vocab = pickle.load(open(os.path.join(\n        opt.vocab_path, '%s_vocab.pkl' % opt.data_name), 'rb'))\n    opt.vocab_size = len(vocab)\n\n    # Load data loaders\n    train_loader, val_loader = data.get_loaders(\n        opt.data_name, vocab, opt.crop_size, opt.batch_size, opt.workers, opt)\n\n    # Construct the model\n    model = VSE(opt)\n\n    # optionally resume from a checkpoint\n    if opt.resume:\n        if os.path.isfile(opt.resume):\n            print(\"=> loading checkpoint '{}'\".format(opt.resume))\n            checkpoint = torch.load(opt.resume)\n            start_epoch = checkpoint['epoch']\n            best_rsum = checkpoint['best_rsum']\n            model.load_state_dict(checkpoint['model'])\n            # Eiters is used to show logs as the continuation of another\n            # training\n            model.Eiters = checkpoint['Eiters']\n            print(\"=> loaded checkpoint '{}' (epoch {}, best_rsum {})\"\n                  .format(opt.resume, start_epoch, best_rsum))\n            validate(opt, val_loader, model)\n        else:\n            print(\"=> no checkpoint found at '{}'\".format(opt.resume))\n\n    # Train the Model\n    best_rsum = 0\n    for epoch in range(opt.num_epochs):\n        adjust_learning_rate(opt, model.optimizer, epoch)\n\n        # train for one epoch\n        train(opt, train_loader, model, epoch, val_loader)\n\n        # evaluate on validation set\n        rsum = validate(opt, val_loader, model)\n\n        # remember best R@ sum and save checkpoint\n        is_best = rsum > best_rsum\n        best_rsum = max(rsum, best_rsum)\n        save_checkpoint({\n            'epoch': epoch + 1,\n            'model': model.state_dict(),\n            'best_rsum': best_rsum,\n            'opt': opt,\n            'Eiters': model.Eiters,\n        }, is_best, prefix=opt.logger_name + '/')\n\n\ndef train(opt, train_loader, model, epoch, val_loader):\n    # average meters to record the training statistics\n    batch_time = AverageMeter()\n    data_time = AverageMeter()\n    train_logger = LogCollector()\n\n    # switch to train mode\n    model.train_start()\n\n    end = time.time()\n    for i, train_data in enumerate(train_loader):\n        if opt.reset_train:\n            # Always reset to train mode, this is not the default behavior\n            model.train_start()\n\n        # measure data loading time\n        data_time.update(time.time() - end)\n\n        # make sure train logger is used\n        model.logger = train_logger\n\n        # Update the model\n        model.train_emb(*train_data)\n\n        # measure elapsed time\n        batch_time.update(time.time() - end)\n        end = time.time()\n\n        # Print log info\n        if model.Eiters % opt.log_step == 0:\n            logging.info(\n                'Epoch: [{0}][{1}/{2}]\\t'\n                '{e_log}\\t'\n                'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\\t'\n                'Data {data_time.val:.3f} ({data_time.avg:.3f})\\t'\n                .format(\n                    epoch, i, len(train_loader), batch_time=batch_time,\n                    data_time=data_time, e_log=str(model.logger)))\n\n        # Record logs in tensorboard\n        tb_logger.log_value('epoch', epoch, step=model.Eiters)\n        tb_logger.log_value('step', i, step=model.Eiters)\n        tb_logger.log_value('batch_time', batch_time.val, step=model.Eiters)\n        tb_logger.log_value('data_time', data_time.val, step=model.Eiters)\n        model.logger.tb_log(tb_logger, step=model.Eiters)\n\n        # validate at every val_step\n        if model.Eiters % opt.val_step == 0:\n            validate(opt, val_loader, model)\n\n\ndef validate(opt, val_loader, model):\n    # compute the encoding for all the validation images and captions\n    img_embs, cap_embs = encode_data(\n        model, val_loader, opt.log_step, logging.info)\n\n    # caption retrieval\n    (r1, r5, r10, medr, meanr) = i2t(img_embs, cap_embs, measure=opt.measure)\n    logging.info(\"Image to text: %.1f, %.1f, %.1f, %.1f, %.1f\" %\n                 (r1, r5, r10, medr, meanr))\n    # image retrieval\n    (r1i, r5i, r10i, medri, meanri) = t2i(\n        img_embs, cap_embs, measure=opt.measure)\n    logging.info(\"Text to image: %.1f, %.1f, %.1f, %.1f, %.1f\" %\n                 (r1i, r5i, r10i, medri, meanri))\n    # sum of recalls to be used for early stopping\n    currscore = r1 + r5 + r10 + r1i + r5i + r10i\n\n    # record metrics in tensorboard\n    tb_logger.log_value('r1', r1, step=model.Eiters)\n    tb_logger.log_value('r5', r5, step=model.Eiters)\n    tb_logger.log_value('r10', r10, step=model.Eiters)\n    tb_logger.log_value('medr', medr, step=model.Eiters)\n    tb_logger.log_value('meanr', meanr, step=model.Eiters)\n    