-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
305 lines (234 loc) · 14.9 KB
/
main.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
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
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
#! /usr/bin/python
# -*- coding: utf8 -*-
import os, time, pickle, random, time
from datetime import datetime
import numpy as np
from time import localtime, strftime
import logging, scipy
import tensorflow as tf
import tensorlayer as tl
from model import SRGAN_g, SRGAN_d, Vgg19_simple_api
from utils import *
from config import config, log_config
###====================== HYPER-PARAMETERS ===========================###
## Adam
batch_size = config.TRAIN.batch_size
lr_init = config.TRAIN.lr_init
beta1 = config.TRAIN.beta1
## initialize G
n_epoch_init = config.TRAIN.n_epoch_init
## adversarial learning (SRGAN)
n_epoch = config.TRAIN.n_epoch
lr_decay = config.TRAIN.lr_decay
decay_every = config.TRAIN.decay_every
ni = int(np.sqrt(batch_size))
def train():
## create folders to save result images and trained model
save_dir_ginit = "samples/{}_ginit".format(tl.global_flag['mode'])
save_dir_gan = "samples/{}_gan".format(tl.global_flag['mode'])
tl.files.exists_or_mkdir(save_dir_ginit)
tl.files.exists_or_mkdir(save_dir_gan)
checkpoint_dir = "checkpoint" # checkpoint_resize_conv
tl.files.exists_or_mkdir(checkpoint_dir)
###====================== PRE-LOAD DATA ===========================###
train_hr_img_list = sorted(tl.files.load_file_list(path=config.TRAIN.hr_img_path, regx='.*.png', printable=False))
train_tr_img_list = sorted(tl.files.load_file_list(path=config.TRAIN.tr_img_path, regx='.*.png', printable=False))
# train_lr_img_list = sorted(tl.files.load_file_list(path=config.TRAIN.lr_img_path, regx='.*.png', printable=False))
# valid_hr_img_list = sorted(tl.files.load_file_list(path=config.VALID.hr_img_path, regx='.*.png', printable=False))
# valid_lr_img_list = sorted(tl.files.load_file_list(path=config.VALID.lr_img_path, regx='.*.png', printable=False))
## If your machine have enough memory, please pre-load the whole train set.
train_hr_imgs = tl.vis.read_images(train_hr_img_list, path=config.TRAIN.hr_img_path, n_threads=32)
train_tr_imgs = tl.vis.read_images(train_tr_img_list, path=config.TRAIN.tr_img_path, n_threads=32)
###========================== DEFINE MODEL ============================###
## train inference
t_image = tf.placeholder('float32', [batch_size, 512, 512, 3], name='t_image_input_to_SRGAN_generator')
t_target_image = tf.placeholder('float32', [batch_size, 512, 512, 3], name='t_target_image')
net_g = SRGAN_g(t_image, is_train=True, reuse=False)
net_d, logits_real = SRGAN_d(t_target_image, is_train=True, reuse=False)
_, logits_fake = SRGAN_d(net_g.outputs, is_train=True, reuse=True)
net_g.print_params(False)
net_g.print_layers()
net_d.print_params(False)
net_d.print_layers()
## vgg inference. 0, 1, 2, 3 BILINEAR NEAREST BICUBIC AREA
t_target_image_224 = tf.image.resize_images(t_target_image, size=[224, 224], method=0,align_corners=False) # resize_target_image_for_vgg # http://tensorlayer.readthedocs.io/en/latest/_modules/tensorlayer/layers.html#UpSampling2dLayer
t_predict_image_224 = tf.image.resize_images(net_g.outputs, size=[224, 224], method=0, align_corners=False) # resize_generate_image_for_vgg
net_vgg, vgg_target_emb = Vgg19_simple_api((t_target_image_224 + 1) / 2, reuse=False)
_, vgg_predict_emb = Vgg19_simple_api((t_predict_image_224 + 1) / 2, reuse=True)
## test inference
net_g_test = SRGAN_g(t_image, is_train=False, reuse=True)
