Fast_Multi_Style_Transfer-tensorflow/engine_multi.py at master · Corentin-pro/Fast_Multi_Style_Transfer-tensorflow · 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
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
import numpy as np
import tensorflow as tf
if tf.__version__.split('.')[0] == '2':
    import tensorflow.compat.v1 as tf
    tf.disable_v2_behavior()
import tensorflow.compat.v1 as tf1

from src.layers import conv_layer, conv_tranpose_layer, pooling, residual_block


class EngineMultiStyle:
    def __init__(self, tf_session: tf1.Session, content_data_size: int, checkpoint_path: str,
                 style_count=16, style_control=None, resize=None, broadcast_style=None, hidden_out=False):
        self.tf_session = tf_session
        self.content_data_szie = content_data_size
        self.checkpoint_path = checkpoint_path

        if broadcast_style is not None:
            self.image_placeholder = tf1.placeholder(
                tf.uint8, shape=[content_data_size, content_data_size, 3],
                name='image_placeholder')
            image_input = tf.broadcast_to(
                self.image_placeholder,
                [broadcast_style, content_data_size, content_data_size, 3])
            self.style_placeholder = tf1.placeholder(
                tf.float32, shape=[None, style_count], name='style_placeholder')
        elif style_control is None:
            self.image_placeholder = tf1.placeholder(
                tf.uint8, shape=[None, content_data_size, content_data_size, 3],
                name='image_placeholder')
            image_input = self.image_placeholder
            self.style_placeholder = tf1.placeholder(
                tf.float32, shape=[None, style_count], name='style_placeholder')
        else:
            self.image_placeholder = tf1.placeholder(
                tf.uint8, shape=[len(style_control), content_data_size, content_data_size, 3], name='image_placeholder')
            image_input = self.image_placeholder
            self.style_placeholder = tf1.placeholder(
                tf.float32, shape=[len(style_control), style_count], name='style_placeholder')
        self.style_control = style_control

        self.network = self.mst_net(
            tf.cast(image_input, tf.float32),
            style_control=self.style_placeholder if style_control is None else self.style_control)
        # self.output = tf.minimum(tf.maximum(self.network, 0), 255)
        self.output = self.network

        if hidden_out:
            def get_features(tensor_name: str, output_name: str, channel_count: int):
                with tf.variable_scope('Normalization'):
                    output = tf_session.graph.get_tensor_by_name(tensor_name)[1:5, :, :, :channel_count]
                    means, variances = tf.nn.moments(output, [0, 1, 2])
                    output = tf.nn.batch_normalization(output, means, variances, 0.5, 1, 1e-6)
                    output = tf.clip_by_value(output, 0, 1)
                return tf.cast(output * 200.0, tf.uint8, name=output_name)

            get_features('conv3/Relu:0', 'input_features_1', 3)
            get_features('res2_a/conv/Relu:0', 'input_features_2', 6)

            with tf.variable_scope('Normalization'):
                res_conv3 = tf_session.graph.get_tensor_by_name('res3_a/conv/Relu:0')
                res_conv3 = tf.reshape(
                    tf1.nn.avg_pool2d(res_conv3[:1], 10, 10, 'VALID'),
                    [(6 * (content_data_size // 256)) * (6 * (content_data_size // 256)) * 128])
                means, variances = tf.nn.moments(res_conv3, [0])
                res_conv3 = tf.nn.batch_normalization(res_conv3, means, variances, 0.5, 1, 1e-6)
                res_conv3 = tf.round(tf.clip_by_value(res_conv3, 0, 1))
            res_conv3 = tf.cast(res_conv3, dtype=tf.int32, name='sub_latent')
            prev_con3 = tf.placeholder(tf.int32, shape=res_conv3.shape, name='prev_latent')
            res_conv3_change = tf.square(prev_con3 - res_conv3)
            tf.cast(res_conv3_change, dtype=tf.uint8, name='latent_change')

            get_features('res4_a/conv/Relu:0', 'output_features_1', 3)
            get_features('res5_b/add:0', 'output_features_2', 6)

        if resize:
            self.output = tf1.image.resize_bilinear(self.output, (resize, resize))
        self.output = tf.cast(self.output, tf.uint8, name='output')

