File size: 10,388 Bytes
2d9a728
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
import os
from torchvision import transforms
from .transforms import *
from .masking_generator import TubeMaskingGenerator, RandomMaskingGenerator
from .mae import VideoMAE
from .mae_multi import VideoMAE_multi
from .kinetics import VideoClsDataset
from .kinetics_sparse import VideoClsDataset_sparse
from .anet import ANetDataset
from .ssv2 import SSVideoClsDataset, SSRawFrameClsDataset
from .hmdb import HMDBVideoClsDataset, HMDBRawFrameClsDataset


class DataAugmentationForVideoMAE(object):
    def __init__(self, args):
        self.input_mean = [0.485, 0.456, 0.406]  # IMAGENET_DEFAULT_MEAN
        self.input_std = [0.229, 0.224, 0.225]  # IMAGENET_DEFAULT_STD
        normalize = GroupNormalize(self.input_mean, self.input_std)
        self.train_augmentation = GroupMultiScaleCrop(args.input_size, [1, .875, .75, .66])
        if args.color_jitter > 0:
            self.transform = transforms.Compose([                            
                self.train_augmentation,
                GroupColorJitter(args.color_jitter),
                GroupRandomHorizontalFlip(flip=args.flip),
                Stack(roll=False),
                ToTorchFormatTensor(div=True),
                normalize,
            ])
        else:
            self.transform = transforms.Compose([                            
                self.train_augmentation,
                GroupRandomHorizontalFlip(flip=args.flip),
                Stack(roll=False),
                ToTorchFormatTensor(div=True),
                normalize,
            ])
        if args.mask_type == 'tube':
            self.masked_position_generator = TubeMaskingGenerator(
                args.window_size, args.mask_ratio
            )
        elif args.mask_type == 'random':
            self.masked_position_generator = RandomMaskingGenerator(
                args.window_size, args.mask_ratio
            )
        elif args.mask_type in 'attention':
            self.masked_position_generator = None

    def __call__(self, images):
        process_data, _ = self.transform(images)
        if self.masked_position_generator is None:
            return process_data, -1
        else:
            return process_data, self.masked_position_generator()

    def __repr__(self):
        repr = "(DataAugmentationForVideoMAE,\n"
        repr += "  transform = %s,\n" % str(self.transform)
        repr += "  Masked position generator = %s,\n" % str(self.masked_position_generator)
        repr += ")"
        return repr


def build_pretraining_dataset(args):
    transform = DataAugmentationForVideoMAE(args)
    dataset = VideoMAE(
        root=None,
        setting=args.data_path,
        prefix=args.prefix,
        split=args.split,
        video_ext='mp4',
        is_color=True,
        modality='rgb',
        num_segments=args.num_segments,
        new_length=args.num_frames,
        new_step=args.sampling_rate,
        transform=transform,
        temporal_jitter=False,
        video_loader=True,
        use_decord=args.use_decord,
        lazy_init=False,
        num_sample=args.num_sample)
    print("Data Aug = %s" % str(transform))
    return dataset


def build_multi_pretraining_dataset(args):
    origianl_flip = args.flip
    transform = DataAugmentationForVideoMAE(args)
    args.flip = False
    transform_ssv2 = DataAugmentationForVideoMAE(args)
    args.flip = origianl_flip

    dataset = VideoMAE_multi(
        root=None,
        setting=args.data_path,
        prefix=args.prefix,
        split=args.split,
        is_color=True,
        modality='rgb',
        num_segments=args.num_segments,
        new_length=args.num_frames,
        new_step=args.sampling_rate,
        transform=transform,
        transform_ssv2=transform_ssv2,
        temporal_jitter=False,
        video_loader=True,
        use_decord=args.use_decord,
        lazy_init=False,
        num_sample=args.num_sample)
    print("Data Aug = %s" % str(transform))
    print("Data Aug for SSV2 = %s" % str(transform_ssv2))
    return dataset


def build_dataset(is_train, test_mode, args):
    print(f'Use Dataset: {args.data_set}')
    if args.data_set in [
            'Kinetics',
            'Kinetics_sparse',
            'mitv1_sparse'
        ]:
        mode = None
        anno_path = None
        if is_train is True:
            mode = 'train'
            anno_path = os.path.join(args.data_path, 'train.csv')
        elif test_mode is True:
            mode = 'test'
            anno_path = os.path.join(args.data_path, 'test.csv') 
        else:  
            mode = 'validation'
            anno_path = os.path.join(args.data_path, 'val.csv') 

        if 'sparse' in args.data_set:
            func = VideoClsDataset_sparse
        else:
            func = VideoClsDataset

