def __init__(
        self,
        task: str,
        dataroot: str,
        annotations_jsonpath: str,
        split: str,
        image_features_reader: ImageFeaturesH5Reader,
        gt_image_features_reader: ImageFeaturesH5Reader,
        tokenizer: BertTokenizer,
        padding_index: int = 0,
        max_seq_length: int = 20,
        max_region_num: int = 60,
        limit_nums=None,
    ):
        self.split = split
        self.refer = REFER(dataroot, dataset=task, splitBy='unc')
        self.ref_ids = self.refer.getRefIds(split=split)
        print('%s refs are in split [%s].' % (len(self.ref_ids), split))

        self.num_labels = 1
        self._image_features_reader = image_features_reader
        self._gt_image_features_reader = gt_image_features_reader
        self._tokenizer = tokenizer

        self._padding_index = padding_index
        self._max_seq_length = max_seq_length
        self.entries = self._load_annotations()

        self.max_region_num = max_region_num

        if not os.path.exists(os.path.join(dataroot, 'cache')):
            os.makedirs(os.path.join(dataroot, 'cache'))

        cache_path = os.path.join(
            dataroot, 'cache', task + '_' + split + '_' + str(max_seq_length) +
            '_' + str(max_region_num) + '_tolist.pkl')
        if not os.path.exists(cache_path):
            self.tokenize()
            self.tensorize()
            cPickle.dump(self.entries, open(cache_path, 'wb'))
        else:
            print('loading entries from %s' % (cache_path))
            self.entries = cPickle.load(open(cache_path, 'rb'))
Пример #2
0
    def __init__(
        self,
        task: str,
        dataroot: str,
        annotations_jsonpath: str,
        split: str,
        image_features_reader: ImageFeaturesH5Reader,
        gt_image_features_reader: ImageFeaturesH5Reader,
        tokenizer: BertTokenizer,
        bert_model,
        clean_datasets,
        padding_index: int = 0,
        max_seq_length: int = 20,
        max_region_num: int = 60,
        target_file: str = '',
    ):
        self.split = split

        if task == "refcocog":
            self.refer = REFER(dataroot, dataset=task, splitBy="umd")
        else:
            self.refer = REFER(dataroot, dataset=task, splitBy="unc")

        if self.split == "mteval":
            self.ref_ids = self.refer.getRefIds(split="train")
        else:
            self.ref_ids = self.refer.getRefIds(split=split)

        print("%s refs are in split [%s]." % (len(self.ref_ids), split))

        self.num_labels = 1
        self._image_features_reader = image_features_reader
        self._gt_image_features_reader = gt_image_features_reader
        self._tokenizer = tokenizer

        self._padding_index = padding_index
        self._max_seq_length = max_seq_length
        self.dataroot = dataroot
        self.entries = self._load_annotations(clean_datasets)

        self.max_region_num = max_region_num

        self.target_dim = int(target_file.split(".")[0][-2:])
        with open(os.path.join(self.dataroot, task, target_file), 'r') as f:
            self.target_data = json.load(f)

        clean_train = "_cleaned" if clean_datasets else ""

        if "roberta" in bert_model:
            cache_path = os.path.join(
                dataroot,
                "cache",
                task + "_" + split + "_" + "roberta" + "_" +
                str(max_seq_length) + "_" + str(max_region_num) + clean_train +
                ".pkl",
            )
        else:
            cache_path = os.path.join(
                dataroot,
                "cache",
                task + "_" + split + "_" + str(max_seq_length) + "_" +
                str(max_region_num) + clean_train + ".pkl",
            )

        if not os.path.exists(cache_path):
            self.tokenize()
            self.tensorize()
            cPickle.dump(self.entries, open(cache_path, "wb"))
        else:
            print("loading entries from %s" % (cache_path))
            self.entries = cPickle.load(open(cache_path, "rb"))
Пример #3
0
class ReferExpressionDataset(Dataset):
    def __init__(
        self,
        task: str,
        dataroot: str,
        annotations_jsonpath: str,
        split: str,
        image_features_reader: ImageFeaturesH5Reader,
        gt_image_features_reader: ImageFeaturesH5Reader,
        tokenizer: BertTokenizer,
        bert_model,
        clean_datasets,
        padding_index: int = 0,
        max_seq_length: int = 20,
        max_region_num: int = 60,
        target_file: str = '',
    ):
        self.split = split

        if task == "refcocog":
            self.refer = REFER(dataroot, dataset=task, splitBy="umd")
        else:
            self.refer = REFER(dataroot, dataset=task, splitBy="unc")

        if self.split == "mteval":
            self.ref_ids = self.refer.getRefIds(split="train")
        else:
            self.ref_ids = self.refer.getRefIds(split=split)

        print("%s refs are in split [%s]." % (len(self.ref_ids), split))

