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
from collections.abc import Sequence
import shelve
from transformers import BertTokenizer

from features import HuggingFaceTransformerExtractor


def get_paths(config):
    """
    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.
    """
    name = config['dataset']['name']
    annotations_path = os.path.join(config['dataset']['data'], name, 'annotations')
    use_restval = config['dataset']['restval']

    roots = {}
    ids = {}
    if 'coco' == name:
        imgdir = config['dataset']['images-path']
        capdir = annotations_path #os.path.join(dataset_root, 'annotations_trainval2014')
        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(annotations_path, 'coco_train_ids.npy'))
        ids['val'] = np.load(os.path.join(annotations_path, 'coco_dev_ids.npy'))[:5000]
        ids['test'] = np.load(os.path.join(annotations_path, 'coco_test_ids.npy'))
        ids['trainrestval'] = (
            ids['train'],
            np.load(os.path.join(annotations_path, 'coco_restval_ids.npy')))
        if use_restval:
            roots['train'] = roots['trainrestval']
            ids['train'] = ids['trainrestval']

    return roots, ids


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

    def __init__(self, root, json, transform=None, ids=None, get_images=True):
        """
        Args:
            root: image directory.
            json: coco annotation file path.
            vocab: vocabulary wrapper.
            transform: transformer for image.
        """
        self.root = root
        self.get_images = get_images
        # 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.transform = transform

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

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

        target = caption
        return image, target, index, img_id

    def get_raw_item(self, index, load_image=True):
        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']
        img = coco.imgs[img_id]
        img_size = np.array([img['width'], img['height']])
        if load_image:
            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, img_size
        else:
            return root, caption, img_id, None, None, img_size

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


class CocoBottomUpFeaturesDataset(CocoDataset):
    def __init__(self, root, json, features_path, split, ids=None, **kwargs):
        super().__init__(root, json, ids=ids)
        # data_path = config['image-model']['data-path']
        self.feats_data_path = os.path.join(features_path, 'cocobu_att')
        self.box_data_path = os.path.join(features_path, 'cocobu_box')
        config = kwargs['config']
        self.load_preextracted = config['text-model']['pre-extracted']
        if self.load_preextracted:
            # TODO: handle different types of preextracted features, not only BERT
            text_extractor = HuggingFaceTransformerExtractor(config, split, finetuned=config['text-model']['fine-tune'])
            self.text_features_db = FeatureSequence(text_extractor)

    def __getitem__(self, index):
        """This function returns a tuple that is further passed to collate_fn
        """
        root, caption, img_id, _, _, img_size = self.get_raw_item(index, load_image=False)
        img_feat_path = os.path.join(self.feats_data_path, '{}.npz'.format(img_id))
        img_box_path = os.path.join(self.box_data_path, '{}.npy'.format(img_id))

        img_feat = np.load(img_feat_path)['feat']
        img_boxes = np.load(img_box_path)

        # normalize boxes
        img_boxes = img_boxes / np.tile(img_size, 2)

        img_feat = torch.Tensor(img_feat)
        img_boxes = torch.Tensor(img_boxes)

        if self.load_preextracted:
            record = self.text_features_db[index]
            features = record['features']
            captions = record['captions']
            wembeddings = record['wembeddings']
            target = (captions, features, wembeddings)
        else:
            target = caption
        # image = (img_feat, img_boxes)
        return img_feat, img_boxes, target, index, img_id


class FeatureSequence(Sequence):
    # Used by VlenPrecompDataset
    def __init__(self, feature_extractor):
        self.feature_extractor = feature_extractor
        # first of all, extract the features
        print('Initializing Feature Sequence for {}'.format(self.feature_extractor))
        db_file = self.feature_extractor.get_db_file()
        if not os.path.isfile('{}.dat'.format(db_file)):
            raise FileNotFoundError("Features file {} not found. Have you run feature extraction for {}?"
                                    .format(db_file, self.feature_extractor))
            # print('Dumping features on file: {}'.format(db_file))
            # self.feature_extractor.extract()
        else:
            print('Loading features file: {}'.format(db_file))

        self.db = shelve.open(db_file, flag='r')

    def __getitem__(self, item):
        return self.db[str(item)]

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


class Collate:
    def __init__(self, config):
        self.vocab_type = config['text-model']['name']
        if self.vocab_type == 'bert':
            self.tokenizer = BertTokenizer.from_pretrained(config['text-model']['pretrain'])

    def __call__(self, 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) or (? > 3, 2048)
                    - 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)
        if len(data[0]) == 5:      # TODO: find a better way to distinguish the two
            images, boxes, captions, ids, img_ids = zip(*data)
        elif len(data[0]) == 4:
            images, captions, ids, img_ids = zip(*data)

