File size: 4,975 Bytes
ce922b3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
from io import BytesIO

import logging
import warnings
import string

import numpy as np
import torch
import base64
from torchvision import transforms

from PIL import Image, ImageFile

from data import data_utils
from data.ofa_dataset import OFADataset

ImageFile.LOAD_TRUNCATED_IMAGES = True
ImageFile.MAX_IMAGE_PIXELS = None
Image.MAX_IMAGE_PIXELS = None

logger = logging.getLogger(__name__)
warnings.filterwarnings("ignore", "(Possibly )?corrupt EXIF data", UserWarning)

IMAGENET_DEFAULT_MEAN = (0.485, 0.456, 0.406)
IMAGENET_DEFAULT_STD = (0.229, 0.224, 0.225)


def collate(samples, pad_idx, eos_idx):
    if len(samples) == 0:
        return {}

    def merge(key):
        return data_utils.collate_tokens(
            [s[key] for s in samples],
            pad_idx,
            eos_idx=eos_idx,
        )

    id = np.array([s["id"] for s in samples])
    src_tokens = merge("source")
    src_lengths = torch.LongTensor([s["source"].ne(pad_idx).long().sum() for s in samples])

    patch_images = torch.stack([sample['patch_image'] for sample in samples], dim=0)
    patch_masks = torch.cat([sample['patch_mask'] for sample in samples])

    prev_output_tokens = None
    target = None
    if samples[0].get("target", None) is not None:
        target = merge("target")
        tgt_lengths = torch.LongTensor([s["target"].ne(pad_idx).long().sum() for s in samples])
        ntokens = tgt_lengths.sum().item()

        if samples[0].get("prev_output_tokens", None) is not None:
            prev_output_tokens = merge("prev_output_tokens")
    else:
        ntokens = src_lengths.sum().item()

    batch = {
        "id": id,
        "nsentences": len(samples),
        "ntokens": ntokens,
        "net_input": {
            "src_tokens": src_tokens,
            "src_lengths": src_lengths,
            "patch_images": patch_images,
            "patch_masks": patch_masks,
            "prev_output_tokens": prev_output_tokens
        },
        "target": target,
    }

    return batch


class CaptionDataset(OFADataset):
    def __init__(
        self,
        split,
        dataset,
        bpe,
        src_dict,
        tgt_dict=None,
        max_src_length=128,
        max_tgt_length=30,
        patch_image_size=224,
        imagenet_default_mean_and_std=False,
        scst=False
    ):
        super().__init__(split, dataset, bpe, src_dict, tgt_dict)
        self.max_src_length = max_src_length
        self.max_tgt_length = max_tgt_length
        self.patch_image_size = patch_image_size
        self.scst = scst

        self.transtab = str.maketrans({key: None for key in string.punctuation})

        if imagenet_default_mean_and_std:
            mean = IMAGENET_DEFAULT_MEAN
            std = IMAGENET_DEFAULT_STD
        else:
            mean = [0.5, 0.5, 0.5]
            std = [0.5, 0.5, 0.5]

        self.patch_resize_transform = transforms.Compose([
            lambda image: image.convert("RGB"),
            transforms.Resize((patch_image_size, patch_image_size), interpolation=Image.BICUBIC),
            transforms.ToTensor(),
            transforms.Normalize(mean=mean, std=std),
        ])

    def __getitem__(self, index):
        uniq_id, image, caption = self.dataset[index]

        image = Image.open(BytesIO(base64.urlsafe_b64decode(image)))
        patch_image = self.patch_resize_transform(image)
        patch_mask = torch.tensor([True])

        if self.split == 'train' and not self.scst:
            caption = caption.translate(self.transtab).strip()
            caption_token_list = caption.strip().split()
            tgt_caption = ' '.join(caption_token_list[:self.max_tgt_length])
        else:
            caption = ' '.join(caption.strip().split())
            caption_list = [cap.translate(self.transtab).strip() for cap in caption.strip().split('&&')]
            tgt_caption = '&&'.join(caption_list)
        src_item = self.encode_text(" what does the image describe?")
        tgt_item = self.encode_text(" {}".format(tgt_caption))

        src_item = torch.cat([self.bos_item, src_item, self.eos_item])
        target_item = torch.cat([tgt_item, self.eos_item])
        prev_output_item = torch.cat([self.bos_item, tgt_item])

        example = {
            "id": uniq_id,
            "source": src_item,
            "patch_image": patch_image,
            "patch_mask": patch_mask,
            "target": target_item,
            "prev_output_tokens": prev_output_item
        }
        return example

    def collater(self, samples, pad_to_length=None):
        """Merge a list of samples to form a mini-batch.
        Args:
            samples (List[dict]): samples to collate
        Returns:
            dict: a mini-batch with the following keys:
        """
        return collate(samples, pad_idx=self.pad, eos_idx=self.eos)