# ###========================== DEFINE TRAIN OPS ==========================###
d_loss1 = tl.cost.sigmoid_cross_entropy(logits_real, tf.ones_like(logits_real), name='d1')
d_loss2 = tl.cost.sigmoid_cross_entropy(logits_fake, tf.zeros_like(logits_fake), name='d2')
d_loss = d_loss1 + d_loss2
g_gan_loss = 1e-3 * tl.cost.sigmoid_cross_entropy(logits_fake, tf.ones_like(logits_fake), name='g')
mse_loss = tl.cost.mean_squared_error(net_g.outputs, t_target_image, is_mean=True)
vgg_loss = 2e-6 * tl.cost.mean_squared_error(vgg_predict_emb.outputs, vgg_target_emb.outputs, is_mean=True)
g_loss = mse_loss + vgg_loss + g_gan_loss
g_vars = tl.layers.get_variables_with_name('SRGAN_g', True, True)
d_vars = tl.layers.get_variables_with_name('SRGAN_d', True, True)
with tf.variable_scope('learning_rate'):
lr_v = tf.Variable(lr_init, trainable=False)
## Pretrain
g_optim_init = tf.train.AdamOptimizer(lr_v, beta1=beta1).minimize(mse_loss, var_list=g_vars)
## SRGAN
g_optim = tf.train.AdamOptimizer(lr_v, beta1=beta1).minimize(g_loss, var_list=g_vars)
d_optim = tf.train.AdamOptimizer(lr_v, beta1=beta1).minimize(d_loss, var_list=d_vars)
###========================== RESTORE MODEL =============================###
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=False))
tl.layers.initialize_global_variables(sess)
###============================= LOAD VGG ===============================###
vgg19_npy_path = "vgg19.npy"
if not os.path.isfile(vgg19_npy_path):
print("Please download vgg19.npz from : https://github.com/machrisaa/tensorflow-vgg")
exit()
npz = np.load(vgg19_npy_path, encoding='latin1').item()
params = []
for val in sorted(npz.items()):
W = np.asarray(val[1][0])
b = np.asarray(val[1][1])
print(" Loading %s: %s, %s" % (val[0], W.shape, b.shape))
params.extend([W, b])
tl.files.assign_params(sess, params, net_vgg)
###============================= TRAINING ===============================###
## use first `batch_size` of train set to have a quick test during training
sample_imgs = train_hr_imgs[0:batch_size]
sample_imgs_tr = train_tr_imgs[0:batch_size]
# sample_imgs = tl.vis.read_images(train_hr_img_list[0:batch_size], path=config.TRAIN.hr_img_path, n_threads=32) # if no pre-load train set
sample_imgs_384 = tl.prepro.threading_data(sample_imgs, fn=crop_sub_imgs_fn, is_random=False)
# print('sample HR sub-image:', sample_imgs_384.shape, sample_imgs_384.min(), sample_imgs_384.max())
sample_imgs_tr_384 = tl.prepro.threading_data(sample_imgs_tr, fn=crop_sub_imgs_fn, is_random=False)
# sample_imgs_96 = tl.prepro.threading_data(sample_imgs_384, fn=downsample_fn)
# print('sample LR sub-image:', sample_imgs_96.shape, sample_imgs_96.min(), sample_imgs_96.max())
# tl.vis.save_images(sample_imgs_96, [ni, ni], save_dir_ginit + '/_train_sample_96.png')
tl.vis.save_images(sample_imgs_384, [ni, ni], save_dir_ginit + '/_train_sample_384.png')
tl.vis.save_images(sample_imgs_tr_384, [ni, ni], save_dir_ginit + '/_train_sample_tr_384.png')
# tl.vis.save_images(sample_imgs_96, [ni, ni], save_dir_gan + '/_train_sample_96.png')
tl.vis.save_images(sample_imgs_384, [ni, ni], save_dir_gan + '/_train_sample_384.png')
tl.vis.save_images(sample_imgs_tr_384, [ni, ni], save_dir_gan + '/_train_sample_tr_384.png')
###========================= initialize G ====================###
## fixed learning rate
saver1 = tf.train.Saver(var_list=tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='SRGAN_g'))
sess.run(tf.assign(lr_v, lr_init))
print(" ** fixed learning rate: %f (for init G)" % lr_init)
for epoch in range(0, n_epoch_init + 1):
epoch_time = time.time()
