/
test.py
234 lines (170 loc) · 8.56 KB
/
test.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
import model
import time
import imageio
from utils import *
class Test(object):
def __init__(self, model_path, save_path,kernel, scale, conf, method_num, num_of_adaptation):
methods=['direct', 'direct', 'bicubic', 'direct']
self.save_results=True
self.max_iters=num_of_adaptation
self.display_iter = 1
self.upscale_method= 'cubic'
self.noise_level = 0.0
self.back_projection=False
self.back_projection_iters=4
self.model_path=model_path
self.save_path=save_path
self.method_num=method_num
self.ds_method=methods[self.method_num]
self.kernel = kernel
self.scale=scale
self.scale_factors = [self.scale, self.scale]
self.build_network(conf)
def build_network(self, conf):
tf.reset_default_graph()
self.lr_decay = tf.placeholder(tf.float32, shape=[], name='learning_rate')
# Input image
self.input= tf.placeholder(tf.float32, shape=[None,None,None,3], name='input')
# Ground truth
self.label = tf.placeholder(tf.float32, shape=[None,None,None,3], name='label')
# parameter variables
self.PARAM=model.Weights(scope='MODEL')
# model class (without feedforward graph)
self.MODEL = model.MODEL(name='MODEL')
# Graph build
self.MODEL.forward(self.input,self.PARAM.weights)
self.output=self.MODEL.output
self.loss_t = tf.losses.absolute_difference(self.label, self.output)
# Optimizer
self.opt = tf.train.GradientDescentOptimizer(learning_rate=self.lr_decay).minimize(self.loss_t)
self.init = tf.global_variables_initializer()
# Variable lists
self.var_list= tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='MODEL')
self.loader=tf.train.Saver(var_list=self.var_list)
self.sess=tf.Session(config=conf)
def initialize(self):
self.sess.run(self.init)
self.loader.restore(self.sess, self.model_path)
print('=============== Load Meta-trained Model parameters... ==============')
self.loss = [None] * self.max_iters
self.mse, self.mse_rec, self.interp_mse, self.interp_rec_mse, self.mse_steps = [], [], [], [], []
self.psnr=[]
self.iter = 0
def __call__(self, img, gt, img_name):
self.img=img
self.gt = modcrop(gt, self.scale)
self.img_name=img_name
print('** Start Adaptation for X', self.scale, os.path.basename(self.img_name), ' **')
# Initialize network
self.initialize()
self.sf = np.array(self.scale_factors)
self.output_shape = np.uint(np.ceil(np.array(self.img.shape[0:2]) * self.scale))
# Train the network
self.quick_test()
print('[*] Baseline ')
self.train()
post_processed_output = self.final_test()
if self.save_results:
if not os.path.exists('%s/%02d' % (self.save_path, self.max_iters)):
os.makedirs('%s/%02d' % (self.save_path, self.max_iters))
imageio.imsave('%s/%02d/%s.png' % (self.save_path, self.max_iters, os.path.basename(self.img_name)[:-4]),
post_processed_output)
print('** Done Adaptation for X', self.scale, os.path.basename(self.img_name),', PSNR: %.4f' % self.psnr[-1], ' **')
print('')
return post_processed_output, self.psnr
def train(self):
self.hr_father = self.img
self.lr_son = imresize(self.img, scale=1/self.scale, kernel=self.kernel, ds_method=self.ds_method)
self.lr_son = np.clip(self.lr_son + np.random.randn(*self.lr_son.shape) * self.noise_level, 0., 1.)
