Full Code of fartashf/vsepp for AI

Repository: fartashf/vsepp
Branch: master
Commit: abe382fd9c75
Files: 8
Total size: 64.3 KB

Directory structure:
gitextract_rkazkyvw/

├── .gitignore
├── LICENSE
├── README.md
├── data.py
├── evaluation.py
├── model.py
├── train.py
└── vocab.py

================================================
FILE CONTENTS
================================================

================================================
FILE: .gitignore
================================================
*.pyc
*.swp
*.ipynb_checkpoints
*.json
*.pth.tar


================================================
FILE: LICENSE
================================================
                                 Apache License
                           Version 2.0, January 2004
                        http://www.apache.org/licenses/

   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION

   1. Definitions.

      "License" shall mean the terms and conditions for use, reproduction,
      and distribution as defined by Sections 1 through 9 of this document.

      "Licensor" shall mean the copyright owner or entity authorized by
      the copyright owner that is granting the License.

      "Legal Entity" shall mean the union of the acting entity and all
      other entities that control, are controlled by, or are under common
      control with that entity. For the purposes of this definition,
      "control" means (i) the power, direct or indirect, to cause the
      direction or management of such entity, whether by contract or
      otherwise, or (ii) ownership of fifty percent (50%) or more of the
      outstanding shares, or (iii) beneficial ownership of such entity.

      "You" (or "Your") shall mean an individual or Legal Entity
      exercising permissions granted by this License.

      "Source" form shall mean the preferred form for making modifications,
      including but not limited to software source code, documentation
      source, and configuration files.

      "Object" form shall mean any form resulting from mechanical
      transformation or translation of a Source form, including but
      not limited to compiled object code, generated documentation,
      and conversions to other media types.

      "Work" shall mean the work of authorship, whether in Source or
      Object form, made available under the License, as indicated by a
      copyright notice that is included in or attached to the work
      (an example is provided in the Appendix below).

      "Derivative Works" shall mean any work, whether in Source or Object
      form, that is based on (or derived from) the Work and for which the
      editorial revisions, annotations, elaborations, or other modifications
      represent, as a whole, an original work of authorship. For the purposes
      of this License, Derivative Works shall not include works that remain
      separable from, or merely link (or bind by name) to the interfaces of,
      the Work and Derivative Works thereof.

      "Contribution" shall mean any work of authorship, including
      the original version of the Work and any modifications or additions
      to that Work or Derivative Works thereof, that is intentionally
      submitted to Licensor for inclusion in the Work by the copyright owner
      or by an individual or Legal Entity authorized to submit on behalf of
      the copyright owner. For the purposes of this definition, "submitted"
      means any form of electronic, verbal, or written communication sent
      to the Licensor or its representatives, including but not limited to
      communication on electronic mailing lists, source code control systems,
      and issue tracking systems that are managed by, or on behalf of, the
      Licensor for the purpose of discussing and improving the Work, but
      excluding communication that is conspicuously marked or otherwise
      designated in writing by the copyright owner as "Not a Contribution."

      "Contributor" shall mean Licensor and any individual or Legal Entity
      on behalf of whom a Contribution has been received by Licensor and
      subsequently incorporated within the Work.

   2. Grant of Copyright License. Subject to the terms and conditions of
      this License, each Contributor hereby grants to You a perpetual,
      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
      copyright license to reproduce, prepare Derivative Works of,
      publicly display, publicly perform, sublicense, and distribute the
      Work and such Derivative Works in Source or Object form.

   3. Grant of Patent License. Subject to the terms and conditions of
      this License, each Contributor hereby grants to You a perpetual,
      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
      (except as stated in this section) patent license to make, have made,
      use, offer to sell, sell, import, and otherwise transfer the Work,
      where such license applies only to those patent claims licensable
      by such Contributor that are necessarily infringed by their
      Contribution(s) alone or by combination of their Contribution(s)
      with the Work to which such Contribution(s) was submitted. If You
      institute patent litigation against any entity (including a
      cross-claim or counterclaim in a lawsuit) alleging that the Work
      or a Contribution incorporated within the Work constitutes direct
      or contributory patent infringement, then any patent licenses
      granted to You under this License for that Work shall terminate
      as of the date such litigation is filed.

   4. Redistribution. You may reproduce and distribute copies of the
      Work or Derivative Works thereof in any medium, with or without
      modifications, and in Source or Object form, provided that You
      meet the following conditions:

      (a) You must give any other recipients of the Work or
          Derivative Works a copy of this License; and

      (b) You must cause any modified files to carry prominent notices
          stating that You changed the files; and

      (c) You must retain, in the Source form of any Derivative Works
          that You distribute, all copyright, patent, trademark, and
          attribution notices from the Source form of the Work,
          excluding those notices that do not pertain to any part of
          the Derivative Works; and

      (d) If the Work includes a "NOTICE" text file as part of its
          distribution, then any Derivative Works that You distribute must
          include a readable copy of the attribution notices contained
          within such NOTICE file, excluding those notices that do not
          pertain to any part of the Derivative Works, in at least one
          of the following places: within a NOTICE text file distributed
          as part of the Derivative Works; within the Source form or
          documentation, if provided along with the Derivative Works; or,
          within a display generated by the Derivative Works, if and
          wherever such third-party notices normally appear. The contents
          of the NOTICE file are for informational purposes only and
          do not modify the License. You may add Your own attribution
          notices within Derivative Works that You distribute, alongside
          or as an addendum to the NOTICE text from the Work, provided
          that such additional attribution notices cannot be construed
          as modifying the License.

      You may add Your own copyright statement to Your modifications and
      may provide additional or different license terms and conditions
      for use, reproduction, or distribution of Your modifications, or
      for any such Derivative Works as a whole, provided Your use,
      reproduction, and distribution of the Work otherwise complies with
      the conditions stated in this License.

   5. Submission of Contributions. Unless You explicitly state otherwise,
      any Contribution intentionally submitted for inclusion in the Work
      by You to the Licensor shall be under the terms and conditions of
      this License, without any additional terms or conditions.
      Notwithstanding the above, nothing herein shall supersede or modify
      the terms of any separate license agreement you may have executed
      with Licensor regarding such Contributions.

   6. Trademarks. This License does not grant permission to use the trade
      names, trademarks, service marks, or product names of the Licensor,
      except as required for reasonable and customary use in describing the
      origin of the Work and reproducing the content of the NOTICE file.

   7. Disclaimer of Warranty. Unless required by applicable law or
      agreed to in writing, Licensor provides the Work (and each
      Contributor provides its Contributions) on an "AS IS" BASIS,
      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
      implied, including, without limitation, any warranties or conditions
      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
      PARTICULAR PURPOSE. You are solely responsible for determining the
      appropriateness of using or redistributing the Work and assume any
      risks associated with Your exercise of permissions under this License.

   8. Limitation of Liability. In no event and under no legal theory,
      whether in tort (including negligence), contract, or otherwise,
      unless required by applicable law (such as deliberate and grossly
      negligent acts) or agreed to in writing, shall any Contributor be
      liable to You for damages, including any direct, indirect, special,
      incidental, or consequential damages of any character arising as a
      result of this License or out of the use or inability to use the
      Work (including but not limited to damages for loss of goodwill,
      work stoppage, computer failure or malfunction, or any and all
      other commercial damages or losses), even if such Contributor
      has been advised of the possibility of such damages.