total_mse_loss, n_iter = 0, 0
## If your machine have enough memory, please pre-load the whole train set.
for idx in range(0, len(train_hr_img_list), batch_size):
step_time = time.time()
filename = train_hr_img_list[idx]
print(filename)
b_imgs_384 = tl.prepro.threading_data(train_hr_imgs[idx:idx + batch_size], fn=crop_sub_imgs_fn, is_random=True)
b_imgs_tr_384 = tl.prepro.threading_data(train_tr_imgs[idx:idx + batch_size], fn=crop_sub_imgs_fn, is_random=True)
# b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)
## update G
errM, _ = sess.run([mse_loss, g_optim_init], {t_image: b_imgs_384, t_target_image: b_imgs_tr_384})
print("Epoch [%2d/%2d] %4d time: %4.4fs, mse: %.8f " % (epoch, n_epoch_init, n_iter, time.time() - step_time, errM))
total_mse_loss += errM
n_iter += 1
log = "[*] Epoch: [%2d/%2d] time: %4.4fs, mse: %.8f" % (epoch, n_epoch_init, time.time() - epoch_time, total_mse_loss / n_iter)
print(log)
## quick evaluation on train set
if (epoch != 0) and (epoch % 2 == 0):
out = sess.run(net_g_test.outputs, {t_image: sample_imgs_384}) #; print('gen sub-image:', out.shape, out.min(), out.max())
print("[*] save images")
tl.vis.save_images(out, [ni, ni], save_dir_ginit + '/train_%d.png' % epoch)
## save model
if (epoch != 0) and (epoch % 2 == 0):
tl.files.save_npz(net_g.all_params, name=checkpoint_dir + '/g_{}_init.npz'.format(tl.global_flag['mode']), sess=sess)
###========================= train GAN (SRGAN) =========================###
for epoch in range(0, n_epoch + 1):
## update learning rate
if epoch != 0 and (epoch % decay_every == 0):
new_lr_decay = lr_decay**(epoch // decay_every)
sess.run(tf.assign(lr_v, lr_init * new_lr_decay))
log = " ** new learning rate: %f (for GAN)" % (lr_init * new_lr_decay)
print(log)
elif epoch == 0:
sess.run(tf.assign(lr_v, lr_init))
log = " ** init lr: %f decay_every_init: %d, lr_decay: %f (for GAN)" % (lr_init, decay_every, lr_decay)
print(log)
epoch_time = time.time()
total_d_loss, total_g_loss, n_iter = 0, 0, 0
## If your machine have enough memory, please pre-load the whole train set.
for idx in range(0, len(train_hr_imgs), batch_size):
step_time = time.time()
b_imgs_384 = tl.prepro.threading_data(train_hr_imgs[idx:idx + batch_size], fn=crop_sub_imgs_fn, is_random=True)
b_imgs_tr_384 = tl.prepro.threading_data(train_tr_imgs[idx:idx + batch_size], fn=crop_sub_imgs_fn, is_random=True)
# b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)
## update D
errD, _ = sess.run([d_loss, d_optim], {t_image: b_imgs_384, t_target_image: b_imgs_tr_384})
## update G
errG, errM, errV, errA, _ = sess.run([g_loss, mse_loss, vgg_loss, g_gan_loss, g_optim], {t_image: b_imgs_384, t_target_image: b_imgs_tr_384})
print("Epoch [%2d/%2d] %4d time: %4.4fs, d_loss: %.8f g_loss: %.8f (mse: %.6f vgg: %.6f adv: %.6f)" %
(epoch, n_epoch, n_iter, time.time() - step_time, errD, errG, errM, errV, errA))
total_d_loss += errD
total_g_loss += errG
n_iter += 1
log = "[*] Epoch: [%2d/%2d] time: %4.4fs, d_loss: %.8f g_loss: %.8f" % (epoch, n_epoch, time.time() - epoch_time, total_d_loss / n_iter,
total_g_loss / n_iter)
print(log)
## quick evaluation on train set
if (epoch != 0) and (epoch % 2 == 0):
out = sess.run(net_g_test.outputs, {t_image: sample_imgs_384}) #; print('gen sub-image:', out.shape, out.min(), out.max())
print("[*] save images")