        self.num_labels = 1
        self._image_features_reader = image_features_reader
        self._gt_image_features_reader = gt_image_features_reader
        self._tokenizer = tokenizer

        self._padding_index = padding_index
        self._max_seq_length = max_seq_length
        self.dataroot = dataroot
        self.entries = self._load_annotations(clean_datasets)

        self.max_region_num = max_region_num

        self.target_dim = int(target_file.split(".")[0][-2:])
        with open(os.path.join(self.dataroot, task, target_file), 'r') as f:
            self.target_data = json.load(f)

        clean_train = "_cleaned" if clean_datasets else ""

        if "roberta" in bert_model:
            cache_path = os.path.join(
                dataroot,
                "cache",
                task + "_" + split + "_" + "roberta" + "_" +
                str(max_seq_length) + "_" + str(max_region_num) + clean_train +
                ".pkl",
            )
        else:
            cache_path = os.path.join(
                dataroot,
                "cache",
                task + "_" + split + "_" + str(max_seq_length) + "_" +
                str(max_region_num) + clean_train + ".pkl",
            )

        if not os.path.exists(cache_path):
            self.tokenize()
            self.tensorize()
            cPickle.dump(self.entries, open(cache_path, "wb"))
        else:
            print("loading entries from %s" % (cache_path))
            self.entries = cPickle.load(open(cache_path, "rb"))

    def _load_annotations(self, clean_datasets):

        # annotations_json: Dict[str, Any] = json.load(open(annotations_jsonpath))

        # Build an index which maps image id with a list of caption annotations.
        entries = []
        remove_ids = []
        if clean_datasets or self.split == "mteval":
            remove_ids = np.load(
                os.path.join(self.dataroot, "cache", "coco_test_ids.npy"))
            remove_ids = [int(x) for x in remove_ids]

        for ref_id in self.ref_ids:
            ref = self.refer.Refs[ref_id]
            image_id = ref["image_id"]
            if self.split == "train" and int(image_id) in remove_ids:
                continue
            elif self.split == "mteval" and int(image_id) not in remove_ids:
                continue
            ref_id = ref["ref_id"]
            refBox = self.refer.getRefBox(ref_id)
            for sent, sent_id in zip(ref["sentences"], ref["sent_ids"]):
                caption = sent["raw"]
                entries.append({
                    "caption": caption,
                    "sent_id": sent_id,
                    "image_id": image_id,
                    "refBox": refBox,
                    "ref_id": ref_id,
                })

        return entries

    def tokenize(self):
        """Tokenizes the captions.

        This will add caption_tokens in each entry of the dataset.
        -1 represents nil, and should be treated as padding_idx in embedding.
        """
        for entry in self.entries:

            # sentence_tokens = self._tokenizer.tokenize(entry["caption"])
            # sentence_tokens = ["[CLS]"] + sentence_tokens + ["[SEP]"]

            # tokens = [
            #     self._tokenizer.vocab.get(w, self._tokenizer.vocab["[UNK]"])
            #     for w in sentence_tokens
            # ]

            tokens = self._tokenizer.encode(entry["caption"])
            tokens = tokens[:self._max_seq_length - 2]
            tokens = self._tokenizer.add_special_tokens_single_sentence(tokens)

            segment_ids = [0] * len(tokens)
            input_mask = [1] * len(tokens)

            if len(tokens) < self._max_seq_length:
                # Note here we pad in front of the sentence
                padding = [self._padding_index
                           ] * (self._max_seq_length - len(tokens))
                tokens = tokens + padding
                input_mask += padding
                segment_ids += padding

            assert_eq(len(tokens), self._max_seq_length)
            entry["token"] = tokens
            entry["input_mask"] = input_mask
            entry["segment_ids"] = segment_ids

    def tensorize(self):

        for entry in self.entries:
            token = torch.from_numpy(np.array(entry["token"]))
            entry["token"] = token

            input_mask = torch.from_numpy(np.array(entry["input_mask"]))
            entry["input_mask"] = input_mask

            segment_ids = torch.from_numpy(np.array(entry["segment_ids"]))
            entry["segment_ids"] = segment_ids

    def __getitem__(self, index):
        entry = self.entries[index]

        image_id = entry["image_id"]
        caption_id = entry["sent_id"]
        ref_box = entry["refBox"]

        ref_box = [
            ref_box[0],
            ref_box[1],
            ref_box[0] + ref_box[2],
            ref_box[1] + ref_box[3],
        ]
        features, num_boxes, boxes, boxes_ori = self._image_features_reader[
            image_id]

        boxes_ori = boxes_ori[:num_boxes]
        boxes = boxes[:num_boxes]
        features = features[:num_boxes]

        if self.split == "train":
            gt_features, gt_num_boxes, gt_boxes, gt_boxes_ori = self._gt_image_features_reader[
                image_id]