   9. Accepting Warranty or Additional Liability. While redistributing
      the Work or Derivative Works thereof, You may choose to offer,
      and charge a fee for, acceptance of support, warranty, indemnity,
      or other liability obligations and/or rights consistent with this
      License. However, in accepting such obligations, You may act only
      on Your own behalf and on Your sole responsibility, not on behalf
      of any other Contributor, and only if You agree to indemnify,
      defend, and hold each Contributor harmless for any liability
      incurred by, or claims asserted against, such Contributor by reason
      of your accepting any such warranty or additional liability.

   END OF TERMS AND CONDITIONS

   APPENDIX: How to apply the Apache License to your work.

      To apply the Apache License to your work, attach the following
      boilerplate notice, with the fields enclosed by brackets "[]"
      replaced with your own identifying information. (Don't include
      the brackets!)  The text should be enclosed in the appropriate
      comment syntax for the file format. We also recommend that a
      file or class name and description of purpose be included on the
      same "printed page" as the copyright notice for easier
      identification within third-party archives.

   Copyright [yyyy] [name of copyright owner]

   Licensed under the Apache License, Version 2.0 (the "License");
   you may not use this file except in compliance with the License.
   You may obtain a copy of the License at

       http://www.apache.org/licenses/LICENSE-2.0

   Unless required by applicable law or agreed to in writing, software
   distributed under the License is distributed on an "AS IS" BASIS,
   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
   See the License for the specific language governing permissions and
   limitations under the License.


================================================
FILE: README.md
================================================
# Improving Visual-Semantic Embeddings with Hard Negatives

Code for the image-caption retrieval methods from
**[VSE++: Improving Visual-Semantic Embeddings with Hard Negatives](https://arxiv.org/abs/1707.05612)**
*, F. Faghri, D. J. Fleet, J. R. Kiros, S. Fidler, Proceedings of the British Machine Vision Conference (BMVC),  2018. (BMVC Spotlight)*

## Dependencies
We recommended to use Anaconda for the following packages.

* Python 2.7 (Checkout branch `python3`)
* [PyTorch](http://pytorch.org/) (>0.2) (Checkout branch `pytorch4.1`)
* [NumPy](http://www.numpy.org/) (>1.12.1)
* [TensorBoard](https://github.com/TeamHG-Memex/tensorboard_logger)
* [pycocotools](https://github.com/cocodataset/cocoapi)
* [torchvision]()
* [matplotlib]()


* Punkt Sentence Tokenizer:
```python
import nltk
nltk.download()
> d punkt
```

## Download data

Download 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/).

```bash
wget http://www.cs.toronto.edu/~faghri/vsepp/vocab.tar
wget http://www.cs.toronto.edu/~faghri/vsepp/data.tar
wget http://www.cs.toronto.edu/~faghri/vsepp/runs.tar
```

We refer to the path of extracted files for `data.tar` as `$DATA_PATH` and 
files for `models.tar` as `$RUN_PATH`. Extract `vocab.tar` to `./vocab` 
directory.

*Update: The vocabulary was originally built using all sets (including test set 
captions). Please see issue #29 for details. Please consider not using test set 
captions if building up on this project.*

## Evaluate pre-trained models

```python
python -c "\
from vocab import Vocabulary
import evaluation
evaluation.evalrank('$RUN_PATH/coco_vse++/model_best.pth.tar', data_path='$DATA_PATH', split='test')"
```

To do cross-validation on MSCOCO, pass `fold5=True` with a model trained using 
`--data_name coco`.

## Training new models
Run `train.py`:

```bash
python train.py --data_path "$DATA_PATH" --data_name coco_precomp --logger_name 
runs/coco_vse++ --max_violation
```

Arguments used to train pre-trained models:

| Method    | Arguments |
| :-------: | :-------: |
| VSE0      | `--no_imgnorm` |
| VSE++     | `--max_violation` |
| Order0    | `--measure order --use_abs --margin .05 --learning_rate .001` |
| Order++   | `--measure order --max_violation` |


## Reference

If you found this code useful, please cite the following paper:

    @article{faghri2018vse++,
      title={VSE++: Improving Visual-Semantic Embeddings with Hard Negatives},
      author={Faghri, Fartash and Fleet, David J and Kiros, Jamie Ryan and Fidler, Sanja},
      booktitle = {Proceedings of the British Machine Vision Conference ({BMVC})},
      url = {https://github.com/fartashf/vsepp},
      year={2018}
    }

## License

[Apache License 2.0](http://www.apache.org/licenses/LICENSE-2.0)


================================================
FILE: data.py
================================================
import torch
import torch.utils.data as data
import torchvision.transforms as transforms
import os
import nltk
from PIL import Image
from pycocotools.coco import COCO
import numpy as np
import json as jsonmod


def get_paths(path, name='coco', use_restval=False):
    """
    Returns paths to images and annotations for the given datasets. For MSCOCO
    indices are also returned to control the data split being used.
    The indices are extracted from the Karpathy et al. splits using this
    snippet:

    >>> import json
    >>> dataset=json.load(open('dataset_coco.json','r'))
    >>> A=[]
    >>> for i in range(len(D['images'])):
    ...   if D['images'][i]['split'] == 'val':
    ...     A+=D['images'][i]['sentids'][:5]
    ...

    :param name: Dataset names
    :param use_restval: If True, the the `restval` data is included in train.
    """
    roots = {}
    ids = {}
    if 'coco' == name:
        imgdir = os.path.join(path, 'images')
        capdir = os.path.join(path, 'annotations')
        roots['train'] = {
            'img': os.path.join(imgdir, 'train2014'),
            'cap': os.path.join(capdir, 'captions_train2014.json')
        }
        roots['val'] = {
            'img': os.path.join(imgdir, 'val2014'),
            'cap': os.path.join(capdir, 'captions_val2014.json')
        }
        roots['test'] = {
            'img': os.path.join(imgdir, 'val2014'),
            'cap': os.path.join(capdir, 'captions_val2014.json')
        }
        roots['trainrestval'] = {
            'img': (roots['train']['img'], roots['val']['img']),
            'cap': (roots['train']['cap'], roots['val']['cap'])
        }
        ids['train'] = np.load(os.path.join(capdir, 'coco_train_ids.npy'))
        ids['val'] = np.load(os.path.join(capdir, 'coco_dev_ids.npy'))[:5000]
        ids['test'] = np.load(os.path.join(capdir, 'coco_test_ids.npy'))
        ids['trainrestval'] = (
            ids['train'],
            np.load(os.path.join(capdir, 'coco_restval_ids.npy')))
        if use_restval:
            roots['train'] = roots['trainrestval']
            ids['train'] = ids['trainrestval']
    elif 'f8k' == name:
        imgdir = os.path.join(path, 'images')
        cap = os.path.join(path, 'dataset_flickr8k.json')
        roots['train'] = {'img': imgdir, 'cap': cap}
        roots['val'] = {'img': imgdir, 'cap': cap}
        roots['test'] = {'img': imgdir, 'cap': cap}
        ids = {'train': None, 'val': None, 'test': None}
    elif 'f30k' == name:
        imgdir = os.path.join(path, 'images')
        cap = os.path.join(path, 'dataset_flickr30k.json')
        roots['train'] = {'img': imgdir, 'cap': cap}
        roots['val'] = {'img': imgdir, 'cap': cap}
        roots['test'] = {'img': imgdir, 'cap': cap}
        ids = {'train': None, 'val': None, 'test': None}

    return roots, ids


class CocoDataset(data.Dataset):
    """COCO Custom Dataset compatible with torch.utils.data.DataLoader."""

