-
Notifications
You must be signed in to change notification settings - Fork 2
/
deep_resnet_perceptual.py
284 lines (229 loc) · 8.07 KB
/
deep_resnet_perceptual.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
# -*- coding: utf-8 -*-
from keras.layers import *
from keras.models import *
from keras.optimizers import *
from keras.callbacks import *
import keras.backend as K
import tensorflow as tf
import matplotlib.pyplot as plt
from data_gen import DataGen
import numpy as np
from sklearn.model_selection import train_test_split
from keras.applications.vgg16 import VGG16
#db_path = r'G:\mix5k.h5'
db_path = r'E:/images_uint8.h5'
train_split = .75
ishape = (16,16)
pshape = (32,32)
strides = (16,16)
bsize = 64
num_epochs=30
reload_data = False
if 'dg' not in globals() or input('Reload data? [Y/N] : ').lower()=='y':
reload_data=True
dg = DataGen(db_path, pshape=pshape, strides=strides)
with dg.load_db() as db:
all_data=dg.get_datasets()
train,test=train_test_split(all_data, train_size=train_split,
shuffle=True, random_state=7)
train_patches = dg.get_patch_list(train)
test_patches = dg.get_patch_list(test, pshape=(64,64), strides=(32,32))
print('Loading Images ....')
dg.load_images()
train_gen = dg.patch_gen(train_patches, bsize)
test_gen = dg.patch_gen(test_patches, bsize)
len_train = len(train_patches)//bsize
len_test = len(test_patches)//bsize
def prop_sigmoid(x):
over = tf.stop_gradient(tf.maximum(x-1., 0.))
under = tf.stop_gradient(tf.maximum(0.-x, 0.))
return x-over+under
K.clear_session()
vgg16 = VGG16(include_top=False, input_shape=(64,64, 3), pooling='avg')
vgg_lrs, vgg_cfg, vgg_wts = zip(*((lr.__class__, lr.get_config(), lr.get_weights()) for lr in vgg16.layers[1:]))
del vgg16
K.clear_session()
in_t = Input(shape=(None,None,3), dtype='int8')
in_tx1 = Lambda(lambda v: tf.cast(v,tf.float32)/255.)(in_t)
x = in_tx1
x = Conv2D(64, (3,3), padding='same', activation=None)(x)
x = Activation('relu')(x)
y = Conv2D(128, (1,1))(x)
x = Concatenate()([x,in_tx1])
x = Conv2D(128, (3,3), padding='same', activation=None)(x)
x = Add()([x,y])
x = Activation('relu')(x)
y = UpSampling2D((2,2))(x)
x = Concatenate()([x,in_tx1])
x = Conv2DTranspose(128, (4,4), padding='same', strides=(2,2),
activation=None)(x)
x = Add()([x,y])
x = Activation('relu')(x)
in_tx2= Lambda(lambda v: tf.image.resize_bicubic(v, tf.shape(v)[1:3]*2))(in_tx1)
y = x
x = Concatenate()([x,in_tx2])
x = Conv2D(128, (3,3), padding='same', activation=None)(x)
y = Add()([x,y])
x = Activation('relu')(x)
y = UpSampling2D((2,2))(x)
x = Concatenate()([x,in_tx2])
x = Conv2DTranspose(128, (4,4), padding='same', strides=(2,2),
activation=None)(x)
x = Add()([x,y])
x = Activation('relu')(x)
in_tx4 = Lambda(lambda v: tf.image.resize_bicubic(v, tf.shape(v)[1:3]*4))(in_tx1)
y = x
x = Concatenate()([x,in_tx4])
x = Conv2D(128, (3,3), padding='same', activation=None)(x)
y = Add()([x,y])
x = Activation('relu')(x)
y = Conv2D(64, (1,1))(x)
x = Concatenate()([x,in_tx4])
x = Conv2D(64, (3,3), padding='same', activation=None)(x)
y = Add()([x,y])
x = Activation('relu')(x)
x = Concatenate()([x,in_tx4])
x = Conv2D(3, (3,3), padding='same', activation=None)(x)
x = Add()([x, in_tx4])
x = Activation(prop_sigmoid, name='out')(x)
out_t = x
vgg_in = Input((None, None, 3))
vgg_outs = []
x = vgg_in
for k, lr, cg, wt in zip(range(8),vgg_lrs,vgg_cfg, vgg_wts):
if 'name' in cg:
del cg['name']
mlr = lr.from_config(cg)
x = mlr(x)
mlr.set_weights(wt)
mlr.trainable=False
vgg_outs.append(x)
vgg_model = Model(vgg_in, vgg_outs)
vgg_model.trainable = False
def tv_loss(y_true, y_pred):
area = tf.cast(tf.reduce_prod(tf.shape(y_pred)[1:3]), tf.float32)
return 0.1*tf.reduce_mean(tf.image.total_variation(y_pred)/area)
def mse_loss(y_true, y_pred):
#print(y_true.dtype)
return tf.reduce_mean((y_true/255.-y_pred)**2)
def percep_loss(y_true, y_pred):
#print(y_true.dtype)
assert y_true.dtype == tf.float32
y = y_true/255.