            # merge two boxes, and assign the labels.
            gt_boxes_ori = gt_boxes_ori[1:gt_num_boxes]
            gt_boxes = gt_boxes[1:gt_num_boxes]
            gt_features = gt_features[1:gt_num_boxes]

            # concatenate the boxes
            mix_boxes_ori = np.concatenate((boxes_ori, gt_boxes_ori), axis=0)
            mix_boxes = np.concatenate((boxes, gt_boxes), axis=0)
            mix_features = np.concatenate((features, gt_features), axis=0)
            mix_num_boxes = min(int(num_boxes + int(gt_num_boxes) - 1),
                                self.max_region_num)
            # given the mix boxes, and ref_box, calculate the overlap.
            mix_target = iou(
                torch.tensor(mix_boxes_ori[:, :4]).float(),
                torch.tensor([ref_box]).float(),
            )
            mix_target[mix_target < 0.5] = 0
        else:
            mix_boxes_ori = boxes_ori
            mix_boxes = boxes
            mix_features = features
            mix_num_boxes = min(int(num_boxes), self.max_region_num)
            mix_target = iou(
                torch.tensor(mix_boxes_ori[:, :4]).float(),
                torch.tensor([ref_box]).float(),
            )

        image_mask = [1] * (mix_num_boxes)
        while len(image_mask) < self.max_region_num:
            image_mask.append(0)

        mix_boxes_pad = np.zeros((self.max_region_num, 5))
        mix_features_pad = np.zeros((self.max_region_num, 2048))

        mix_boxes_pad[:mix_num_boxes] = mix_boxes[:mix_num_boxes]
        mix_features_pad[:mix_num_boxes] = mix_features[:mix_num_boxes]

        # appending the target feature.
        features = torch.tensor(mix_features_pad).float()
        image_mask = torch.tensor(image_mask).long()
        spatials = torch.tensor(mix_boxes_pad).float()

        target = torch.zeros((self.max_region_num, 1)).float()
        target[:mix_num_boxes] = mix_target[:mix_num_boxes]

        spatials_ori = torch.tensor(mix_boxes_ori).float()
        co_attention_mask = torch.zeros(
            (self.max_region_num, self._max_seq_length))

        caption = entry["token"]
        input_mask = entry["input_mask"]
        segment_ids = entry["segment_ids"]

        indices, values = self.target_data.get(str(image_id),
                                               {}).get(str(caption_id))
        grid_vec = torch.zeros(self.target_dim * self.target_dim)
        for idx in range(len(indices)):
            index = indices[idx]
            grid_vec[index] = values[idx]

        grid_vec = torch.FloatTensor(grid_vec)

        return (
            features,
            spatials,
            image_mask,
            caption,
            target,
            input_mask,
            segment_ids,
            co_attention_mask,
            image_id,
            grid_vec,
        )

    def __len__(self):
        return len(self.entries)
    def __init__(
        self,
        task: str,
        dataroot: str,
        annotations_jsonpath: str,
        split: str,
        image_features_reader: ImageFeaturesH5Reader,
        gt_image_features_reader: ImageFeaturesH5Reader,
        tokenizer: BertTokenizer,
        bert_model,
        clean_datasets,
        padding_index: int = 0,
        max_seq_length: int = 20,
        max_region_num: int = 60,
    ):
        self.split = split

        if task == 'refcocog':
            self.refer = REFER(dataroot, dataset=task, splitBy='umd')
        else:
            self.refer = REFER(dataroot, dataset=task, splitBy='unc')

        if self.split == 'mteval':
            self.ref_ids = self.refer.getRefIds(split='train')
        else:
            self.ref_ids = self.refer.getRefIds(split=split)

        print('%s refs are in split [%s].' % (len(self.ref_ids), split))

        self.num_labels = 1
        self._image_features_reader = image_features_reader
        self._gt_image_features_reader = gt_image_features_reader
        self._tokenizer = tokenizer

        self._padding_index = padding_index
        self._max_seq_length = max_seq_length
        self.dataroot = dataroot
        self.entries = self._load_annotations(clean_datasets)

        self.max_region_num = max_region_num

        clean_train = '_cleaned_tolist' if clean_datasets else ''

        if 'roberta' in bert_model:
            cache_path = os.path.join(
                dataroot,
                'cache',
                task + '_' + split + '_' + 'roberta' + '_' +
                str(max_seq_length) + '_' + str(max_region_num) + clean_train +
                '.pkl',
            )
        else:
            cache_path = os.path.join(
                dataroot,
                'cache',
                task + '_' + split + '_' + str(max_seq_length) + '_' +
                str(max_region_num) + clean_train + '.pkl',
            )