    def __init__(self, root, json, vocab, transform=None, ids=None):
        """
        Args:
            root: image directory.
            json: coco annotation file path.
            vocab: vocabulary wrapper.
            transform: transformer for image.
        """
        self.root = root
        # when using `restval`, two json files are needed
        if isinstance(json, tuple):
            self.coco = (COCO(json[0]), COCO(json[1]))
        else:
            self.coco = (COCO(json),)
            self.root = (root,)
        # if ids provided by get_paths, use split-specific ids
        if ids is None:
            self.ids = list(self.coco.anns.keys())
        else:
            self.ids = ids

        # if `restval` data is to be used, record the break point for ids
        if isinstance(self.ids, tuple):
            self.bp = len(self.ids[0])
            self.ids = list(self.ids[0]) + list(self.ids[1])
        else:
            self.bp = len(self.ids)
        self.vocab = vocab
        self.transform = transform

    def __getitem__(self, index):
        """This function returns a tuple that is further passed to collate_fn
        """
        vocab = self.vocab
        root, caption, img_id, path, image = self.get_raw_item(index)

        if self.transform is not None:
            image = self.transform(image)

        # Convert caption (string) to word ids.
        tokens = nltk.tokenize.word_tokenize(
            str(caption).lower().decode('utf-8'))
        caption = []
        caption.append(vocab('<start>'))
        caption.extend([vocab(token) for token in tokens])
        caption.append(vocab('<end>'))
        target = torch.Tensor(caption)
        return image, target, index, img_id

    def get_raw_item(self, index):
        if index < self.bp:
            coco = self.coco[0]
            root = self.root[0]
        else:
            coco = self.coco[1]
            root = self.root[1]
        ann_id = self.ids[index]
        caption = coco.anns[ann_id]['caption']
        img_id = coco.anns[ann_id]['image_id']
        path = coco.loadImgs(img_id)[0]['file_name']
        image = Image.open(os.path.join(root, path)).convert('RGB')

        return root, caption, img_id, path, image

    def __len__(self):
        return len(self.ids)


class FlickrDataset(data.Dataset):
    """
    Dataset loader for Flickr30k and Flickr8k full datasets.
    """

    def __init__(self, root, json, split, vocab, transform=None):
        self.root = root
        self.vocab = vocab
        self.split = split
        self.transform = transform
        self.dataset = jsonmod.load(open(json, 'r'))['images']
        self.ids = []
        for i, d in enumerate(self.dataset):
            if d['split'] == split:
                self.ids += [(i, x) for x in range(len(d['sentences']))]

    def __getitem__(self, index):
        """This function returns a tuple that is further passed to collate_fn
        """
        vocab = self.vocab
        root = self.root
        ann_id = self.ids[index]
        img_id = ann_id[0]
        caption = self.dataset[img_id]['sentences'][ann_id[1]]['raw']
        path = self.dataset[img_id]['filename']

        image = Image.open(os.path.join(root, path)).convert('RGB')
        if self.transform is not None:
            image = self.transform(image)

        # Convert caption (string) to word ids.
        tokens = nltk.tokenize.word_tokenize(
            str(caption).lower().decode('utf-8'))
        caption = []
        caption.append(vocab('<start>'))
        caption.extend([vocab(token) for token in tokens])
        caption.append(vocab('<end>'))
        target = torch.Tensor(caption)
        return image, target, index, img_id

    def __len__(self):
        return len(self.ids)


class PrecompDataset(data.Dataset):
    """
    Load precomputed captions and image features
    Possible options: f8k, f30k, coco, 10crop
    """

    def __init__(self, data_path, data_split, vocab):
        self.vocab = vocab
        loc = data_path + '/'

        # Captions
        self.captions = []
        with open(loc+'%s_caps.txt' % data_split, 'rb') as f:
            for line in f:
                self.captions.append(line.strip())

        # Image features
        self.images = np.load(loc+'%s_ims.npy' % data_split)
        self.length = len(self.captions)
        # rkiros data has redundancy in images, we divide by 5, 10crop doesn't
        if self.images.shape[0] != self.length:
            self.im_div = 5
        else:
            self.im_div = 1
        # the development set for coco is large and so validation would be slow
        if data_split == 'dev':
            self.length = 5000

    def __getitem__(self, index):
        # handle the image redundancy
        img_id = index/self.im_div
        image = torch.Tensor(self.images[img_id])
        caption = self.captions[index]
        vocab = self.vocab

        # Convert caption (string) to word ids.
        tokens = nltk.tokenize.word_tokenize(
            str(caption).lower().decode('utf-8'))
        caption = []
        caption.append(vocab('<start>'))
        caption.extend([vocab(token) for token in tokens])
        caption.append(vocab('<end>'))
        target = torch.Tensor(caption)
        return image, target, index, img_id

    def __len__(self):
        return self.length


def collate_fn(data):
    """Build mini-batch tensors from a list of (image, caption) tuples.
    Args:
        data: list of (image, caption) tuple.
            - image: torch tensor of shape (3, 256, 256).
            - caption: torch tensor of shape (?); variable length.

    Returns:
        images: torch tensor of shape (batch_size, 3, 256, 256).
        targets: torch tensor of shape (batch_size, padded_length).
        lengths: list; valid length for each padded caption.
    """
    # Sort a data list by caption length
    data.sort(key=lambda x: len(x[1]), reverse=True)
    images, captions, ids, img_ids = zip(*data)

    # Merge images (convert tuple of 3D tensor to 4D tensor)
    images = torch.stack(images, 0)

    # Merget captions (convert tuple of 1D tensor to 2D tensor)
    lengths = [len(cap) for cap in captions]
    targets = torch.zeros(len(captions), max(lengths)).long()
    for i, cap in enumerate(captions):
        end = lengths[i]
        targets[i, :end] = cap[:end]

    return images, targets, lengths, ids


def get_loader_single(data_name, split, root, json, vocab, transform,
                      batch_size=100, shuffle=True,
                      num_workers=2, ids=None, collate_fn=collate_fn):
    """Returns torch.utils.data.DataLoader for custom coco dataset."""
    if 'coco' in data_name:
        # COCO custom dataset
        dataset = CocoDataset(root=root,
                              json=json,
                              vocab=vocab,
                              transform=transform, ids=ids)
    elif 'f8k' in data_name or 'f30k' in data_name:
        dataset = FlickrDataset(root=root,
                                split=split,
                                json=json,
                                vocab=vocab,
                                transform=transform)

    # Data loader
    data_loader = torch.utils.data.DataLoader(dataset=dataset,
                                              batch_size=batch_size,
                                              shuffle=shuffle,
                                              pin_memory=True,
                                              num_workers=num_workers,
                                              collate_fn=collate_fn)
    return data_loader


def get_precomp_loader(data_path, data_split, vocab, opt, batch_size=100,
                       shuffle=True, num_workers=2):
    """Returns torch.utils.data.DataLoader for custom coco dataset."""
    dset = PrecompDataset(data_path, data_split, vocab)

    data_loader = torch.utils.data.DataLoader(dataset=dset,
                                              batch_size=batch_size,
                                              shuffle=shuffle,
                                              pin_memory=True,
                                              collate_fn=collate_fn)
    return data_loader


def get_transform(data_name, split_name, opt):
    normalizer = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                      std=[0.229, 0.224, 0.225])
    t_list = []
    if split_name == 'train':
        t_list = [transforms.RandomResizedCrop(opt.crop_size),
                  transforms.RandomHorizontalFlip()]
    elif split_name == 'val':
        t_list = [transforms.Resize(256), transforms.CenterCrop(224)]
    elif split_name == 'test':
        t_list = [transforms.Resize(256), transforms.CenterCrop(224)]

    t_end = [transforms.ToTensor(), normalizer]
    transform = transforms.Compose(t_list + t_end)
    return transform


