File size: 5,476 Bytes
c99d198
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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

"""API factory."""

import functools
import os.path as osp

import albumentations as alb
import torch
from xirl import models
from xirl import transforms

TRANSFORMS = {
    "random_resized_crop":
        functools.partial(
            alb.RandomResizedCrop, scale=(0.8, 1.0), ratio=(0.75, 1.333),
            p=1.0),
    "center_crop":
        functools.partial(alb.CenterCrop, p=1.0),
    "global_resize":
        functools.partial(alb.Resize, p=1.0),
    "grayscale":
        functools.partial(alb.ToGray, p=0.2),
    "vertical_flip":
        functools.partial(alb.VerticalFlip, p=0.5),
    "horizontal_flip":
        functools.partial(alb.HorizontalFlip, p=0.5),
    "gaussian_blur":
        functools.partial(
            alb.GaussianBlur,
            blur_limit=(13, 13),
            sigma_limit=(1.0, 2.0),
            p=0.2,
        ),
    "color_jitter":
        functools.partial(
            alb.ColorJitter,
            brightness=0.4,
            contrast=0.4,
            hue=0.1,
            saturation=0.1,
            p=0.8,
        ),
    "rotate":
        functools.partial(alb.Rotate, limit=(-5, 5), border_mode=0, p=0.5),
    "normalize":
        functools.partial(
            alb.Normalize,
            mean=transforms.PretrainedMeans.IMAGENET,
            std=transforms.PretrainedStds.IMAGENET,
            p=1.0,
        ),
}

MODELS = {
    "resnet18_linear": models.Resnet18LinearEncoderNet,
    "resnet18_clip_linear": models.Resnet18LinearEncoderAndTextEncoderNet,
    "resnet18_classifier": models.GoalClassifier,
    "resnet18_features": models.Resnet18RawImageNetFeaturesNet,
    "resnet18_linear_ae": models.Resnet18LinearEncoderAutoEncoderNet,
}

def model_from_config(config):
    """Create a model from a config."""
    kwargs = {
        "num_ctx_frames": config.frame_sampler.num_context_frames,
        "normalize_embeddings": config.model.normalize_embeddings,
        "learnable_temp": config.model.learnable_temp,
    }
    if config.model.model_type == "resnet18_linear":
        kwargs["embedding_size"] = config.model.embedding_size
    elif config.model.model_type == "resnet18_clip_linear":
        kwargs["embedding_size"] = config.model.embedding_size
    elif config.model.model_type == "resnet18_linear_ae":
        kwargs["embedding_size"] = config.model.embedding_size
    return MODELS[config.model.model_type](**kwargs)

def create_transform(name, *args, **kwargs):
    """Create an image augmentation from its name and args."""
    # pylint: disable=invalid-name
    if "::" in name:
        # e.g., `rotate::{'limit': (-45, 45)}`
        name, __kwargs = name.split("::")
        _kwargs = eval(__kwargs)  # pylint: disable=eval-used
    else:
        _kwargs = {}
    _kwargs.update(kwargs)
    return TRANSFORMS[name](*args, **_kwargs)

def dataset_from_config(config, downstream, split, debug):
    """Create a video dataset from a config."""
    dataset_path = osp.join(config.data.root, split)

    image_size = config.data_augmentation.image_size
    if isinstance(image_size, int):
        image_size = (image_size, image_size)
    image_size = tuple(image_size)

    # Note(kevin): We used to disable data augmentation on all downstream
    # dataloaders. I've decided to keep them for train downstream loaders.
    if debug:
        # The minimum data augmentation we want to keep is resizing when
        # debugging.
        aug_names = ["global_resize"]
    else:
        if split == "train":
            aug_names = config.data_augmentation.train_transforms
        else:
            aug_names = config.data_augmentation.eval_transforms

    # Create a list of data augmentation callables.
    aug_funcs = []
    for name in aug_names:
        if "resize" in name or "crop" in name:
            aug_funcs.append(create_transform(name, *image_size))
        else:
            aug_funcs.append(create_transform(name))

    augmentor = transforms.VideoAugmentor({SequenceType.FRAMES: aug_funcs})

    # Restrict action classes if they have been provided. Else, load all
    # from the data directory.
    c_action_class = (
        config.data.downstream_action_class
        if downstream else config.data.pretrain_action_class
    )
    if c_action_class:
        action_classes = c_action_class
    else:
        action_classes = get_subdirs(
            dataset_path,
            basename=True,
            nonempty=True,
            sort_lexicographical=True,
        )

    # We need to separate out the dataclasses for each action class when
    # creating downstream datasets.
    if downstream:
        dataset = {}
        for action_class in action_classes:
            frame_sampler = frame_sampler_from_config(config, downstream=True)
            single_class_dataset = VideoDataset(
                dataset_path,
                frame_sampler,
                seed=config.seed,
                augmentor=augmentor,
                max_vids_per_class=config.data.max_vids_per_class,
            )
            single_class_dataset.restrict_subdirs(action_class)
            dataset[action_class] = single_class_dataset
    else:
        frame_sampler = frame_sampler_from_config(config, downstream=False)
        dataset = VideoDataset(
            dataset_path,
            frame_sampler,
            seed=config.seed,
            augmentor=augmentor,
            max_vids_per_class=config.data.max_vids_per_class,
        )
        dataset.restrict_subdirs(action_classes)

    return dataset