        if not os.path.exists(cache_path):
            self.tokenize()
            self.tensorize()
            cPickle.dump(self.entries, open(cache_path, 'wb'))
        else:
            print('loading entries from %s' % (cache_path))
            self.entries = cPickle.load(open(cache_path, 'rb'))
class ReferExpressionDataset(Dataset):
    def __init__(
        self,
        task: str,
        dataroot: str,
        annotations_jsonpath: str,
        split: str,
        image_features_reader: ImageFeaturesH5Reader,
        gt_image_features_reader: ImageFeaturesH5Reader,
        tokenizer: BertTokenizer,
        bert_model,
        clean_datasets,
        padding_index: int = 0,
        max_seq_length: int = 20,
        max_region_num: int = 60,
    ):
        self.split = split

        if task == 'refcocog':
            self.refer = REFER(dataroot, dataset=task, splitBy='umd')
        else:
            self.refer = REFER(dataroot, dataset=task, splitBy='unc')

        if self.split == 'mteval':
            self.ref_ids = self.refer.getRefIds(split='train')
        else:
            self.ref_ids = self.refer.getRefIds(split=split)

        print('%s refs are in split [%s].' % (len(self.ref_ids), split))

        self.num_labels = 1
        self._image_features_reader = image_features_reader
        self._gt_image_features_reader = gt_image_features_reader
        self._tokenizer = tokenizer

        self._padding_index = padding_index
        self._max_seq_length = max_seq_length
        self.dataroot = dataroot
        self.entries = self._load_annotations(clean_datasets)

        self.max_region_num = max_region_num

        clean_train = '_cleaned_tolist' if clean_datasets else ''

        if 'roberta' in bert_model:
            cache_path = os.path.join(
                dataroot,
                'cache',
                task + '_' + split + '_' + 'roberta' + '_' +
                str(max_seq_length) + '_' + str(max_region_num) + clean_train +
                '.pkl',
            )
        else:
            cache_path = os.path.join(
                dataroot,
                'cache',
                task + '_' + split + '_' + str(max_seq_length) + '_' +
                str(max_region_num) + clean_train + '.pkl',
            )

        if not os.path.exists(cache_path):
            self.tokenize()
            self.tensorize()
            cPickle.dump(self.entries, open(cache_path, 'wb'))
        else:
            print('loading entries from %s' % (cache_path))
            self.entries = cPickle.load(open(cache_path, 'rb'))

    def _load_annotations(self, clean_datasets):

        # annotations_json: Dict[str, Any] = json.load(open(annotations_jsonpath))

        # Build an index which maps image id with a list of caption annotations.
        entries = []
        remove_ids = []
        if clean_datasets or self.split == 'mteval':
            remove_ids = np.load(
                os.path.join(self.dataroot, 'cache', 'coco_test_ids.npy'))
            remove_ids = [int(x) for x in remove_ids]

        for ref_id in self.ref_ids:
            ref = self.refer.Refs[ref_id]
            image_id = ref['image_id']
            if self.split == 'train' and int(image_id) in remove_ids:
                continue
            elif self.split == 'mteval' and int(image_id) not in remove_ids:
                continue
            ref_id = ref['ref_id']
            refBox = self.refer.getRefBox(ref_id)
            for sent, sent_id in zip(ref['sentences'], ref['sent_ids']):
                caption = sent['raw']
                entries.append({
                    'caption': caption,
                    'sent_id': sent_id,
                    'image_id': image_id,
                    'refBox': refBox,
                    'ref_id': ref_id,
                })

        return entries

    def tokenize(self):
        """Tokenizes the captions.

        This will add caption_tokens in each entry of the dataset.
        -1 represents nil, and should be treated as padding_idx in embedding.
        """
        for entry in self.entries:

            # sentence_tokens = self._tokenizer.tokenize(entry["caption"])
            # sentence_tokens = ["[CLS]"] + sentence_tokens + ["[SEP]"]

            # tokens = [
            #     self._tokenizer.vocab.get(w, self._tokenizer.vocab["[UNK]"])
            #     for w in sentence_tokens
            # ]

            tokens = self._tokenizer.encode(entry['caption'])
            tokens = tokens[:self._max_seq_length - 2]
            tokens = self._tokenizer.add_special_tokens_single_sentence(tokens)

            segment_ids = [0] * len(tokens)
            input_mask = [1] * len(tokens)

            if len(tokens) < self._max_seq_length:
                # Note here we pad in front of the sentence
                padding = [self._padding_index
                           ] * (self._max_seq_length - len(tokens))
                tokens = tokens + padding
                input_mask += padding
                segment_ids += padding

            assert_eq(len(tokens), self._max_seq_length)
            entry['token'] = tokens
            entry['input_mask'] = input_mask
            entry['segment_ids'] = segment_ids

    def tensorize(self):