def get_loaders(data_name, vocab, crop_size, batch_size, workers, opt):
    dpath = os.path.join(opt.data_path, data_name)
    if opt.data_name.endswith('_precomp'):
        train_loader = get_precomp_loader(dpath, 'train', vocab, opt,
                                          batch_size, True, workers)
        val_loader = get_precomp_loader(dpath, 'dev', vocab, opt,
                                        batch_size, False, workers)
    else:
        # Build Dataset Loader
        roots, ids = get_paths(dpath, data_name, opt.use_restval)

        transform = get_transform(data_name, 'train', opt)
        train_loader = get_loader_single(opt.data_name, 'train',
                                         roots['train']['img'],
                                         roots['train']['cap'],
                                         vocab, transform, ids=ids['train'],
                                         batch_size=batch_size, shuffle=True,
                                         num_workers=workers,
                                         collate_fn=collate_fn)

        transform = get_transform(data_name, 'val', opt)
        val_loader = get_loader_single(opt.data_name, 'val',
                                       roots['val']['img'],
                                       roots['val']['cap'],
                                       vocab, transform, ids=ids['val'],
                                       batch_size=batch_size, shuffle=False,
                                       num_workers=workers,
                                       collate_fn=collate_fn)

    return train_loader, val_loader


def get_test_loader(split_name, data_name, vocab, crop_size, batch_size,
                    workers, opt):
    dpath = os.path.join(opt.data_path, data_name)
    if opt.data_name.endswith('_precomp'):
        test_loader = get_precomp_loader(dpath, split_name, vocab, opt,
                                         batch_size, False, workers)
    else:
        # Build Dataset Loader
        roots, ids = get_paths(dpath, data_name, opt.use_restval)

        transform = get_transform(data_name, split_name, opt)
        test_loader = get_loader_single(opt.data_name, split_name,
                                        roots[split_name]['img'],
                                        roots[split_name]['cap'],
                                        vocab, transform, ids=ids[split_name],
                                        batch_size=batch_size, shuffle=False,
                                        num_workers=workers,
                                        collate_fn=collate_fn)

    return test_loader


================================================
FILE: evaluation.py
================================================
from __future__ import print_function
import os
import pickle

import numpy
from data import get_test_loader
import time
import numpy as np
from vocab import Vocabulary  # NOQA
import torch
from model import VSE, order_sim
from collections import OrderedDict


class AverageMeter(object):
    """Computes and stores the average and current value"""

    def __init__(self):
        self.reset()

    def reset(self):
        self.val = 0
        self.avg = 0
        self.sum = 0
        self.count = 0

    def update(self, val, n=0):
        self.val = val
        self.sum += val * n
        self.count += n
        self.avg = self.sum / (.0001 + self.count)

    def __str__(self):
        """String representation for logging
        """
        # for values that should be recorded exactly e.g. iteration number
        if self.count == 0:
            return str(self.val)
        # for stats
        return '%.4f (%.4f)' % (self.val, self.avg)


class LogCollector(object):
    """A collection of logging objects that can change from train to val"""

    def __init__(self):
        # to keep the order of logged variables deterministic
        self.meters = OrderedDict()

    def update(self, k, v, n=0):
        # create a new meter if previously not recorded
        if k not in self.meters:
            self.meters[k] = AverageMeter()
        self.meters[k].update(v, n)

    def __str__(self):
        """Concatenate the meters in one log line
        """
        s = ''
        for i, (k, v) in enumerate(self.meters.iteritems()):
            if i > 0:
                s += '  '
            s += k + ' ' + str(v)
        return s

    def tb_log(self, tb_logger, prefix='', step=None):
        """Log using tensorboard
        """
        for k, v in self.meters.iteritems():
            tb_logger.log_value(prefix + k, v.val, step=step)


def encode_data(model, data_loader, log_step=10, logging=print):
    """Encode all images and captions loadable by `data_loader`
    """
    batch_time = AverageMeter()
    val_logger = LogCollector()

    # switch to evaluate mode
    model.val_start()

    end = time.time()

    # numpy array to keep all the embeddings
    img_embs = None
    cap_embs = None
    for i, (images, captions, lengths, ids) in enumerate(data_loader):
        # make sure val logger is used
        model.logger = val_logger

        # compute the embeddings
        img_emb, cap_emb = model.forward_emb(images, captions, lengths,
                                             volatile=True)

        # initialize the numpy arrays given the size of the embeddings
        if img_embs is None:
            img_embs = np.zeros((len(data_loader.dataset), img_emb.size(1)))
            cap_embs = np.zeros((len(data_loader.dataset), cap_emb.size(1)))

        # preserve the embeddings by copying from gpu and converting to numpy
        img_embs[ids] = img_emb.data.cpu().numpy().copy()
        cap_embs[ids] = cap_emb.data.cpu().numpy().copy()

        # measure accuracy and record loss
        model.forward_loss(img_emb, cap_emb)

        # measure elapsed time
        batch_time.update(time.time() - end)
        end = time.time()

        if i % log_step == 0:
            logging('Test: [{0}/{1}]\t'
                    '{e_log}\t'
                    'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
                    .format(
                        i, len(data_loader), batch_time=batch_time,
                        e_log=str(model.logger)))
        del images, captions

    return img_embs, cap_embs


def evalrank(model_path, data_path=None, split='dev', fold5=False):
    """
    Evaluate a trained model on either dev or test. If `fold5=True`, 5 fold
    cross-validation is done (only for MSCOCO). Otherwise, the full data is
    used for evaluation.
    """
    # load model and options
    checkpoint = torch.load(model_path)
    opt = checkpoint['opt']
    if data_path is not None:
        opt.data_path = data_path

    # load vocabulary used by the model
    with open(os.path.join(opt.vocab_path,
                           '%s_vocab.pkl' % opt.data_name), 'rb') as f:
        vocab = pickle.load(f)
    opt.vocab_size = len(vocab)

    # construct model
    model = VSE(opt)

    # load model state
    model.load_state_dict(checkpoint['model'])

    print('Loading dataset')
    data_loader = get_test_loader(split, opt.data_name, vocab, opt.crop_size,
                                  opt.batch_size, opt.workers, opt)

    print('Computing results...')
    img_embs, cap_embs = encode_data(model, data_loader)
    print('Images: %d, Captions: %d' %
          (img_embs.shape[0] / 5, cap_embs.shape[0]))

    if not fold5:
        # no cross-validation, full evaluation
        r, rt = i2t(img_embs, cap_embs, measure=opt.measure, return_ranks=True)
        ri, rti = t2i(img_embs, cap_embs,
                      measure=opt.measure, return_ranks=True)
        ar = (r[0] + r[1] + r[2]) / 3
        ari = (ri[0] + ri[1] + ri[2]) / 3
        rsum = r[0] + r[1] + r[2] + ri[0] + ri[1] + ri[2]
        print("rsum: %.1f" % rsum)
        print("Average i2t Recall: %.1f" % ar)
        print("Image to text: %.1f %.1f %.1f %.1f %.1f" % r)
        print("Average t2i Recall: %.1f" % ari)
        print("Text to image: %.1f %.1f %.1f %.1f %.1f" % ri)
    else:
        # 5fold cross-validation, only for MSCOCO
        results = []
        for i in range(5):
            r, rt0 = i2t(img_embs[i * 5000:(i + 1) * 5000],
                         cap_embs[i * 5000:(i + 1) *
                                  5000], measure=opt.measure,
                         return_ranks=True)
            print("Image to text: %.1f, %.1f, %.1f, %.1f, %.1f" % r)
            ri, rti0 = t2i(img_embs[i * 5000:(i + 1) * 5000],
                           cap_embs[i * 5000:(i + 1) *
                                    5000], measure=opt.measure,
                           return_ranks=True)
            if i == 0:
                rt, rti = rt0, rti0
            print("Text to image: %.1f, %.1f, %.1f, %.1f, %.1f" % ri)
            ar = (r[0] + r[1] + r[2]) / 3
            ari = (ri[0] + ri[1] + ri[2]) / 3
            rsum = r[0] + r[1] + r[2] + ri[0] + ri[1] + ri[2]
            print("rsum: %.1f ar: %.1f ari: %.1f" % (rsum, ar, ari))
            results += [list(r) + list(ri) + [rsum, ar, ari]]