x = y_pred
yo = vgg_model(y)
xo = vgg_model(x)
return (#0.01*tf.reduce_mean((yo[4]-xo[4])**2)
# +0.01*tf.reduce_mean((yo[3]-xo[3])**2)
+0.01*tf.reduce_mean((yo[7]-xo[7])**2)
# +0.001*tf.reduce_mean((yo[0]-xo[0])**2)
# +tf.reduce_mean((y-x)**2)
) + tv_loss(y_true,y_pred)
model = Model(in_t, out_t)
print(model.summary())
opt = Adam(lr=1e-4)
model.compile(opt, percep_loss, metrics=[mse_loss, tv_loss])
def on_ep_end(epoch, log):
plt.figure(figsize=(15,5))
X, Y = next(test_gen)
Y_pred = model.predict_on_batch(X)
for k in range(15):
c, r = k//5, k%5
dp = r*3+c
sp = r*11+c*4
plt.subplot(5, 11, 1 + sp)
plt.axis('off')
plt.imshow(X[dp])
if not r: plt.title('LR')
plt.subplot(5, 11, 2 + sp)
plt.axis('off')
plt.imshow(np.clip(Y_pred[dp],0.,1.))
if not r: plt.title('HRPred')
plt.subplot(5, 11, 3 + sp)
plt.axis('off')
plt.imshow(Y[dp])
if not r: plt.title('HR')
plt.pause(.01)
figs=plt.get_fignums()
if len(figs)>15:
for ff in figs[0:len(figs)-15]: plt.close(ff)
ep_end = LambdaCallback(on_epoch_end=on_ep_end)
mchk = ModelCheckpoint('deep_resnet_checkpoint.h5',save_best_only=True)
redlr = ReduceLROnPlateau(factor=.75, patience=3, verbose=2, cooldown=2, min_lr=1e-5)
#model.load_weights('htv_training_highly_perceptual_deep_resnet_incremental_v1_32x32_linear_downsampling_mix5k_validationdata.h5')
model.fit_generator(train_gen, steps_per_epoch=len_train//10, epochs=num_epochs,
validation_data=test_gen, validation_steps=len_test//10,
callbacks=[ep_end, mchk])
model.save('htv_training_highly_perceptual_deep_resnet_incremental_v1_32x32_linear_downsampling_mix5k_validationdata.h5')
#import cv2
#img=dg.image_dict[test[3]]
#imgl=cv2.resize(img, None, fx=.25, fy=.25)
#imgs=model.predict_on_batch(imgl[None,...])[0]
#
#plt.figure(figsize=[7,7]), plt.axis('off')
#plt.imshow(imgl)
#plt.title('LO-RES')
#plt.tight_layout()
#
#plt.figure(figsize=[20,7])
#
#plt.subplot(131),plt.axis('off')
#plt.imshow(cv2.resize(imgl, None, fx=4, fy=4, interpolation=cv2.INTER_CUBIC))
#plt.title('4xBI-CUBIC')
#
#plt.subplot(132),plt.axis('off')
#plt.imshow(imgs)
#plt.title('4xSUPERRESOLUTION')
#
#plt.subplot(133),plt.axis('off')
#plt.imshow(img)
#plt.title('ORIGINAL HIRES')
#plt.tight_layout()
#
#
import cv2
img=dg.image_dict[test[5]]
imgl=cv2.resize(img, None, fx=.25, fy=.25)
imgs=model.predict_on_batch(imgl[None,...])[0]
imgr=resize(img, imgs.shape).astype('float32')
plt.figure(figsize=[7,7]), plt.axis('off')
plt.imshow(imgl)
plt.title('LO-RES')
plt.tight_layout()
plt.figure(figsize=[20,7])
plt.subplot(131),plt.axis('off')
imgc=np.clip(cv2.resize(imgl/255., None, fx=4, fy=4, interpolation=cv2.INTER_CUBIC), 0.,1.)
plt.imshow(imgc)
plt.title('4xBI-CUBIC')
plt.subplot(132),plt.axis('off')
plt.imshow(imgs)
plt.title('4xSUPERRESOLUTION')
plt.subplot(133),plt.axis('off')
plt.imshow(imgr)
plt.title('ORIGINAL HIRES')
plt.tight_layout()
num_img = len(test)
bic_psnr_arr, bic_ssim_arr = [], [];
sup_psnr_arr, sup_ssim_arr = [], [];
for k in range(num_img):
if not k%10:
print('{}/{}'.format(k,num_img))
img=dg.image_dict[test[k]]
if img.shape[0]*img.shape[1]>400*500:
continue
imgl=cv2.resize(img, None, fx=.25, fy=.25)
imgs=model.predict_on_batch(imgl[None,...])[0]
imgr=resize(img, imgs.shape).astype('float32')
imgc=np.clip(cv2.resize(imgl/255., None, fx=4, fy=4, interpolation=cv2.INTER_CUBIC), 0.,1.).astype('float32')
bic_psnr_arr.append(compare_psnr(imgr, imgc))
sup_psnr_arr.append(compare_psnr(imgr, imgs))
bic_ssim_arr.append(compare_ssim(imgr, imgc, multichannel=True, data_range=1.0, gaussian_weights=True))
sup_ssim_arr.append(compare_ssim(imgr, imgs, multichannel=True, data_range=1.0, gaussian_weights=True))
sup_avg_psnr=np.mean(sup_psnr_arr)
sup_avg_ssim = np.mean(sup_ssim_arr)
bic_avg_psnr = np.mean(bic_psnr_arr)
bic_avg_ssim = np.mean(bic_ssim_arr)
print('Bi-cubic PSNR:', bic_avg_psnr)
print('Superresolution PSNR:', sup_avg_psnr)
print('Bi-cubic SSIM:', bic_avg_ssim)
print('Superresolution SSIM:', sup_avg_ssim)