        for entry in self.entries:
            token = torch.from_numpy(np.array(entry['token']))
            entry['token'] = token

            input_mask = torch.from_numpy(np.array(entry['input_mask']))
            entry['input_mask'] = input_mask

            segment_ids = torch.from_numpy(np.array(entry['segment_ids']))
            entry['segment_ids'] = segment_ids

    def __getitem__(self, index):
        entry = self.entries[index]

        image_id = entry['image_id']
        ref_box = entry['refBox']

        ref_box = [
            ref_box[0],
            ref_box[1],
            ref_box[0] + ref_box[2],
            ref_box[1] + ref_box[3],
        ]
        features, num_boxes, boxes, boxes_ori = self._image_features_reader[
            image_id]

        boxes_ori = boxes_ori[:num_boxes]
        boxes = boxes[:num_boxes]
        features = features[:num_boxes]

        if self.split == 'train':
            gt_features, gt_num_boxes, gt_boxes, gt_boxes_ori = self._gt_image_features_reader[
                image_id]

            # merge two boxes, and assign the labels.
            gt_boxes_ori = gt_boxes_ori[1:gt_num_boxes]
            gt_boxes = gt_boxes[1:gt_num_boxes]
            gt_features = gt_features[1:gt_num_boxes]

            # concatenate the boxes
            mix_boxes_ori = np.concatenate((boxes_ori, gt_boxes_ori), axis=0)
            mix_boxes = np.concatenate((boxes, gt_boxes), axis=0)
            mix_features = np.concatenate((features, gt_features), axis=0)
            mix_num_boxes = min(int(num_boxes + int(gt_num_boxes) - 1),
                                self.max_region_num)
            # given the mix boxes, and ref_box, calculate the overlap.
            mix_target = iou(
                torch.tensor(mix_boxes_ori[:, :4]).float(),
                torch.tensor([ref_box]).float(),
            )
            mix_target[mix_target < 0.5] = 0
        else:
            mix_boxes_ori = boxes_ori
            mix_boxes = boxes
            mix_features = features
            mix_num_boxes = min(int(num_boxes), self.max_region_num)
            mix_target = iou(
                torch.tensor(mix_boxes_ori[:, :4]).float(),
                torch.tensor([ref_box]).float(),
            )

        image_mask = [1] * (mix_num_boxes)
        while len(image_mask) < self.max_region_num:
            image_mask.append(0)

        mix_boxes_pad = np.zeros((self.max_region_num, 5))
        mix_features_pad = np.zeros((self.max_region_num, 2048))

        mix_boxes_pad[:mix_num_boxes] = mix_boxes[:mix_num_boxes]
        mix_features_pad[:mix_num_boxes] = mix_features[:mix_num_boxes]

        # appending the target feature.
        features = torch.tensor(mix_features_pad).float()
        image_mask = torch.tensor(image_mask).long()
        spatials = torch.tensor(mix_boxes_pad).float()

        target = torch.zeros((self.max_region_num, 1)).float()
        target[:mix_num_boxes] = mix_target[:mix_num_boxes]

        # spatials_ori = torch.tensor(mix_boxes_ori).float()
        co_attention_mask = torch.zeros(
            (self.max_region_num, self._max_seq_length))

        caption = entry['token']
        input_mask = entry['input_mask']
        segment_ids = entry['segment_ids']

        device = co_attention_mask.device
        caption = torch.tensor(caption, device=device)
        input_mask = torch.tensor(input_mask, device=device)
        segment_ids = torch.tensor(segment_ids, device=device)

        return (
            features,
            spatials,
            image_mask,
            caption,
            target,
            input_mask,
            segment_ids,
            co_attention_mask,
            image_id,
        )

    def __len__(self):
        return len(self.entries)
Пример #6
0
    def __init__(
        self,
        task: str,
        dataroot: str,
        annotations_jsonpath: str,
        split: str,
        image_features_reader: ImageFeaturesH5Reader,
        gt_image_features_reader: ImageFeaturesH5Reader,
        tokenizer: AutoTokenizer,
        bert_model,
        padding_index: int = 0,
        max_seq_length: int = 20,
        max_region_num: int = 60,
        num_locs=5,
        add_global_imgfeat=None,
        append_mask_sep=False,
    ):
        self.split = split

        if task == "refcocog":
            self.refer = REFER(dataroot, dataset=task, splitBy="umd")
        else:
            self.refer = REFER(dataroot, dataset=task, splitBy="unc")

        if self.split == "mteval":
            self.ref_ids = self.refer.getRefIds(split="train")
        else:
            self.ref_ids = self.refer.getRefIds(split=split)

        print("%s refs are in split [%s]." % (len(self.ref_ids), split))

        self.num_labels = 1
        self._image_features_reader = image_features_reader
        self._gt_image_features_reader = gt_image_features_reader
        self._tokenizer = tokenizer