        print("-----------------------------------")
        print("Mean metrics: ")
        mean_metrics = tuple(np.array(results).mean(axis=0).flatten())
        print("rsum: %.1f" % (mean_metrics[10] * 6))
        print("Average i2t Recall: %.1f" % mean_metrics[11])
        print("Image to text: %.1f %.1f %.1f %.1f %.1f" %
              mean_metrics[:5])
        print("Average t2i Recall: %.1f" % mean_metrics[12])
        print("Text to image: %.1f %.1f %.1f %.1f %.1f" %
              mean_metrics[5:10])

    torch.save({'rt': rt, 'rti': rti}, 'ranks.pth.tar')


def i2t(images, captions, npts=None, measure='cosine', return_ranks=False):
    """
    Images->Text (Image Annotation)
    Images: (5N, K) matrix of images
    Captions: (5N, K) matrix of captions
    """
    if npts is None:
        npts = images.shape[0] / 5
    index_list = []

    ranks = numpy.zeros(npts)
    top1 = numpy.zeros(npts)
    for index in range(npts):

        # Get query image
        im = images[5 * index].reshape(1, images.shape[1])

        # Compute scores
        if measure == 'order':
            bs = 100
            if index % bs == 0:
                mx = min(images.shape[0], 5 * (index + bs))
                im2 = images[5 * index:mx:5]
                d2 = order_sim(torch.Tensor(im2).cuda(),
                               torch.Tensor(captions).cuda())
                d2 = d2.cpu().numpy()
            d = d2[index % bs]
        else:
            d = numpy.dot(im, captions.T).flatten()
        inds = numpy.argsort(d)[::-1]
        index_list.append(inds[0])

        # Score
        rank = 1e20
        for i in range(5 * index, 5 * index + 5, 1):
            tmp = numpy.where(inds == i)[0][0]
            if tmp < rank:
                rank = tmp
        ranks[index] = rank
        top1[index] = inds[0]

    # Compute metrics
    r1 = 100.0 * len(numpy.where(ranks < 1)[0]) / len(ranks)
    r5 = 100.0 * len(numpy.where(ranks < 5)[0]) / len(ranks)
    r10 = 100.0 * len(numpy.where(ranks < 10)[0]) / len(ranks)
    medr = numpy.floor(numpy.median(ranks)) + 1
    meanr = ranks.mean() + 1
    if return_ranks:
        return (r1, r5, r10, medr, meanr), (ranks, top1)
    else:
        return (r1, r5, r10, medr, meanr)


def t2i(images, captions, npts=None, measure='cosine', return_ranks=False):
    """
    Text->Images (Image Search)
    Images: (5N, K) matrix of images
    Captions: (5N, K) matrix of captions
    """
    if npts is None:
        npts = images.shape[0] / 5
    ims = numpy.array([images[i] for i in range(0, len(images), 5)])

    ranks = numpy.zeros(5 * npts)
    top1 = numpy.zeros(5 * npts)
    for index in range(npts):

        # Get query captions
        queries = captions[5 * index:5 * index + 5]

        # Compute scores
        if measure == 'order':
            bs = 100
            if 5 * index % bs == 0:
                mx = min(captions.shape[0], 5 * index + bs)
                q2 = captions[5 * index:mx]
                d2 = order_sim(torch.Tensor(ims).cuda(),
                               torch.Tensor(q2).cuda())
                d2 = d2.cpu().numpy()

            d = d2[:, (5 * index) % bs:(5 * index) % bs + 5].T
        else:
            d = numpy.dot(queries, ims.T)
        inds = numpy.zeros(d.shape)
        for i in range(len(inds)):
            inds[i] = numpy.argsort(d[i])[::-1]
            ranks[5 * index + i] = numpy.where(inds[i] == index)[0][0]
            top1[5 * index + i] = inds[i][0]

    # Compute metrics
    r1 = 100.0 * len(numpy.where(ranks < 1)[0]) / len(ranks)
    r5 = 100.0 * len(numpy.where(ranks < 5)[0]) / len(ranks)
    r10 = 100.0 * len(numpy.where(ranks < 10)[0]) / len(ranks)
    medr = numpy.floor(numpy.median(ranks)) + 1
    meanr = ranks.mean() + 1
    if return_ranks:
        return (r1, r5, r10, medr, meanr), (ranks, top1)
    else:
        return (r1, r5, r10, medr, meanr)


================================================
FILE: model.py
================================================
import torch
import torch.nn as nn
import torch.nn.init
import torchvision.models as models
from torch.autograd import Variable
from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence
import torch.backends.cudnn as cudnn
from torch.nn.utils.clip_grad import clip_grad_norm
import numpy as np
from collections import OrderedDict


def l2norm(X):
    """L2-normalize columns of X
    """
    norm = torch.pow(X, 2).sum(dim=1, keepdim=True).sqrt()
    X = torch.div(X, norm)
    return X


def EncoderImage(data_name, img_dim, embed_size, finetune=False,
                 cnn_type='vgg19', use_abs=False, no_imgnorm=False):
    """A wrapper to image encoders. Chooses between an encoder that uses
    precomputed image features, `EncoderImagePrecomp`, or an encoder that
    computes image features on the fly `EncoderImageFull`.
    """
    if data_name.endswith('_precomp'):
        img_enc = EncoderImagePrecomp(
            img_dim, embed_size, use_abs, no_imgnorm)
    else:
        img_enc = EncoderImageFull(
            embed_size, finetune, cnn_type, use_abs, no_imgnorm)

    return img_enc


# tutorials/09 - Image Captioning
class EncoderImageFull(nn.Module):

    def __init__(self, embed_size, finetune=False, cnn_type='vgg19',
                 use_abs=False, no_imgnorm=False):
        """Load pretrained VGG19 and replace top fc layer."""
        super(EncoderImageFull, self).__init__()
        self.embed_size = embed_size
        self.no_imgnorm = no_imgnorm
        self.use_abs = use_abs