tl.vis.save_images(out, [ni, ni], save_dir_gan + '/train_%d.png' % epoch)
## save model
if (epoch != 0) and (epoch % 2 == 0):
tl.files.save_npz(net_g.all_params, name=checkpoint_dir + '/g_{}.npz'.format(tl.global_flag['mode']), sess=sess)
tl.files.save_npz(net_d.all_params, name=checkpoint_dir + '/d_{}.npz'.format(tl.global_flag['mode']), sess=sess)
saver1.save(sess, checkpoint_dir + '/SRGAN_X4')
def evaluate():
print('here 1')
## create folders to save result images
save_dir = "samples/{}".format(tl.global_flag['mode'])
tl.files.exists_or_mkdir(save_dir)
checkpoint_dir = "checkpoint"
###====================== PRE-LOAD DATA ===========================###
#valid_hr_img_list = sorted(tl.files.load_file_list(path=config.VALID.hr_img_path, regx='.*.png', printable=False))
#valid_lr_img_list = sorted(tl.files.load_file_list(path=config.VALID.lr_img_path, regx='.*.png', printable=False))
train_hr_img_list = sorted(tl.files.load_file_list(path=config.TRAIN.hr_img_path, regx='.*.png', printable=False))
train_tr_img_list = sorted(tl.files.load_file_list(path=config.TRAIN.tr_img_path, regx='.*.png', printable=False))
train_tr_imgs = tl.vis.read_images(train_tr_img_list, path=config.TRAIN.tr_img_path, n_threads=32)
train_hr_imgs = tl.vis.read_images(train_hr_img_list, path=config.TRAIN.hr_img_path, n_threads=32)
###========================== DEFINE MODEL ============================###
for imid in range(len(train_hr_img_list)): # 0: 企鹅 81: 蝴蝶 53: 鸟 64: 古堡
filename = train_hr_img_list[imid]
splitfilename = filename.split()
splitfilename = splitfilename[0]
train_hr_img = train_hr_imgs[imid]
# valid_lr_img = get_imgs_fn('test.png', 'data2017/') # if you want to test your own image
train_hr_img = (train_hr_img / 127.5) - 1 # rescale to ［－1, 1]
# print(valid_lr_img.min(), valid_lr_img.max())
size = train_hr_img.shape
# t_image = tf.placeholder('float32', [None, size[0], size[1], size[2]], name='input_image') # the old version of TL need to specify the image size
t_image = tf.placeholder('float32', [1, None, None, 3], name='input_image')
net_g = SRGAN_g(t_image, is_train=False, reuse=tf.AUTO_REUSE)
###========================== RESTORE G =============================###
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=False))
tl.layers.initialize_global_variables(sess)
tl.files.load_and_assign_npz(sess=sess, name=checkpoint_dir + '/g_srgan.npz', network=net_g)
###======================= EVALUATION =============================###
start_time = time.time()
out = sess.run(net_g.outputs, {t_image: [train_hr_img]})
print("took: %4.4fs" % (time.time() - start_time))
print("LR size: %s / generated HR size: %s" % (size, out.shape)) # LR size: (339, 510, 3) / gen HR size: (1, 1356, 2040, 3)
print("[*] save images")
path_fname = os.path.join(save_dir, splitfilename)
# tl.vis.save_image(out[0], save_dir + '/valid_gen_%d.png' % imid)
tl.vis.save_image(out[0],path_fname)
#tl.vis.save_image(train_lr_img, save_dir + '/valid_lr.png')
#tl.vis.save_image(train_hr_img, save_dir + '/valid_hr.png')
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--mode', type=str, default='srgan', help='srgan, evaluate')
args = parser.parse_args()
tl.global_flag['mode'] = args.mode
if tl.global_flag['mode'] == 'srgan':
train()
elif tl.global_flag['mode'] == 'evaluate':
evaluate()
else:
raise Exception("Unknow --mode")