        self._padding_index = padding_index
        self._max_seq_length = max_seq_length
        self.dataroot = dataroot
        self.entries = self._load_annotations()

        self._max_region_num = max_region_num + int(
            add_global_imgfeat is not None)
        self._num_locs = num_locs
        self._add_global_imgfeat = add_global_imgfeat

        if "roberta" in bert_model:
            cache_path = os.path.join(
                dataroot,
                "cache",
                task + "_" + split + "_" + "roberta" + "_" +
                str(max_seq_length) + "_" + str(max_region_num) + ".pkl",
            )
        else:
            cache_path = os.path.join(
                dataroot,
                "cache",
                task + "_" + split + "_" + str(max_seq_length) + "_" +
                str(max_region_num) + ".pkl",
            )

        if not os.path.exists(cache_path):
            self.tokenize()
            self.tensorize()
            cPickle.dump(self.entries, open(cache_path, "wb"))
        else:
            print("loading entries from %s" % (cache_path))
            self.entries = cPickle.load(open(cache_path, "rb"))
Пример #7
0
class ReferExpressionDataset(Dataset):
    def __init__(
        self,
        task: str,
        dataroot: str,
        annotations_jsonpath: str,
        split: str,
        image_features_reader: ImageFeaturesH5Reader,
        gt_image_features_reader: ImageFeaturesH5Reader,
        tokenizer: AutoTokenizer,
        bert_model,
        padding_index: int = 0,
        max_seq_length: int = 20,
        max_region_num: int = 60,
        num_locs=5,
        add_global_imgfeat=None,
        append_mask_sep=False,
    ):
        self.split = split

        if task == "refcocog":
            self.refer = REFER(dataroot, dataset=task, splitBy="umd")
        else:
            self.refer = REFER(dataroot, dataset=task, splitBy="unc")

        if self.split == "mteval":
            self.ref_ids = self.refer.getRefIds(split="train")
        else:
            self.ref_ids = self.refer.getRefIds(split=split)

        print("%s refs are in split [%s]." % (len(self.ref_ids), split))

        self.num_labels = 1
        self._image_features_reader = image_features_reader
        self._gt_image_features_reader = gt_image_features_reader
        self._tokenizer = tokenizer

        self._padding_index = padding_index
        self._max_seq_length = max_seq_length
        self.dataroot = dataroot
        self.entries = self._load_annotations()

        self._max_region_num = max_region_num + int(
            add_global_imgfeat is not None)
        self._num_locs = num_locs
        self._add_global_imgfeat = add_global_imgfeat

        if "roberta" in bert_model:
            cache_path = os.path.join(
                dataroot,
                "cache",
                task + "_" + split + "_" + "roberta" + "_" +
                str(max_seq_length) + "_" + str(max_region_num) + ".pkl",
            )
        else:
            cache_path = os.path.join(
                dataroot,
                "cache",
                task + "_" + split + "_" + str(max_seq_length) + "_" +
                str(max_region_num) + ".pkl",
            )

        if not os.path.exists(cache_path):
            self.tokenize()
            self.tensorize()
            cPickle.dump(self.entries, open(cache_path, "wb"))
        else:
            print("loading entries from %s" % (cache_path))
            self.entries = cPickle.load(open(cache_path, "rb"))

    def _load_annotations(self):
        # Build an index which maps image id with a list of caption annotations.
        entries = []
        remove_ids = []
        if self.split == "mteval":
            remove_ids = np.load(
                os.path.join(self.dataroot, "cache", "coco_test_ids.npy"))
            remove_ids = [int(x) for x in remove_ids]

        for ref_id in self.ref_ids:
            ref = self.refer.Refs[ref_id]
            image_id = ref["image_id"]
            if self.split == "train" and int(image_id) in remove_ids:
                continue
            elif self.split == "mteval" and int(image_id) not in remove_ids:
                continue
            ref_id = ref["ref_id"]
            refBox = self.refer.getRefBox(ref_id)
            for sent, sent_id in zip(ref["sentences"], ref["sent_ids"]):
                caption = sent["raw"]
                entries.append({
                    "caption": caption,
                    "sent_id": sent_id,
                    "image_id": image_id,
                    "refBox": refBox,
                    "ref_id": ref_id,
                })

        return entries

    def tokenize(self):
        """Tokenizes the captions.