        # Load a pre-trained model
        self.cnn = self.get_cnn(cnn_type, True)

        # For efficient memory usage.
        for param in self.cnn.parameters():
            param.requires_grad = finetune

        # Replace the last fully connected layer of CNN with a new one
        if cnn_type.startswith('vgg'):
            self.fc = nn.Linear(self.cnn.classifier._modules['6'].in_features,
                                embed_size)
            self.cnn.classifier = nn.Sequential(
                *list(self.cnn.classifier.children())[:-1])
        elif cnn_type.startswith('resnet'):
            self.fc = nn.Linear(self.cnn.module.fc.in_features, embed_size)
            self.cnn.module.fc = nn.Sequential()

        self.init_weights()

    def get_cnn(self, arch, pretrained):
        """Load a pretrained CNN and parallelize over GPUs
        """
        if pretrained:
            print("=> using pre-trained model '{}'".format(arch))
            model = models.__dict__[arch](pretrained=True)
        else:
            print("=> creating model '{}'".format(arch))
            model = models.__dict__[arch]()

        if arch.startswith('alexnet') or arch.startswith('vgg'):
            model.features = nn.DataParallel(model.features)
            model.cuda()
        else:
            model = nn.DataParallel(model).cuda()

        return model

    def load_state_dict(self, state_dict):
        """
        Handle the models saved before commit pytorch/vision@989d52a
        """
        if 'cnn.classifier.1.weight' in state_dict:
            state_dict['cnn.classifier.0.weight'] = state_dict[
                'cnn.classifier.1.weight']
            del state_dict['cnn.classifier.1.weight']
            state_dict['cnn.classifier.0.bias'] = state_dict[
                'cnn.classifier.1.bias']
            del state_dict['cnn.classifier.1.bias']
            state_dict['cnn.classifier.3.weight'] = state_dict[
                'cnn.classifier.4.weight']
            del state_dict['cnn.classifier.4.weight']
            state_dict['cnn.classifier.3.bias'] = state_dict[
                'cnn.classifier.4.bias']
            del state_dict['cnn.classifier.4.bias']

        super(EncoderImageFull, self).load_state_dict(state_dict)

    def init_weights(self):
        """Xavier initialization for the fully connected layer
        """
        r = np.sqrt(6.) / np.sqrt(self.fc.in_features +
                                  self.fc.out_features)
        self.fc.weight.data.uniform_(-r, r)
        self.fc.bias.data.fill_(0)

    def forward(self, images):
        """Extract image feature vectors."""
        features = self.cnn(images)

        # normalization in the image embedding space
        features = l2norm(features)

        # linear projection to the joint embedding space
        features = self.fc(features)

        # normalization in the joint embedding space
        if not self.no_imgnorm:
            features = l2norm(features)

        # take the absolute value of the embedding (used in order embeddings)
        if self.use_abs:
            features = torch.abs(features)

        return features


class EncoderImagePrecomp(nn.Module):

    def __init__(self, img_dim, embed_size, use_abs=False, no_imgnorm=False):
        super(EncoderImagePrecomp, self).__init__()
        self.embed_size = embed_size
        self.no_imgnorm = no_imgnorm
        self.use_abs = use_abs

        self.fc = nn.Linear(img_dim, embed_size)

        self.init_weights()

    def init_weights(self):
        """Xavier initialization for the fully connected layer
        """
        r = np.sqrt(6.) / np.sqrt(self.fc.in_features +
                                  self.fc.out_features)
        self.fc.weight.data.uniform_(-r, r)
        self.fc.bias.data.fill_(0)

    def forward(self, images):
        """Extract image feature vectors."""
        # assuming that the precomputed features are already l2-normalized

        features = self.fc(images)

        # normalize in the joint embedding space
        if not self.no_imgnorm:
            features = l2norm(features)

        # take the absolute value of embedding (used in order embeddings)
        if self.use_abs:
            features = torch.abs(features)

        return features

    def load_state_dict(self, state_dict):
        """Copies parameters. overwritting the default one to
        accept state_dict from Full model
        """
        own_state = self.state_dict()
        new_state = OrderedDict()
        for name, param in state_dict.items():
            if name in own_state:
                new_state[name] = param

        super(EncoderImagePrecomp, self).load_state_dict(new_state)


# tutorials/08 - Language Model
# RNN Based Language Model
class EncoderText(nn.Module):

    def __init__(self, vocab_size, word_dim, embed_size, num_layers,
                 use_abs=False):
        super(EncoderText, self).__init__()
        self.use_abs = use_abs
        self.embed_size = embed_size

        # word embedding
        self.embed = nn.Embedding(vocab_size, word_dim)

        # caption embedding
        self.rnn = nn.GRU(word_dim, embed_size, num_layers, batch_first=True)

        self.init_weights()

    def init_weights(self):
        self.embed.weight.data.uniform_(-0.1, 0.1)

    def forward(self, x, lengths):
        """Handles variable size captions
        """
        # Embed word ids to vectors
        x = self.embed(x)
        packed = pack_padded_sequence(x, lengths, batch_first=True)

        # Forward propagate RNN
        out, _ = self.rnn(packed)

        # Reshape *final* output to (batch_size, hidden_size)
        padded = pad_packed_sequence(out, batch_first=True)
        I = torch.LongTensor(lengths).view(-1, 1, 1)
        I = Variable(I.expand(x.size(0), 1, self.embed_size)-1).cuda()
        out = torch.gather(padded[0], 1, I).squeeze(1)

        # normalization in the joint embedding space
        out = l2norm(out)

        # take absolute value, used by order embeddings
        if self.use_abs:
            out = torch.abs(out)

        return out


def cosine_sim(im, s):
    """Cosine similarity between all the image and sentence pairs
    """
    return im.mm(s.t())


def order_sim(im, s):
    """Order embeddings similarity measure $max(0, s-im)$
    """
    YmX = (s.unsqueeze(1).expand(s.size(0), im.size(0), s.size(1))
           - im.unsqueeze(0).expand(s.size(0), im.size(0), s.size(1)))
    score = -YmX.clamp(min=0).pow(2).sum(2).sqrt().t()
    return score


class ContrastiveLoss(nn.Module):
    """
    Compute contrastive loss
    """

    def __init__(self, margin=0, measure=False, max_violation=False):
        super(ContrastiveLoss, self).__init__()
        self.margin = margin
        if measure == 'order':
            self.sim = order_sim
        else:
            self.sim = cosine_sim

        self.max_violation = max_violation

    def forward(self, im, s):
        # compute image-sentence score matrix
        scores = self.sim(im, s)
        diagonal = scores.diag().view(im.size(0), 1)
        d1 = diagonal.expand_as(scores)
        d2 = diagonal.t().expand_as(scores)

        # compare every diagonal score to scores in its column
        # caption retrieval
        cost_s = (self.margin + scores - d1).clamp(min=0)
        # compare every diagonal score to scores in its row
        # image retrieval
        cost_im = (self.margin + scores - d2).clamp(min=0)

        # clear diagonals
        mask = torch.eye(scores.size(0)) > .5
        I = Variable(mask)
        if torch.cuda.is_available():
            I = I.cuda()
        cost_s = cost_s.masked_fill_(I, 0)
        cost_im = cost_im.masked_fill_(I, 0)

        # keep the maximum violating negative for each query
        if self.max_violation:
            cost_s = cost_s.max(1)[0]
            cost_im = cost_im.max(0)[0]

        return cost_s.sum() + cost_im.sum()


class VSE(object):
    """
    rkiros/uvs model
    """

    def __init__(self, opt):
        # tutorials/09 - Image Captioning
        # Build Models
        self.grad_clip = opt.grad_clip
        self.img_enc = EncoderImage(opt.data_name, opt.img_dim, opt.embed_size,
                                    opt.finetune, opt.cnn_type,
                                    use_abs=opt.use_abs,
                                    no_imgnorm=opt.no_imgnorm)
        self.txt_enc = EncoderText(opt.vocab_size, opt.word_dim,
                                   opt.embed_size, opt.num_layers,
                                   use_abs=opt.use_abs)
        if torch.cuda.is_available():
            self.img_enc.cuda()
            self.txt_enc.cuda()
            cudnn.benchmark = True

