File size: 8,012 Bytes
872b0a0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import random

import numpy as np
import tensorflow as tf

from .augment import random_crop


def resize_input(t, height, width, resized_h, resized_w):
    # Undo old resizing and apply bilinear
    t = tf.reshape(t, [resized_h, resized_w, 3])
    t = tf.expand_dims(tf.image.resize_image_with_crop_or_pad(t, height, width), 0)
    return tf.image.resize_bilinear(t, [resized_h, resized_w])


def resize_output_crop(t, height, width, channels):
    _, oldh, oldw, c = tf.unstack(tf.shape(t))
    t = tf.reshape(t, [oldh, oldw, c])
    t = tf.image.resize_image_with_crop_or_pad(t, height, width)
    return tf.reshape(t, [1, height, width, channels])


def resize_output(t, height, width, channels):
    return tf.image.resize_bilinear(t, [height, width])


def resize_output_flow(t, height, width, channels):
    batch, old_height, old_width, _ = tf.unstack(tf.shape(t), num=4)
    t = tf.image.resize_bilinear(t, [height, width])
    u, v = tf.unstack(t, axis=3)
    u *= tf.cast(width, tf.float32) / tf.cast(old_width, tf.float32)
    v *= tf.cast(height, tf.float32) / tf.cast(old_height, tf.float32)
    return tf.reshape(tf.stack([u, v], axis=3), [batch, height, width, 2])


def frame_name_to_num(name):
    stripped = name.split('.')[0].lstrip('0')
    if stripped == '':
        return 0
    return int(stripped)


class Input():
    mean = [104.920005, 110.1753, 114.785955]
    stddev = 1 / 0.0039216

    def __init__(self, data, batch_size, dims, *,
                 num_threads=1, normalize=True,
                 skipped_frames=False):
        assert len(dims) == 2
        self.data = data
        self.dims = dims
        self.batch_size = batch_size
        self.num_threads = num_threads
        self.normalize = normalize
        self.skipped_frames = skipped_frames

    def _resize_crop_or_pad(self, tensor):
        height, width = self.dims
        # return tf.image.resize_bilinear(tf.expand_dims(tensor, 0), [height, width])
        return tf.image.resize_image_with_crop_or_pad(tensor, height, width)

    def _resize_image_fixed(self, image):
        height, width = self.dims
        return tf.reshape(self._resize_crop_or_pad(image), [height, width, 3])

    def _normalize_image(self, image):
        return (image - self.mean) / self.stddev

    def _preprocess_image(self, image):
        image = self._resize_image_fixed(image)
        if self.normalize:
            image = self._normalize_image(image)
        return image

    def _input_images(self, image_dir, hold_out_inv=None):
        """Assumes that paired images are next to each other after ordering the
        files.
        """
        image_dir = os.path.join(self.data.current_dir, image_dir)

        filenames_1 = []
        filenames_2 = []
        image_files = os.listdir(image_dir)
        image_files.sort()

        assert len(image_files) % 2 == 0, 'expected pairs of images'

        for i in range(len(image_files) // 2):
            filenames_1.append(os.path.join(image_dir, image_files[i * 2]))
            filenames_2.append(os.path.join(image_dir, image_files[i * 2 + 1]))

        if hold_out_inv is not None:
            filenames = list(zip(filenames_1, filenames_2))
            random.seed(0)
            random.shuffle(filenames)
            filenames = filenames[:hold_out_inv]

            filenames_1, filenames_2 = zip(*filenames)
            filenames_1 = list(filenames_1)
            filenames_2 = list(filenames_2)

        input_1 = read_png_image(filenames_1, 1)
        input_2 = read_png_image(filenames_2, 1)
        image_1 = self._preprocess_image(input_1)
        image_2 = self._preprocess_image(input_2)
        return tf.shape(input_1), image_1, image_2

    def _input_test(self, image_dir, hold_out_inv=None):
        input_shape, im1, im2 = self._input_images(image_dir, hold_out_inv)
        return tf.train.batch(
            [im1, im2, input_shape],
            batch_size=self.batch_size,
            num_threads=self.num_threads,
            allow_smaller_final_batch=True)

    def get_normalization(self):
        return self.mean, self.stddev

    def input_raw(self, swap_images=True, sequence=True,
                  needs_crop=True, shift=0, seed=0,
                  center_crop=False, skip=0):
        """Constructs input of raw data.

        Args:
            sequence: Assumes that image file order in data_dirs corresponds to
                temporal order, if True. Otherwise, assumes uncorrelated pairs of
                images in lexicographical ordering.
            shift: number of examples to shift the input queue by.
                Useful to resume training.
            swap_images: for each pair (im1, im2), also include (im2, im1)
            seed: seed for filename shuffling.
        Returns:
            image_1: batch of first images
            image_2: batch of second images
        """
        if not isinstance(skip, list):
            skip = [skip]

        data_dirs = self.data.get_raw_dirs()
        height, width = self.dims
        #assert batch_size % 2 == 0

        filenames = []
        for dir_path in data_dirs:
            files = os.listdir(dir_path)
            files.sort()
            if sequence:
                steps = [1 + s for s in skip]
                stops = [len(files) - s for s in steps]
            else:
                steps = [2]
                stops = [len(files)]
                assert len(files) % 2 == 0
            for step, stop in zip(steps, stops):
                for i in range(0, stop, step):
                    if self.skipped_frames and sequence:
                        assert step == 1
                        num_first = frame_name_to_num(files[i])
                        num_second = frame_name_to_num(files[i+1])
                        if num_first + 1 != num_second:
                            continue
                    fn1 = os.path.join(dir_path, files[i])
                    fn2 = os.path.join(dir_path, files[i + 1])
                    filenames.append((fn1, fn2))

        random.seed(seed)
        random.shuffle(filenames)
        print("Training on {} frame pairs.".format(len(filenames)))

        filenames_extended = []
        for fn1, fn2 in filenames:
            filenames_extended.append((fn1, fn2))
            if swap_images:
                filenames_extended.append((fn2, fn1))

        shift = shift % len(filenames_extended)
        filenames_extended = list(np.roll(filenames_extended, shift))


        filenames_1, filenames_2 = zip(*filenames_extended)
        filenames_1 = list(filenames_1)
        filenames_2 = list(filenames_2)

        with tf.variable_scope('train_inputs'):
            image_1 = read_png_image(filenames_1)
            image_2 = read_png_image(filenames_2)

            if needs_crop:
                #if center_crop:
                #    image_1 = tf.image.resize_image_with_crop_or_pad(image_1, height, width)
                #    image_2 = tf.image.resize_image_with_crop_or_pad(image_1, height, width)
                #else:
                image_1, image_2 = random_crop([image_1, image_2], [height, width, 3])
            else:
                image_1 = tf.reshape(image_1, [height, width, 3])
                image_2 = tf.reshape(image_2, [height, width, 3])

            if self.normalize:
                image_1 = self._normalize_image(image_1)
                image_2 = self._normalize_image(image_2)

            return tf.train.batch(
                [image_1, image_2],
                batch_size=self.batch_size,
                num_threads=self.num_threads)


def read_png_image(filenames, num_epochs=None):
    """Given a list of filenames, constructs a reader op for images."""
    filename_queue = tf.train.string_input_producer(filenames,
        shuffle=False, capacity=len(filenames))
    reader = tf.WholeFileReader()
    _, value = reader.read(filename_queue)
    image_uint8 = tf.image.decode_png(value, channels=3)
    image = tf.cast(image_uint8, tf.float32)
    return image