        This will add caption_tokens in each entry of the dataset.
        -1 represents nil, and should be treated as padding_idx in embedding.
        """
        for entry in self.entries:
            tokens = self._tokenizer.encode(entry["caption"])
            tokens = [tokens[0]] + tokens[1:-1][:self._max_seq_length -
                                                2] + [tokens[-1]]

            segment_ids = [0] * len(tokens)
            input_mask = [1] * len(tokens)

            if len(tokens) < self._max_seq_length:
                # Note here we pad in front of the sentence
                padding = [self._padding_index
                           ] * (self._max_seq_length - len(tokens))
                tokens = tokens + padding
                input_mask += padding
                segment_ids += padding

            assert_eq(len(tokens), self._max_seq_length)
            entry["token"] = tokens
            entry["input_mask"] = input_mask
            entry["segment_ids"] = segment_ids

    def tensorize(self):
        for entry in self.entries:
            token = torch.from_numpy(np.array(entry["token"]))
            entry["token"] = token

            input_mask = torch.from_numpy(np.array(entry["input_mask"]))
            entry["input_mask"] = input_mask

            segment_ids = torch.from_numpy(np.array(entry["segment_ids"]))
            entry["segment_ids"] = segment_ids

    def __getitem__(self, index):
        entry = self.entries[index]

        image_id = entry["image_id"]
        ref_box = entry["refBox"]

        ref_box = [
            ref_box[0],
            ref_box[1],
            ref_box[0] + ref_box[2],
            ref_box[1] + ref_box[3],
        ]
        features, num_boxes, boxes, boxes_ori = self._image_features_reader[
            image_id]

        boxes_ori = boxes_ori[:num_boxes]
        boxes = boxes[:num_boxes]
        features = features[:num_boxes]

        mix_boxes_ori = boxes_ori
        mix_boxes = boxes
        mix_features = features
        mix_num_boxes = min(int(num_boxes), self._max_region_num)
        mix_target = iou(
            torch.tensor(mix_boxes_ori[:, :4]).float(),
            torch.tensor([ref_box]).float(),
        )

        image_mask = [1] * (mix_num_boxes)
        while len(image_mask) < self._max_region_num:
            image_mask.append(0)

        mix_boxes_pad = np.zeros((self._max_region_num, self._num_locs))
        mix_features_pad = np.zeros((self._max_region_num, 2048))

        mix_boxes_pad[:mix_num_boxes] = mix_boxes[:mix_num_boxes]
        mix_features_pad[:mix_num_boxes] = mix_features[:mix_num_boxes]

        # appending the target feature.
        features = torch.tensor(mix_features_pad).float()
        image_mask = torch.tensor(image_mask).long()
        spatials = torch.tensor(mix_boxes_pad).float()

        target = torch.zeros((self._max_region_num, 1)).float()
        target[:mix_num_boxes] = mix_target[:mix_num_boxes]

        spatials_ori = torch.tensor(mix_boxes_ori).float()

        caption = entry["token"]
        input_mask = entry["input_mask"]
        segment_ids = entry["segment_ids"]

        return features, spatials, image_mask, caption, target, input_mask, segment_ids, image_id

    def __len__(self):
        return len(self.entries)
Пример #8
0
class ReferExpressionDataset(Dataset):
    def __init__(self,
                 task: str,
                 dataroot: str,
                 annotations_jsonpath: str,
                 split: str,
                 image_features_reader: ImageFeaturesH5Reader,
                 gt_image_features_reader: ImageFeaturesH5Reader,
                 tokenizer: BertTokenizer,
                 padding_index: int = 0,
                 max_seq_length: int = 20,
                 max_region_num: int = 60):
        self.split = split
        self.refer = REFER(dataroot, dataset=task, splitBy='unc')
        self.ref_ids = self.refer.getRefIds(split=split)
        print('%s refs are in split [%s].' % (len(self.ref_ids), split))

        self.num_labels = 1
        self._image_features_reader = image_features_reader
        self._gt_image_features_reader = gt_image_features_reader
        self._tokenizer = tokenizer

        self._padding_index = padding_index
        self._max_seq_length = max_seq_length
        self.entries = self._load_annotations()

        self.max_region_num = max_region_num

        if not os.path.exists(os.path.join(dataroot, "cache")):
            os.makedirs(os.path.join(dataroot, "cache"))

        cache_path = os.path.join(
            dataroot, "cache", task + '_' + split + '_' + str(max_seq_length) +
            "_" + str(max_region_num) + '.pkl')
        if not os.path.exists(cache_path):
            self.tokenize()
            self.tensorize()
            cPickle.dump(self.entries, open(cache_path, 'wb'))
        else:
            print('loading entries from %s' % (cache_path))
            self.entries = cPickle.load(open(cache_path, "rb"))

    def _load_annotations(self):

        # annotations_json: Dict[str, Any] = json.load(open(annotations_jsonpath))

        # Build an index which maps image id with a list of caption annotations.
        entries = []

        for ref_id in self.ref_ids:
            ref = self.refer.Refs[ref_id]
            image_id = ref['image_id']
            ref_id = ref['ref_id']
            refBox = self.refer.getRefBox(ref_id)
            for sent, sent_id in zip(ref['sentences'], ref['sent_ids']):
                caption = sent['raw']
                entries.append(
                    {"caption": caption, 'sent_id':sent_id, 'image_id':image_id, \
                    "refBox": refBox, 'ref_id': ref_id}
                    )

        return entries

    def tokenize(self):
        """Tokenizes the captions.