        # Loss and Optimizer
        self.criterion = ContrastiveLoss(margin=opt.margin,
                                         measure=opt.measure,
                                         max_violation=opt.max_violation)
        params = list(self.txt_enc.parameters())
        params += list(self.img_enc.fc.parameters())
        if opt.finetune:
            params += list(self.img_enc.cnn.parameters())
        self.params = params

        self.optimizer = torch.optim.Adam(params, lr=opt.learning_rate)

        self.Eiters = 0

    def state_dict(self):
        state_dict = [self.img_enc.state_dict(), self.txt_enc.state_dict()]
        return state_dict

    def load_state_dict(self, state_dict):
        self.img_enc.load_state_dict(state_dict[0])
        self.txt_enc.load_state_dict(state_dict[1])

    def train_start(self):
        """switch to train mode
        """
        self.img_enc.train()
        self.txt_enc.train()

    def val_start(self):
        """switch to evaluate mode
        """
        self.img_enc.eval()
        self.txt_enc.eval()

    def forward_emb(self, images, captions, lengths, volatile=False):
        """Compute the image and caption embeddings
        """
        # Set mini-batch dataset
        images = Variable(images, volatile=volatile)
        captions = Variable(captions, volatile=volatile)
        if torch.cuda.is_available():
            images = images.cuda()
            captions = captions.cuda()

        # Forward
        img_emb = self.img_enc(images)
        cap_emb = self.txt_enc(captions, lengths)
        return img_emb, cap_emb

    def forward_loss(self, img_emb, cap_emb, **kwargs):
        """Compute the loss given pairs of image and caption embeddings
        """
        loss = self.criterion(img_emb, cap_emb)
        self.logger.update('Le', loss.data[0], img_emb.size(0))
        return loss

    def train_emb(self, images, captions, lengths, ids=None, *args):
        """One training step given images and captions.
        """
        self.Eiters += 1
        self.logger.update('Eit', self.Eiters)
        self.logger.update('lr', self.optimizer.param_groups[0]['lr'])

        # compute the embeddings
        img_emb, cap_emb = self.forward_emb(images, captions, lengths)

        # measure accuracy and record loss
        self.optimizer.zero_grad()
        loss = self.forward_loss(img_emb, cap_emb)

        # compute gradient and do SGD step
        loss.backward()
        if self.grad_clip > 0:
            clip_grad_norm(self.params, self.grad_clip)
        self.optimizer.step()


================================================
FILE: train.py
================================================
import pickle
import os
import time
import shutil

import torch

import data
from vocab import Vocabulary  # NOQA
from model import VSE
from evaluation import i2t, t2i, AverageMeter, LogCollector, encode_data

import logging
import tensorboard_logger as tb_logger

import argparse


def main():
    # Hyper Parameters
    parser = argparse.ArgumentParser()
    parser.add_argument('--data_path', default='/w/31/faghri/vsepp_data/',
                        help='path to datasets')
    parser.add_argument('--data_name', default='precomp',
                        help='{coco,f8k,f30k,10crop}_precomp|coco|f8k|f30k')
    parser.add_argument('--vocab_path', default='./vocab/',
                        help='Path to saved vocabulary pickle files.')
    parser.add_argument('--margin', default=0.2, type=float,
                        help='Rank loss margin.')
    parser.add_argument('--num_epochs', default=30, type=int,
                        help='Number of training epochs.')
    parser.add_argument('--batch_size', default=128, type=int,
                        help='Size of a training mini-batch.')
    parser.add_argument('--word_dim', default=300, type=int,
                        help='Dimensionality of the word embedding.')
    parser.add_argument('--embed_size', default=1024, type=int,
                        help='Dimensionality of the joint embedding.')
    parser.add_argument('--grad_clip', default=2., type=float,
                        help='Gradient clipping threshold.')
    parser.add_argument('--crop_size', default=224, type=int,
                        help='Size of an image crop as the CNN input.')
    parser.add_argument('--num_layers', default=1, type=int,
                        help='Number of GRU layers.')
    parser.add_argument('--learning_rate', default=.0002, type=float,
                        help='Initial learning rate.')
    parser.add_argument('--lr_update', default=15, type=int,
                        help='Number of epochs to update the learning rate.')
    parser.add_argument('--workers', default=10, type=int,
                        help='Number of data loader workers.')
    parser.add_argument('--log_step', default=10, type=int,
                        help='Number of steps to print and record the log.')
    parser.add_argument('--val_step', default=500, type=int,
                        help='Number of steps to run validation.')
    parser.add_argument('--logger_name', default='runs/runX',
                        help='Path to save the model and Tensorboard log.')
    parser.add_argument('--resume', default='', type=str, metavar='PATH',
                        help='path to latest checkpoint (default: none)')
    parser.add_argument('--max_violation', action='store_true',
                        help='Use max instead of sum in the rank loss.')
    parser.add_argument('--img_dim', default=4096, type=int,
                        help='Dimensionality of the image embedding.')
    parser.add_argument('--finetune', action='store_true',
                        help='Fine-tune the image encoder.')
    parser.add_argument('--cnn_type', default='vgg19',
                        help="""The CNN used for image encoder
                        (e.g. vgg19, resnet152)""")
    parser.add_argument('--use_restval', action='store_true',
                        help='Use the restval data for training on MSCOCO.')
    parser.add_argument('--measure', default='cosine',
                        help='Similarity measure used (cosine|order)')
    parser.add_argument('--use_abs', action='store_true',
                        help='Take the absolute value of embedding vectors.')
    parser.add_argument('--no_imgnorm', action='store_true',
                        help='Do not normalize the image embeddings.')
    parser.add_argument('--reset_train', action='store_true',
                        help='Ensure the training is always done in '
                        'train mode (Not recommended).')
    opt = parser.parse_args()
    print(opt)

    logging.basicConfig(format='%(asctime)s %(message)s', level=logging.INFO)
    tb_logger.configure(opt.logger_name, flush_secs=5)

    # Load Vocabulary Wrapper
    vocab = pickle.load(open(os.path.join(
        opt.vocab_path, '%s_vocab.pkl' % opt.data_name), 'rb'))
    opt.vocab_size = len(vocab)

    # Load data loaders
    train_loader, val_loader = data.get_loaders(
        opt.data_name, vocab, opt.crop_size, opt.batch_size, opt.workers, opt)

    # Construct the model
    model = VSE(opt)

    # optionally resume from a checkpoint
    if opt.resume:
        if os.path.isfile(opt.resume):
            print("=> loading checkpoint '{}'".format(opt.resume))
            checkpoint = torch.load(opt.resume)
            start_epoch = checkpoint['epoch']
            best_rsum = checkpoint['best_rsum']
            model.load_state_dict(checkpoint['model'])
            # Eiters is used to show logs as the continuation of another
            # training
            model.Eiters = checkpoint['Eiters']
            print("=> loaded checkpoint '{}' (epoch {}, best_rsum {})"
                  .format(opt.resume, start_epoch, best_rsum))
            validate(opt, val_loader, model)
        else:
            print("=> no checkpoint found at '{}'".format(opt.resume))

    # Train the Model
    best_rsum = 0
    for epoch in range(opt.num_epochs):
        adjust_learning_rate(opt, model.optimizer, epoch)

        # train for one epoch
        train(opt, train_loader, model, epoch, val_loader)

        # evaluate on validation set
        rsum = validate(opt, val_loader, model)

        # remember best R@ sum and save checkpoint
        is_best = rsum > best_rsum
        best_rsum = max(rsum, best_rsum)
        save_checkpoint({
            'epoch': epoch + 1,
            'model': model.state_dict(),
            'best_rsum': best_rsum,
            'opt': opt,
            'Eiters': model.Eiters,
        }, is_best, prefix=opt.logger_name + '/')


def train(opt, train_loader, model, epoch, val_loader):
    # average meters to record the training statistics
    batch_time = AverageMeter()
    data_time = AverageMeter()
    train_logger = LogCollector()