t1=time.time()
for self.iter in range(self.max_iters):
if self.method_num == 0:
'''direct'''
if self.iter==0:
self.learning_rate=2e-2
elif self.iter < 4:
self.learning_rate=1e-2
else:
self.learning_rate=5e-3
elif self.method_num == 1:
'''Multi-scale'''
if self.iter < 3:
self.learning_rate=1e-2
else:
self.learning_rate=5e-3
elif self.method_num == 2:
'''bicubic'''
if self.iter == 0:
self.learning_rate = 0.01
elif self.iter < 3:
self.learning_rate = 0.01
else:
self.learning_rate = 0.001
elif self.method_num == 3:
''''scale 4'''
if self.iter ==0:
self.learning_rate=1e-2
elif self.iter < 5:
self.learning_rate=5e-3
else:
self.learning_rate=1e-3
self.train_output = self.forward_backward_pass(self.lr_son, self.hr_father)
# Display information
if self.iter % self.display_iter == 0:
print('Scale: ', self.scale, ', iteration: ', (self.iter+1), ', loss: ', self.loss[self.iter])
# Test network during adaptation
# if self.iter % self.display_iter == 0:
# output=self.quick_test()
# if self.iter==0:
# imageio.imsave('%s/%02d/01/%s.png' % (self.save_path, self.method_num, os.path.basename(self.img_name)[:-4]), output)
# if self.iter==9:
# imageio.imsave('%s/%02d/10/%s_%d.png' % (self.save_path, self.method_num, os.path.basename(self.img_name)[:-4], self.iter), output)
t2 = time.time()
print('%.2f seconds' % (t2 - t1))
def forward_pass(self, input, output_shape=None):
ILR = imresize(input, self.scale, output_shape, self.upscale_method)
feed_dict = {self.input : ILR[None,:,:,:]}
output_=self.sess.run(self.output, feed_dict)
return np.clip(np.squeeze(output_), 0., 1.)
def forward_backward_pass(self, input, hr_father):
ILR = imresize(input, self.scale, hr_father.shape, self.upscale_method)
HR = hr_father[None, :, :, :]
# Create feed dict
feed_dict = {self.input: ILR[None,:,:,:], self.label: HR, self.lr_decay: self.learning_rate}
# Run network
_, self.loss[self.iter], train_output = self.sess.run([self.opt, self.loss_t, self.output], feed_dict=feed_dict)
return np.clip(np.squeeze(train_output), 0., 1.)
def hr2lr(self, hr):
lr = imresize(hr, 1.0 / self.scale, kernel=self.kernel, ds_method=self.ds_method)
return np.clip(lr + np.random.randn(*lr.shape) * self.noise_level, 0., 1.)
def quick_test(self):
# 1. True MSE
self.sr = self.forward_pass(self.img, self.gt.shape)
self.mse = self.mse + [np.mean((self.gt - self.sr)**2)]
'''Shave'''
scale=int(self.scale)
PSNR=psnr(rgb2y(np.round(np.clip(self.gt*255., 0.,255.)).astype(np.uint8))[scale:-scale, scale:-scale],
rgb2y(np.round(np.clip(self.sr*255., 0., 255.)).astype(np.uint8))[scale:-scale, scale:-scale])
# PSNR=psnr(rgb2y(np.round(np.clip(self.gt*255., 0.,255.)).astype(np.uint8)), rgb2y(np.round(np.clip(self.sr*255., 0., 255.)).astype(np.uint8)))
self.psnr.append(PSNR)
# 2. Reconstruction MSE
self.reconstruct_output = self.forward_pass(self.hr2lr(self.img), self.img.shape)
self.mse_rec.append(np.mean((self.img - self.reconstruct_output)**2))
processed_output=np.round(np.clip(self.sr*255, 0., 255.)).astype(np.uint8)
print('iteration: ', self.iter, 'recon mse:', self.mse_rec[-1], ', true mse:', (self.mse[-1] if self.mse else None), ', PSNR: %.4f' % PSNR)
return processed_output
def final_test(self):
output = self.forward_pass(self.img, self.gt.shape)
if self.back_projection == True:
for bp_iter in range(self.back_projection_iters):
output = back_projection(output, self.img, down_kernel=self.kernel,
up_kernel=self.upscale_method, sf=self.scale, ds_method=self.ds_method)
processed_output=np.round(np.clip(output*255, 0., 255.)).astype(np.uint8)
'''Shave'''
scale=int(self.scale)
PSNR=psnr(rgb2y(np.round(np.clip(self.gt*255., 0.,255.)).astype(np.uint8))[scale:-scale, scale:-scale],
rgb2y(processed_output)[scale:-scale, scale:-scale])
# PSNR=psnr(rgb2y(np.round(np.clip(self.gt*255., 0.,255.)).astype(np.uint8)),
# rgb2y(processed_output))
self.psnr.append(PSNR)
return processed_output