        This will add caption_tokens in each entry of the dataset.
        -1 represents nil, and should be treated as padding_idx in embedding.
        """
        for entry in self.entries:

            sentence_tokens = self._tokenizer.tokenize(entry["caption"])
            sentence_tokens = ["[CLS]"] + sentence_tokens + ["[SEP]"]

            tokens = [
                self._tokenizer.vocab.get(w, self._tokenizer.vocab["[UNK]"])
                for w in sentence_tokens
            ]

            tokens = tokens[:self._max_seq_length]
            segment_ids = [0] * len(tokens)
            input_mask = [1] * len(tokens)

            if len(tokens) < self._max_seq_length:
                # Note here we pad in front of the sentence
                padding = [self._padding_index
                           ] * (self._max_seq_length - len(tokens))
                tokens = tokens + padding
                input_mask += padding
                segment_ids += padding

            assert_eq(len(tokens), self._max_seq_length)
            entry["token"] = tokens
            entry["input_mask"] = input_mask
            entry["segment_ids"] = segment_ids

    def tensorize(self):

        for entry in self.entries:
            token = torch.from_numpy(np.array(entry["token"]))
            entry["token"] = token

            input_mask = torch.from_numpy(np.array(entry["input_mask"]))
            entry["input_mask"] = input_mask

            segment_ids = torch.from_numpy(np.array(entry["segment_ids"]))
            entry["segment_ids"] = segment_ids

    def __getitem__(self, index):
        entry = self.entries[index]

        image_id = entry["image_id"]
        ref_box = entry["refBox"]

        ref_box = [
            ref_box[0], ref_box[1], ref_box[0] + ref_box[2],
            ref_box[1] + ref_box[3]
        ]
        features, num_boxes, boxes, boxes_ori = self._image_features_reader[
            image_id]

        boxes_ori = boxes_ori[:num_boxes]
        boxes = boxes[:num_boxes]
        features = features[:num_boxes]

        if self.split == 'train':
            gt_features, gt_num_boxes, gt_boxes, gt_boxes_ori = self._gt_image_features_reader[
                image_id]

            # merge two boxes, and assign the labels.
            gt_boxes_ori = gt_boxes_ori[1:gt_num_boxes]
            gt_boxes = gt_boxes[1:gt_num_boxes]
            gt_features = gt_features[1:gt_num_boxes]

            # concatenate the boxes
            mix_boxes_ori = np.concatenate((boxes_ori, gt_boxes_ori), axis=0)
            mix_boxes = np.concatenate((boxes, gt_boxes), axis=0)
            mix_features = np.concatenate((features, gt_features), axis=0)
            mix_num_boxes = min(int(num_boxes + int(gt_num_boxes) - 1),
                                self.max_region_num)
            # given the mix boxes, and ref_box, calculate the overlap.
            mix_target = iou(
                torch.tensor(mix_boxes_ori[:, :4]).float(),
                torch.tensor([ref_box]).float())
            mix_target[mix_target < 0.5] = 0

        else:
            mix_boxes_ori = boxes_ori
            mix_boxes = boxes
            mix_features = features
            mix_num_boxes = min(int(num_boxes), self.max_region_num)
            mix_target = iou(
                torch.tensor(mix_boxes_ori[:, :4]).float(),
                torch.tensor([ref_box]).float())

        image_mask = [1] * (mix_num_boxes)
        while len(image_mask) < self.max_region_num:
            image_mask.append(0)

        mix_boxes_pad = np.zeros((self.max_region_num, 5))
        mix_features_pad = np.zeros((self.max_region_num, 2048))

        mix_boxes_pad[:mix_num_boxes] = mix_boxes[:mix_num_boxes]
        mix_features_pad[:mix_num_boxes] = mix_features[:mix_num_boxes]

        # appending the target feature.
        features = torch.tensor(mix_features_pad).float()
        image_mask = torch.tensor(image_mask).long()
        spatials = torch.tensor(mix_boxes_pad).float()

        target = torch.zeros((self.max_region_num, 1)).float()
        target[:mix_num_boxes] = mix_target

        spatials_ori = torch.tensor(mix_boxes_ori).float()
        co_attention_mask = torch.zeros(
            (self.max_region_num, self._max_seq_length))

        caption = entry["token"]
        input_mask = entry["input_mask"]
        segment_ids = entry["segment_ids"]

        return features, spatials, image_mask, caption, target, input_mask, segment_ids, co_attention_mask, image_id

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