        preextracted_captions = type(captions[0]) is tuple
        if preextracted_captions:
            # they are pre-extracted features
            captions, cap_features, wembeddings = zip(*captions)
            cap_lengths = [len(cap) for cap in cap_features]
            captions = [torch.LongTensor(c) for c in captions]
            cap_features = [torch.FloatTensor(f) for f in cap_features]
            wembeddings = [torch.FloatTensor(w) for w in wembeddings]
        else:
            # captions is a list of strings to be processed
            if self.vocab_type == 'bert':
                cap_lengths = [len(self.tokenizer.tokenize(c)) + 2 for c in
                           captions]  # + 2 in order to account for begin and end tokens
                max_len = max(cap_lengths)
                captions_ids = [torch.LongTensor(self.tokenizer.encode(c, max_length=max_len, pad_to_max_length=True))
                                for c in captions]

            captions = captions_ids
        # Merge images (convert tuple of 3D tensor to 4D tensor)
        preextracted_images = not (images[0].shape[0] == 3)
        if not preextracted_images:
            # they are images
            images = torch.stack(images, 0)
        else:
            # they are image features, variable length
            feat_lengths = [f.shape[0] + 1 for f in images] # +1 because the first region feature is reserved as CLS
            feat_dim = images[0].shape[1]
            img_features = torch.zeros(len(images), max(feat_lengths), feat_dim)
            for i, img in enumerate(images):
                end = feat_lengths[i]
                img_features[i, 1:end] = img

            box_lengths = [b.shape[0] + 1 for b in boxes]  # +1 because the first region feature is reserved as CLS
            assert box_lengths == feat_lengths
            out_boxes = torch.zeros(len(boxes), max(box_lengths), 4)
            for i, box in enumerate(boxes):
                end = box_lengths[i]
                out_boxes[i, 1:end] = box

        # Merget captions (convert tuple of 1D tensor to 2D tensor)
        if preextracted_captions:
            captions_t = torch.zeros(len(captions), max(cap_lengths)).long()
            features_t = torch.zeros(len(cap_features), max(cap_lengths), cap_features[0].shape[1])
            wembeddings_t = torch.zeros(len(wembeddings), max(cap_lengths), wembeddings[0].shape[1])
            for i, (cap, feats, wembs, l) in enumerate(zip(captions, cap_features, wembeddings, cap_lengths)):
                captions_t[i, :l] = cap[:l]
                features_t[i, :l] = feats[:l]
                wembeddings_t[i, :l] = wembs[:l]
            targets = (captions_t, features_t, wembeddings_t)
        else:
            targets = torch.zeros(len(captions), max(cap_lengths)).long()
            for i, cap in enumerate(captions):
                end = cap_lengths[i]
                targets[i, :end] = cap[:end]

        if not preextracted_images:
            return images, targets, None, cap_lengths, None, ids
        else:
            # features = features.permute(0, 2, 1)
            return img_features, targets, feat_lengths, cap_lengths, out_boxes, ids


def get_loader_single(data_name, split, root, json, transform, preextracted_root=None,
                      batch_size=100, shuffle=True,
                      num_workers=2, ids=None, collate_fn=None, **kwargs):
    """Returns torch.utils.data.DataLoader for custom coco dataset."""
    if 'coco' in data_name:
        if preextracted_root is not None:
            dataset = CocoBottomUpFeaturesDataset(root=root,
                                                  json=json,
                                                  features_path=preextracted_root, split=split,
                                                  ids=ids, **kwargs)
        else:
            # COCO custom dataset
            dataset = CocoDataset(root=root,
                                  json=json,
                                  transform=transform, ids=ids)

    # 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_transform(data_name, split_name, config):
    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(config['image-model']['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(config, workers, batch_size=None):
    data_name = config['dataset']['name']
    if batch_size is None:
        batch_size = config['training']['bs']
    collate_fn = Collate(config)

    # Build Dataset Loader
    roots, ids = get_paths(config)

    transform = get_transform(data_name, 'train', config)
    preextracted_root = config['image-model']['pre-extracted-features-root'] \
        if 'pre-extracted-features-root' in config['image-model'] else None

    train_loader = get_loader_single(data_name, 'train',
                                     roots['train']['img'],
                                     roots['train']['cap'],
                                     transform, ids=ids['train'],
                                     preextracted_root=preextracted_root,
                                     batch_size=batch_size, shuffle=True,
                                     num_workers=workers,
                                     collate_fn=collate_fn, config=config)

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

    return train_loader, val_loader


def get_test_loader(config, vocab, workers, split_name='test', batch_size=None):
    data_name = config['dataset']['name']
    if batch_size is None:
        batch_size = config['training']['bs']
    collate_fn = Collate(config)

    # Build Dataset Loader
    roots, ids = get_paths(config)

    preextracted_root = config['image-model']['pre-extracted-features-root'] \
        if 'pre-extracted-features-root' in config['image-model'] else None

    transform = get_transform(data_name, split_name, config)
    test_loader = get_loader_single(data_name, split_name,
                                    roots[split_name]['img'],
                                    roots[split_name]['cap'],
                                    transform, ids=ids[split_name],
                                    preextracted_root=preextracted_root,
                                    batch_size=batch_size, shuffle=False,
                                    num_workers=workers,
                                    collate_fn=collate_fn, config=config)
    return test_loader