        dataset = func(
            anno_path=anno_path,
            prefix=args.prefix,
            split=args.split,
            mode=mode,
            clip_len=args.num_frames,
            frame_sample_rate=args.sampling_rate,
            num_segment=1,
            test_num_segment=args.test_num_segment,
            test_num_crop=args.test_num_crop,
            num_crop=1 if not test_mode else 3,
            keep_aspect_ratio=True,
            crop_size=args.input_size,
            short_side_size=args.short_side_size,
            new_height=256,
            new_width=320,
            args=args)
        
        nb_classes = args.nb_classes
    
    elif args.data_set == 'SSV2':
        mode = None
        anno_path = None
        if is_train is True:
            mode = 'train'
            anno_path = os.path.join(args.data_path, 'train.csv')
        elif test_mode is True:
            mode = 'test'
            anno_path = os.path.join(args.data_path, 'test.csv') 
        else:  
            mode = 'validation'
            anno_path = os.path.join(args.data_path, 'val.csv') 

        if args.use_decord:
            func = SSVideoClsDataset
        else:
            func = SSRawFrameClsDataset

        dataset = func(
            anno_path=anno_path,
            prefix=args.prefix,
            split=args.split,
            mode=mode,
            clip_len=1,
            num_segment=args.num_frames,
            test_num_segment=args.test_num_segment,
            test_num_crop=args.test_num_crop,
            num_crop=1 if not test_mode else 3,
            keep_aspect_ratio=True,
            crop_size=args.input_size,
            short_side_size=args.short_side_size,
            new_height=256,
            new_width=320,
            filename_tmpl=args.filename_tmpl,
            args=args)
        nb_classes = 174

    elif args.data_set == 'UCF101':
        mode = None
        anno_path = None
        if is_train is True:
            mode = 'train'
            anno_path = os.path.join(args.data_path, 'train.csv')
        elif test_mode is True:
            mode = 'test'
            anno_path = os.path.join(args.data_path, 'test.csv') 
        else:  
            mode = 'validation'
            anno_path = os.path.join(args.data_path, 'val.csv') 

        dataset = VideoClsDataset(
            anno_path=anno_path,
            prefix=args.prefix,
            split=args.split,
            mode=mode,
            clip_len=args.num_frames,
            frame_sample_rate=args.sampling_rate,
            num_segment=1,
            test_num_segment=args.test_num_segment,
            test_num_crop=args.test_num_crop,
            num_crop=1 if not test_mode else 3,
            keep_aspect_ratio=True,
            crop_size=args.input_size,
            short_side_size=args.short_side_size,
            new_height=256,
            new_width=320,
            args=args)
        nb_classes = 101
    
    elif args.data_set == 'HMDB51':
        mode = None
        anno_path = None
        if is_train is True:
            mode = 'train'
            anno_path = os.path.join(args.data_path, 'train.csv')
        elif test_mode is True:
            mode = 'test'
            anno_path = os.path.join(args.data_path, 'test.csv') 
        else:  
            mode = 'validation'
            anno_path = os.path.join(args.data_path, 'val.csv') 

        if args.use_decord:
            func = HMDBVideoClsDataset
        else:
            func = HMDBRawFrameClsDataset

        dataset = func(
            anno_path=anno_path,
            prefix=args.prefix,
            split=args.split,
            mode=mode,
            clip_len=1,
            num_segment=args.num_frames,
            test_num_segment=args.test_num_segment,
            test_num_crop=args.test_num_crop,
            num_crop=1 if not test_mode else 3,
            keep_aspect_ratio=True,
            crop_size=args.input_size,
            short_side_size=args.short_side_size,
            new_height=256,
            new_width=320,
            filename_tmpl=args.filename_tmpl,
            args=args)
        nb_classes = 51

    elif args.data_set in [
        'ANet', 
        'HACS',
        'ANet_interval', 
        'HACS_interval'
    ]:
        mode = None
        anno_path = None
        if is_train is True:
            mode = 'train'
            anno_path = os.path.join(args.data_path, 'train.csv')
        elif test_mode is True:
            mode = 'test'
            anno_path = os.path.join(args.data_path, 'test.csv') 
        else:  
            mode = 'validation'
            anno_path = os.path.join(args.data_path, 'val.csv') 

        if 'interval' in args.data_set:
            func = ANetDataset
        else:
            func = VideoClsDataset_sparse

        dataset = func(
            anno_path=anno_path,
            prefix=args.prefix,
            split=args.split,
            mode=mode,
            clip_len=args.num_frames,
            frame_sample_rate=args.sampling_rate,
            num_segment=1,
            test_num_segment=args.test_num_segment,
            test_num_crop=args.test_num_crop,
            num_crop=1 if not test_mode else 3,
            keep_aspect_ratio=True,
            crop_size=args.input_size,
            short_side_size=args.short_side_size,
            new_height=256,
            new_width=320,
            args=args)
        nb_classes = args.nb_classes

    else:
        print(f'Wrong: {args.data_set}')
        raise NotImplementedError()
    assert nb_classes == args.nb_classes
    print("Number of the class = %d" % args.nb_classes)

    return dataset, nb_classes