temporal-ssl/PreprocessorVideoSSL.py at master · sjenni/temporal-ssl · GitHub

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
import tensorflow as tf

slim = tf.contrib.slim


class PreprocessorTransform:
    def __init__(self, seq_length, n_speeds=1, crop_size=(256, 256), resize_shape=(256, 256), src_shape=(256, 256),
                 flip_prob=0.5, transforms=('foba', 'shuffle', 'warp'), augment_color=False, rand_grey=0.):
        self.crop_size = crop_size
        self.src_shape = src_shape
        self.resize_shape = resize_shape
        self.seq_length = seq_length
        self.n_speeds = n_speeds
        self.flip_prob = flip_prob
        self.transforms = transforms
        self.augment_color = augment_color
        self.rand_grey = rand_grey
        self.load_shape = (seq_length,) + src_shape + (3,)
        self.out_shape = (len(self.transforms) + 1, seq_length) + crop_size + (3,)

    def color_scale_vid(self, vid):
        vid = tf.cast(vid, tf.float32) / 255.
        vid = tf.map_fn(self.color, vid)
        return vid

    def load_inds(self, example, inds):
        images = tf.map_fn(lambda i: tf.image.decode_jpeg(example["frames"].values[i]), inds, dtype=tf.uint8)
        images = tf.reverse(images, axis=[-1])  # To RGB
        images = tf.reshape(images, [-1, self.load_shape[1], self.load_shape[2], 3])
        return images

    def sub_sample_seq(self, vid, n_frames, skip):
        max_frame = n_frames - self.seq_length * skip + 1
        random_offset = tf.random.uniform(shape=(), minval=0, maxval=max_frame, dtype=tf.int64)
        inds = tf.range(random_offset, random_offset + self.seq_length * skip, delta=skip)
        sub_inds = tf.gather(vid, inds)
        return sub_inds

    def process_train(self, example):
        full_inds = tf.range(0, example["num_frames"])

        # Choose a speed
        max_speed = tf.minimum(tf.cast(log2(example["num_frames"] / self.seq_length), tf.int64), self.n_speeds-1)
        speed_label = tf.cond(tf.greater(max_speed, 0),
                              lambda: tf.random.uniform(shape=(), minval=0, maxval=max_speed+1, dtype=tf.int64),
                              lambda: tf.constant(0, tf.int64))
        eff_skip = 2 ** speed_label

        # Speed type video
        vid_ori_inds = self.sub_sample_seq(full_inds, example["num_frames"], eff_skip)
        vid_orig = self.load_inds(example, vid_ori_inds)
        vid_orig.set_shape(self.load_shape)
        vids_transformed = [vid_orig]

        # Periodic (forward/backward) transformation
        if 'foba' in self.transforms:
            sub_len = tf.minimum(example["num_frames"], (self.seq_length - 1) * eff_skip)
            crop_inds = tf.random_crop(full_inds, [sub_len])
            inds = tf.range(0, (self.seq_length - 2) * eff_skip, delta=eff_skip)
            start = tf.cond(tf.greater(eff_skip, 1),
                            lambda: tf.constant(1, tf.int64),
                            lambda: tf.constant(0, tf.int64))
            random_offset = tf.random.uniform(shape=(), minval=start, maxval=eff_skip + 1 - start, dtype=tf.int64)
            inds_foba = tf.concat([inds, random_offset + tf.reverse(inds, [0])], 0)
            inds_foba = tf.random_crop(inds_foba, [self.seq_length])
            vid_foba_inds = tf.gather(crop_inds, inds_foba)
            vid_foba = self.load_inds(example, vid_foba_inds)
            vid_foba.set_shape(self.load_shape)
            vids_transformed.append(vid_foba)

