-
Notifications
You must be signed in to change notification settings - Fork 39
/
Copy pathgenerate_header_and_model.py
191 lines (152 loc) · 7.98 KB
/
generate_header_and_model.py
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
import tensorflow as tf
import numpy as np
import argparse
import os
import enum
from models.model_espcn import ESPCN
from models.model_srcnn import SRCNN
from models.model_vespcn import VESPCN
from models.model_vsrnet import VSRnet
from collections import OrderedDict
@enum.unique
class Padding(enum.Enum):
Valid = 0
Same = 1
Same_clamp_to_edge = 2
def get_arguments():
parser = argparse.ArgumentParser(description='generate c header with model weights and binary model file')
parser.add_argument('--model', type=str, default='srcnn', choices=['srcnn', 'espcn', 'vespcn', 'vsrnet'],
help='What model to use for generation')
parser.add_argument('--output_folder', type=str, default='./',
help='where to put generated files')
parser.add_argument('--ckpt_path', default=None,
help='Path to the model checkpoint, from which weights are loaded')
parser.add_argument('--use_mc', action='store_true',
help='Whether motion compensation is used in video super resolution model')
parser.add_argument('--scale_factor', type=int, default=2, choices=[2, 3, 4],
help='What scale factor was used for chosen model')
return parser.parse_args()
def dump_to_file(file, values, name):
file.write('\nstatic const float ' + name + '[] = {\n')
values_flatten = values.flatten()
max_len = 0
for value in values_flatten:
if len(str(value)) > max_len:
max_len = len(str(value))
counter = 0
for i in range(len(values_flatten)):
counter += 1
if counter == 4:
file.write(str(values_flatten[i]) + 'f')
if i != len(values_flatten) - 1:
file.write(',')
file.write('\n')
counter = 0
else:
if counter == 1:
file.write(' ')
file.write(str(values_flatten[i]) + 'f')
if i != len(values_flatten) - 1:
file.write(',')
file.write(' ' * (1 + max_len - len(str(values_flatten[i]))))
if counter != 0:
file.write('\n')
file.write('};\n')
file.write('\nstatic const long int ' + name + '_dims[] = {\n')
for i in range(len(values.shape)):
file.write(' ')
file.write(str(values.shape[i]))
if i != len(values.shape) - 1:
file.write(',\n')
file.write('\n};\n')
def write_conv_layer(kernel, bias, dilation_rate, padding, activation, model_file):
kernel = np.transpose(kernel, [3, 0, 1, 2])
np.array([1, dilation_rate, padding.value, activation, kernel.shape[3], kernel.shape[0], kernel.shape[1]], dtype=np.uint32).tofile(model_file)
kernel.tofile(model_file)
bias.tofile(model_file)
def write_depth_to_space_layer(block_size, model_file):
np.array([2, block_size], dtype=np.uint32).tofile(model_file)
def prepare_native_mf_srcnn(weights, model_file):
np.array([3], dtype=np.uint32).tofile(model_file)
write_conv_layer(weights['srcnn/conv1/kernel:0'], weights['srcnn/conv1/bias:0'], 1, Padding.Same_clamp_to_edge, 0, model_file)
write_conv_layer(weights['srcnn/conv2/kernel:0'], weights['srcnn/conv2/bias:0'], 1, Padding.Same_clamp_to_edge, 0, model_file)
write_conv_layer(weights['srcnn/conv3/kernel:0'], weights['srcnn/conv3/bias:0'], 1, Padding.Same_clamp_to_edge, 0, model_file)
def prepare_native_mf_espcn(weights, model_file, scale_factor):
np.array([4], dtype=np.uint32).tofile(model_file)
write_conv_layer(weights['espcn/conv1/kernel:0'], weights['espcn/conv1/bias:0'], 1, Padding.Same_clamp_to_edge, 1, model_file)
write_conv_layer(weights['espcn/conv2/kernel:0'], weights['espcn/conv2/bias:0'], 1, Padding.Same_clamp_to_edge, 1, model_file)
write_conv_layer(weights['espcn/conv3/kernel:0'], weights['espcn/conv3/bias:0'], 1, Padding.Same_clamp_to_edge, 2, model_file)