tb_logger.log_value('r1i', r1i, step=model.Eiters)\n    tb_logger.log_value('r5i', r5i, step=model.Eiters)\n    tb_logger.log_value('r10i', r10i, step=model.Eiters)\n    tb_logger.log_value('medri', medri, step=model.Eiters)\n    tb_logger.log_value('meanri', meanri, step=model.Eiters)\n    tb_logger.log_value('rsum', currscore, step=model.Eiters)\n\n    return currscore\n\n\ndef save_checkpoint(state, is_best, filename='checkpoint.pth.tar', prefix=''):\n    torch.save(state, prefix + filename)\n    if is_best:\n        shutil.copyfile(prefix + filename, prefix + 'model_best.pth.tar')\n\n\ndef adjust_learning_rate(opt, optimizer, epoch):\n    \"\"\"Sets the learning rate to the initial LR\n       decayed by 10 every 30 epochs\"\"\"\n    lr = opt.learning_rate * (0.1 ** (epoch // opt.lr_update))\n    for param_group in optimizer.param_groups:\n        param_group['lr'] = lr\n\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\n\nif __name__ == '__main__':\n    main()\n"
  },
  {
    "path": "vocab.py",
    "content": "# Create a vocabulary wrapper\nimport nltk\nimport pickle\nfrom collections import Counter\nfrom pycocotools.coco import COCO\nimport json\nimport argparse\nimport os\n\nannotations = {\n    'coco_precomp': ['train_caps.txt', 'dev_caps.txt'],\n    'coco': ['annotations/captions_train2014.json',\n             'annotations/captions_val2014.json'],\n    'f8k_precomp': ['train_caps.txt', 'dev_caps.txt'],\n    '10crop_precomp': ['train_caps.txt', 'dev_caps.txt'],\n    'f30k_precomp': ['train_caps.txt', 'dev_caps.txt'],\n    'f8k': ['dataset_flickr8k.json'],\n    'f30k': ['dataset_flickr30k.json'],\n}\n\n\nclass Vocabulary(object):\n    \"\"\"Simple vocabulary wrapper.\"\"\"\n\n    def __init__(self):\n        self.word2idx = {}\n        self.idx2word = {}\n        self.idx = 0\n\n    def add_word(self, word):\n        if word not in self.word2idx:\n            self.word2idx[word] = self.idx\n            self.idx2word[self.idx] = word\n            self.idx += 1\n\n    def __call__(self, word):\n        if word not in self.word2idx:\n            return self.word2idx['<unk>']\n        return self.word2idx[word]\n\n    def __len__(self):\n        return len(self.word2idx)\n\n\ndef from_coco_json(path):\n    coco = COCO(path)\n    ids = coco.anns.keys()\n    captions = []\n    for i, idx in enumerate(ids):\n        captions.append(str(coco.anns[idx]['caption']))\n\n    return captions\n\n\ndef from_flickr_json(path):\n    dataset = json.load(open(path, 'r'))['images']\n    captions = []\n    for i, d in enumerate(dataset):\n        captions += [str(x['raw']) for x in d['sentences']]\n\n    return captions\n\n\ndef from_txt(txt):\n    captions = []\n    with open(txt, 'rb') as f:\n        for line in f:\n            captions.append(line.strip())\n    return captions\n\n\ndef build_vocab(data_path, data_name, jsons, threshold):\n    \"\"\"Build a simple vocabulary wrapper.\"\"\"\n    counter = Counter()\n    for path in jsons[data_name]:\n        full_path = os.path.join(os.path.join(data_path, data_name), path)\n        if data_name == 'coco':\n            captions = from_coco_json(full_path)\n        elif data_name == 'f8k' or data_name == 'f30k':\n            captions = from_flickr_json(full_path)\n        else:\n            captions = from_txt(full_path)\n        for i, caption in enumerate(captions):\n            tokens = nltk.tokenize.word_tokenize(\n                caption.lower().decode('utf-8'))\n            counter.update(tokens)\n\n            if i % 1000 == 0:\n                print(\"[%d/%d] tokenized the captions.\" % (i, len(captions)))\n\n    # Discard if the occurrence of the word is less than min_word_cnt.\n    words = [word for word, cnt in counter.items() if cnt >= threshold]\n\n    # Create a vocab wrapper and add some special tokens.\n    vocab = Vocabulary()\n    vocab.add_word('<pad>')\n    vocab.add_word('<start>')\n    vocab.add_word('<end>')\n    vocab.add_word('<unk>')\n\n    # Add words to the vocabulary.\n    for i, word in enumerate(words):\n        vocab.add_word(word)\n    return vocab\n\n\ndef main(data_path, data_name):\n    vocab = build_vocab(data_path, data_name, jsons=annotations, threshold=4)\n    with open('./vocab/%s_vocab.pkl' % data_name, 'wb') as f:\n        pickle.dump(vocab, f, pickle.HIGHEST_PROTOCOL)\n    print(\"Saved vocabulary file to \", './vocab/%s_vocab.pkl' % data_name)\n\n\nif __name__ == '__main__':\n    parser = argparse.ArgumentParser()\n    parser.add_argument('--data_path', default='/w/31/faghri/vsepp_data/')\n    parser.add_argument('--data_name', default='coco',\n                        help='{coco,f8k,f30k,10crop}_precomp|coco|f8k|f30k')\n    opt = parser.parse_args()\n    main(opt.data_path, opt.data_name)\n"
  }
]