        train_writer = tf.summary.FileWriter('engine', self.tf_session.graph, flush_secs=20)

        self.saver = tf1.train.Saver(var_list=tf1.trainable_variables())
        self.saver.restore(self.tf_session, self.checkpoint_path)

    def mst_net(self, x, style_control=None, reuse=False):
        with tf1.variable_scope(tf1.get_variable_scope(), reuse=reuse):
            # batch_size, height, width, channels = x.get_shape().as_list()

            x = conv_layer(x, 32, 9, 1, style_control=style_control, name='conv1')
            x = conv_layer(x, 64, 3, 2, style_control=style_control, name='conv2')
            x = conv_layer(x, 128, 3, 2, style_control=style_control, name='conv3')
            x = residual_block(x, 3, style_control=style_control, name='res1')
            x = residual_block(x, 3, style_control=style_control, name='res2')
            x = residual_block(x, 3, style_control=style_control, name='res3')
            x = residual_block(x, 3, style_control=style_control, name='res4')
            x = residual_block(x, 3, style_control=style_control, name='res5')
            x = conv_tranpose_layer(x, 64, 3, 2, style_control=style_control, name='up_conv1',
                                    engine=True)
            x = pooling(x)
            x = conv_tranpose_layer(x, 32, 3, 2, style_control=style_control, name='up_conv2',
                                    engine=True)
            x = pooling(x)
            x = conv_layer(x, 3, 9, 1, relu=False, style_control=style_control, name='output')
            preds = tf.nn.tanh(x) * 127 + 128
        return preds

    def predict(self, images, style_control=None):
        if style_control is None:
            return self.tf_session.run(
                self.output, feed_dict={self.image_placeholder: images})
        return self.tf_session.run(
            self.output, feed_dict={
                self.image_placeholder: images,
                self.style_placeholder: style_control})


def main():
    import argparse
    import time

    from PIL import Image
    import numpy as np

    parser = argparse.ArgumentParser()
    parser.add_argument('input_image', help='Path to image to parse')
    parser.add_argument('checkpoint_path', help='Path to checkpoint to load')
    parser.add_argument('--batch-size', type=int, default=1, help='Batch size to use')
    parser.add_argument('--style', nargs='+',
                        default=[1.] * 16, help='List of weights for style')
    parser.add_argument('--input-size', type=int, default=256, help='Shape of input to use (depends on checkpoint)')
    parser.add_argument('--output-size', type=int, default=None, help='Shape of output image')
    parser.add_argument('--dynamic_style', action='store_true', help='Use dynamic style control')
    arguments = parser.parse_args()

    style_control = np.asarray([[float(value) for value in arguments.style] for _ in range(arguments.batch_size)])

    input_image = Image.open(arguments.input_image).convert('RGB')
    input_image = input_image.resize((arguments.input_size, arguments.input_size))
    input_image = np.asarray(input_image)
    input_image = np.asarray([input_image] * arguments.batch_size)

    gpu_options = tf1.GPUOptions(allow_growth=True)
    session_config = tf1.ConfigProto(gpu_options=gpu_options)

    with tf1.Session(config=session_config).as_default() as session:
        if arguments.dynamic_style:
            engine = EngineMultiStyle(
                session, arguments.input_size, arguments.checkpoint_path, resize=arguments.output_size)
            output = engine.predict(input_image, style_control=style_control)[0]

            start_time = time.time()
            prediction_count = 30
            for _ in range(prediction_count):
                output = engine.predict(input_image, style_control=style_control)[0]
            time_spent = time.time() - start_time
        else:
            engine = EngineMultiStyle(
                session, arguments.input_size, arguments.checkpoint_path,
                style_control=style_control, resize=arguments.output_size)
            output = engine.predict(input_image)[0]

            start_time = time.time()
            prediction_count = 30
            for _ in range(prediction_count):
                output = engine.predict(input_image)[0]
            time_spent = time.time() - start_time

        print('{} predictions in {:.03f}s => {:.02f}FPS ({:.02f} batch/s)'.format(
            prediction_count, time_spent,
            (prediction_count * arguments.batch_size) / time_spent,
            prediction_count / time_spent))

        Image.fromarray(output).save('output.png')


if __name__ == '__main__':
    main()