write_depth_to_space_layer(scale_factor, model_file)
def prepare_native_mf_vespcn(weights, model_file, scale_factor):
np.array([6], dtype=np.uint32).tofile(model_file)
write_conv_layer(weights['vespcn/conv1/kernel:0'], weights['vespcn/conv1/bias:0'], 1, Padding.Same_clamp_to_edge, 0, model_file)
write_conv_layer(weights['vespcn/conv2/kernel:0'], weights['vespcn/conv2/bias:0'], 1, Padding.Same_clamp_to_edge, 0, model_file)
write_conv_layer(weights['vespcn/conv3/kernel:0'], weights['vespcn/conv3/bias:0'], 1, Padding.Same_clamp_to_edge, 0, model_file)
write_conv_layer(weights['vespcn/conv4/kernel:0'], weights['vespcn/conv4/bias:0'], 1, Padding.Same_clamp_to_edge, 0, model_file)
write_conv_layer(weights['vespcn/conv5/kernel:0'], weights['vespcn/conv5/bias:0'], 1, Padding.Same_clamp_to_edge, 0, model_file)
write_depth_to_space_layer(scale_factor, model_file)
def prepare_native_mf_vsrnet(weights, model_file):
np.array([3], dtype=np.uint32).tofile(model_file)
write_conv_layer(weights['vsrnet/conv1/kernel:0'], weights['vsrnet/conv1/bias:0'], 1, Padding.Same_clamp_to_edge, 0, model_file)
write_conv_layer(weights['vsrnet/conv2/kernel:0'], weights['vsrnet/conv2/bias:0'], 1, Padding.Same_clamp_to_edge, 0, model_file)
write_conv_layer(weights['vsrnet/conv3/kernel:0'], weights['vsrnet/conv3/bias:0'], 1, Padding.Same_clamp_to_edge, 0, model_file)
def main():
args = get_arguments()
if not os.path.exists(args.output_folder):
os.mkdir(args.output_folder)
if args.ckpt_path is None:
print("Path to the checkpoint file was not provided")
exit(1)
if args.model == 'srcnn':
model = SRCNN(args)
elif args.model == 'espcn':
model = ESPCN(args)
elif args.model == 'vespcn':
model = VESPCN(args)
elif args.model == 'vsrnet':
model = VSRnet(args)
else:
exit(1)
with tf.Session() as sess:
input_ph = model.get_placeholder()
predicted = model.load_model(input_ph)
if args.model == 'vespcn':
predicted = predicted[2]
predicted = tf.identity(predicted, name='y')
if os.path.isdir(args.ckpt_path):
args.ckpt_path = tf.train.latest_checkpoint(args.ckpt_path)
saver = tf.train.Saver()
saver.restore(sess, args.ckpt_path)
with open(os.path.join(args.output_folder, args.model + '.model'), 'wb') as native_mf:
weights = model.get_model_weights(sess)
if args.model == 'srcnn':
prepare_native_mf_srcnn(weights, native_mf)
elif args.model == 'espcn':
prepare_native_mf_espcn(weights, native_mf, args.scale_factor)
elif args.model == 'vespcn':
prepare_native_mf_vespcn(weights, native_mf, args.scale_factor)
elif args.model == 'vsrnet':
prepare_native_mf_vsrnet(weights, native_mf)
with open(os.path.join(args.output_folder, 'dnn_' + args.model + '.h'), 'w') as header:
header.write('/**\n')
header.write(' * @file\n')
header.write(' * Default cnn weights for x' + str(args.scale_factor) + ' upscaling with ' +
args.model + ' model.\n')
header.write(' */\n\n')
header.write('#ifndef AVFILTER_DNN_' + args.model.upper() + '_H\n')
header.write('#define AVFILTER_DNN_' + args.model.upper() + '_H\n')
variables = tf.trainable_variables()
var_dict = OrderedDict()
for variable in variables:
var_name = variable.name.split(':')[0].replace('/', '_')
value = variable.eval()
if 'kernel' in var_name:
value = np.transpose(value, axes=(3, 0, 1, 2))
var_dict[var_name] = value
for name, value in var_dict.items():
dump_to_file(header, value, name)
header.write('#endif\n')
output_graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, ['y'])
tf.train.write_graph(output_graph_def, args.output_folder, args.model + '.pb', as_text=False)
if __name__ == '__main__':
main()