        # Random transformation
        if 'shuffle' in self.transforms:
            vid_shuffle_inds = self.sub_sample_seq(full_inds, example["num_frames"], 1)
            shuffle_inds = tf.random_shuffle(tf.range(0, self.seq_length))
            vid_shuffle_inds = tf.gather(vid_shuffle_inds, shuffle_inds)
            vid_shuffle = self.load_inds(example, vid_shuffle_inds)
            vid_shuffle.set_shape(self.load_shape)
            vids_transformed.append(vid_shuffle)

        # Warp transformation
        if 'warp' in self.transforms:
            sub_len = tf.cond(tf.greater(max_speed, 1),
                              lambda: tf.cast(self.seq_length * 2 ** max_speed, tf.int64),
                              lambda: example["num_frames"])
            max_offset = 2 ** max_speed
            off_sets = tf.random.uniform(shape=(self.seq_length,), minval=1, maxval=max_offset + 1, dtype=tf.int64)
            inds_warp_v3 = tf.cumsum(off_sets)

            # Special treatment for speed0
            inds_warp_v1 = tf.random_shuffle(tf.range(0, sub_len))
            inds_warp_v1 = inds_warp_v1[:self.seq_length]
            inds_warp_v1 = tf.sort(inds_warp_v1)
            inds_warp = tf.cond(tf.greater(max_speed, 1), lambda: inds_warp_v3, lambda: inds_warp_v1)

            vid_warp_inds = tf.random_crop(full_inds, [sub_len])
            vid_warp_inds = tf.gather(vid_warp_inds, inds_warp)
            vid_warp = self.load_inds(example, vid_warp_inds)
            vid_warp.set_shape(self.load_shape)
            vids_transformed.append(vid_warp)

        vids_transformed = [self.resize_crop(self.color_scale_vid(v)) for v in vids_transformed]
        vids_transformed = tf.stack(vids_transformed)

        return vids_transformed, speed_label, example

    def flip_lr(self, vid):
        flipped_vids = tf.map_fn(lambda x: flip_video(x, self.flip_prob), vid)
        return flipped_vids

    def color(self, image, bright_max_delta=32. / 255., lower_sat=0.5, upper_sat=1.5):
        if self.augment_color:
            image = tf.cond(tf.random_uniform(shape=(), minval=0.0, maxval=1.0) > 0.5,
                            true_fn=lambda: tf.image.random_saturation(
                                tf.image.random_brightness(image, max_delta=bright_max_delta),
                                lower=lower_sat, upper=upper_sat),
                            false_fn=lambda: tf.image.random_brightness(
                                tf.image.random_saturation(image, lower=lower_sat, upper=upper_sat),
                                max_delta=bright_max_delta))

        if self.rand_grey > 0.:
            image = tf.cond(tf.random_uniform(shape=(), minval=0.0, maxval=1.0) > self.rand_grey,
                            true_fn=lambda: image,
                            false_fn=lambda: tf.tile(tf.image.rgb_to_grayscale(image), [1, 1, 3]))

        # Scale to [-1, 1]
        image = tf.cast(image, tf.float32) * 2. - 1.
        image = tf.clip_by_value(image, -1., 1.)
        return image

    def random_crop(self, vid):
        in_sz = vid.get_shape().as_list()
        return tf.image.random_crop(vid, [in_sz[0], self.crop_size[0], self.crop_size[1], in_sz[-1]])

    def resize_crop(self, vid):
        if not self.resize_shape == self.src_shape:
            resized_video = tf.compat.v1.image.resize_bilinear(vid, self.resize_shape)
            return self.random_crop(resized_video)
        else:
            return self.random_crop(vid)

    def augment_train(self, video):
        video = self.flip_lr(video)
        return video


def flip_video(vid, flip_prob):
    return tf.cond(tf.random.uniform(shape=(), minval=0.0, maxval=1.0) > flip_prob,
                   true_fn=lambda: vid,
                   false_fn=lambda: tf.reverse(vid, [3]))


def log2(x):
    numerator = tf.log(x)
    denominator = tf.log(tf.constant(2, dtype=numerator.dtype))
    return numerator / denominator