    # switch to train mode
    model.train_start()

    end = time.time()
    for i, train_data in enumerate(train_loader):
        if opt.reset_train:
            # Always reset to train mode, this is not the default behavior
            model.train_start()

        # measure data loading time
        data_time.update(time.time() - end)

        # make sure train logger is used
        model.logger = train_logger

        # Update the model
        model.train_emb(*train_data)

        # measure elapsed time
        batch_time.update(time.time() - end)
        end = time.time()

        # Print log info
        if model.Eiters % opt.log_step == 0:
            logging.info(
                'Epoch: [{0}][{1}/{2}]\t'
                '{e_log}\t'
                'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
                'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
                .format(
                    epoch, i, len(train_loader), batch_time=batch_time,
                    data_time=data_time, e_log=str(model.logger)))

        # Record logs in tensorboard
        tb_logger.log_value('epoch', epoch, step=model.Eiters)
        tb_logger.log_value('step', i, step=model.Eiters)
        tb_logger.log_value('batch_time', batch_time.val, step=model.Eiters)
        tb_logger.log_value('data_time', data_time.val, step=model.Eiters)
        model.logger.tb_log(tb_logger, step=model.Eiters)

        # validate at every val_step
        if model.Eiters % opt.val_step == 0:
            validate(opt, val_loader, model)


def validate(opt, val_loader, model):
    # compute the encoding for all the validation images and captions
    img_embs, cap_embs = encode_data(
        model, val_loader, opt.log_step, logging.info)

    # caption retrieval
    (r1, r5, r10, medr, meanr) = i2t(img_embs, cap_embs, measure=opt.measure)
    logging.info("Image to text: %.1f, %.1f, %.1f, %.1f, %.1f" %
                 (r1, r5, r10, medr, meanr))
    # image retrieval
    (r1i, r5i, r10i, medri, meanri) = t2i(
        img_embs, cap_embs, measure=opt.measure)
    logging.info("Text to image: %.1f, %.1f, %.1f, %.1f, %.1f" %
                 (r1i, r5i, r10i, medri, meanri))
    # sum of recalls to be used for early stopping
    currscore = r1 + r5 + r10 + r1i + r5i + r10i

    # record metrics in tensorboard
    tb_logger.log_value('r1', r1, step=model.Eiters)
    tb_logger.log_value('r5', r5, step=model.Eiters)
    tb_logger.log_value('r10', r10, step=model.Eiters)
    tb_logger.log_value('medr', medr, step=model.Eiters)
    tb_logger.log_value('meanr', meanr, step=model.Eiters)
    tb_logger.log_value('r1i', r1i, step=model.Eiters)
    tb_logger.log_value('r5i', r5i, step=model.Eiters)
    tb_logger.log_value('r10i', r10i, step=model.Eiters)
    tb_logger.log_value('medri', medri, step=model.Eiters)
    tb_logger.log_value('meanri', meanri, step=model.Eiters)
    tb_logger.log_value('rsum', currscore, step=model.Eiters)

    return currscore


def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', prefix=''):
    torch.save(state, prefix + filename)
    if is_best:
        shutil.copyfile(prefix + filename, prefix + 'model_best.pth.tar')


def adjust_learning_rate(opt, optimizer, epoch):
    """Sets the learning rate to the initial LR
       decayed by 10 every 30 epochs"""
    lr = opt.learning_rate * (0.1 ** (epoch // opt.lr_update))
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr


def accuracy(output, target, topk=(1,)):
    """Computes the precision@k for the specified values of k"""
    maxk = max(topk)
    batch_size = target.size(0)

    _, pred = output.topk(maxk, 1, True, True)
    pred = pred.t()
    correct = pred.eq(target.view(1, -1).expand_as(pred))

    res = []
    for k in topk:
        correct_k = correct[:k].view(-1).float().sum(0)
        res.append(correct_k.mul_(100.0 / batch_size))
    return res


if __name__ == '__main__':
    main()


================================================
FILE: vocab.py
================================================
# Create a vocabulary wrapper
import nltk
import pickle
from collections import Counter
from pycocotools.coco import COCO
import json
import argparse
import os

annotations = {
    'coco_precomp': ['train_caps.txt', 'dev_caps.txt'],
    'coco': ['annotations/captions_train2014.json',
             'annotations/captions_val2014.json'],
    'f8k_precomp': ['train_caps.txt', 'dev_caps.txt'],
    '10crop_precomp': ['train_caps.txt', 'dev_caps.txt'],
    'f30k_precomp': ['train_caps.txt', 'dev_caps.txt'],
    'f8k': ['dataset_flickr8k.json'],
    'f30k': ['dataset_flickr30k.json'],
}


class Vocabulary(object):
    """Simple vocabulary wrapper."""

    def __init__(self):
        self.word2idx = {}
        self.idx2word = {}
        self.idx = 0

    def add_word(self, word):
        if word not in self.word2idx:
            self.word2idx[word] = self.idx
            self.idx2word[self.idx] = word
            self.idx += 1

    def __call__(self, word):
        if word not in self.word2idx:
            return self.word2idx['<unk>']
        return self.word2idx[word]

    def __len__(self):
        return len(self.word2idx)


def from_coco_json(path):
    coco = COCO(path)
    ids = coco.anns.keys()
    captions = []
    for i, idx in enumerate(ids):
        captions.append(str(coco.anns[idx]['caption']))

    return captions


def from_flickr_json(path):
    dataset = json.load(open(path, 'r'))['images']
    captions = []
    for i, d in enumerate(dataset):
        captions += [str(x['raw']) for x in d['sentences']]

    return captions


def from_txt(txt):
    captions = []
    with open(txt, 'rb') as f:
        for line in f:
            captions.append(line.strip())
    return captions


def build_vocab(data_path, data_name, jsons, threshold):
    """Build a simple vocabulary wrapper."""
    counter = Counter()
    for path in jsons[data_name]:
        full_path = os.path.join(os.path.join(data_path, data_name), path)
        if data_name == 'coco':
            captions = from_coco_json(full_path)
        elif data_name == 'f8k' or data_name == 'f30k':
            captions = from_flickr_json(full_path)
        else:
            captions = from_txt(full_path)
        for i, caption in enumerate(captions):
            tokens = nltk.tokenize.word_tokenize(
                caption.lower().decode('utf-8'))
            counter.update(tokens)

            if i % 1000 == 0:
                print("[%d/%d] tokenized the captions." % (i, len(captions)))

    # Discard if the occurrence of the word is less than min_word_cnt.
    words = [word for word, cnt in counter.items() if cnt >= threshold]

    # Create a vocab wrapper and add some special tokens.
    vocab = Vocabulary()
    vocab.add_word('<pad>')
    vocab.add_word('<start>')
    vocab.add_word('<end>')
    vocab.add_word('<unk>')

    # Add words to the vocabulary.
    for i, word in enumerate(words):
        vocab.add_word(word)
    return vocab


def main(data_path, data_name):
    vocab = build_vocab(data_path, data_name, jsons=annotations, threshold=4)
    with open('./vocab/%s_vocab.pkl' % data_name, 'wb') as f:
        pickle.dump(vocab, f, pickle.HIGHEST_PROTOCOL)
    print("Saved vocabulary file to ", './vocab/%s_vocab.pkl' % data_name)


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--data_path', default='/w/31/faghri/vsepp_data/')
    parser.add_argument('--data_name', default='coco',
                        help='{coco,f8k,f30k,10crop}_precomp|coco|f8k|f30k')
    opt = parser.parse_args()
    main(opt.data_path, opt.data_name)