File size: 6,780 Bytes
ee21b96
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
# Copyright 2022 The OFA-Sys Team. 
# All rights reserved.
# This source code is licensed under the Apache 2.0 license 
# found in the LICENSE file in the root directory.

from io import BytesIO

import logging
import warnings
import functools

import numpy as np
import torch
import base64
from torchvision import transforms
from timm.data import create_transform
from utils.vision_helper import RandomAugment

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])

    conf = None
    if samples[0].get("conf", None) is not None:
        conf = torch.cat([s['conf'] for s in samples], dim=0)

    ref_dict = None
    if samples[0].get("ref_dict", None) is not None:
        ref_dict = np.array([s['ref_dict'] for s in samples])

    constraint_masks = None
    if samples[0].get("constraint_mask", None) is not None:
        constraint_masks = merge("constraint_mask")

    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
        },
        "conf": conf,
        "ref_dict": ref_dict,
        "constraint_masks": constraint_masks,
        "target": target,
    }

    return batch


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

        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]

        if self.split != 'train':
            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),
            ])
            logger.info("val split, do not use random augmentation.")
        else:
            self.patch_resize_transform = create_transform(
                input_size=patch_image_size,
                is_training=True,
                color_jitter=0.4,
                auto_augment='rand-m9-mstd0.5-inc1',
                interpolation='bicubic',
                re_prob=0.25,
                re_mode='pixel',
                re_count=1,
                mean=mean,
                std=std,
            )
            self.patch_resize_transform = transforms.Compose(functools.reduce(lambda x, y:x + y, [
                [lambda image: image.convert("RGB"),],
                self.patch_resize_transform.transforms[:2],
                [self.patch_resize_transform.transforms[2]],
                [RandomAugment(2, 7, isPIL=True, augs=['Identity', 'AutoContrast', 'Equalize', 'Brightness', 'Sharpness', 'ShearX', 'ShearY', 'TranslateX', 'TranslateY', 'Rotate']), ],
                self.patch_resize_transform.transforms[3:],
            ]))
            logger.info("train split, use random augmentation.")

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

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

        src_item = self.encode_text(' what does the image describe?')
        tgt_item = self.encode_text(" {}".format(label_name))
        ref_dict = {label_name: 1.0}

        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": index,
            "source": src_item,
            "patch_image": patch_image,
            "patch_mask": patch_mask,
            "target": target_item,
            "prev_output_tokens": prev_output_item,
            "ref_dict": ref_dict,
        }
        if self.constraint_trie is not None:
            constraint_mask = torch.zeros((len(prev_output_item), len(self.tgt_dict))).bool()
            for i in range(len(prev_output_item)):
                constraint_prefix_token = prev_output_item[:i+1].tolist()
                constraint_nodes = self.constraint_trie.get_next_layer(constraint_prefix_token)
                constraint_mask[i][constraint_nodes] = True
            example["constraint_mask"] = constraint_mask
        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 containing the data of the task
        """
        return collate(samples, pad_idx=self.pad, eos_idx=self.eos)