-
Notifications
You must be signed in to change notification settings - Fork 14
/
style_generate.py
167 lines (137 loc) · 7.42 KB
/
style_generate.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
import os
import numpy as np
import argparse
import cv2 as cv
from time import time
import keras.backend as K
from utils.img_process import generate_train, get_region_mask
from utils.vgg16_gray import VGG16Gray
from utils.compare_patch import compare_patch
def save_train_feat(photo_path, sketch_path, weight_path, feature_layers, save_path='./Data/train_sketch_feat.hdf5', img_size=(288, 288)):
photo, sketch, _ = generate_train(photo_path, sketch_path, img_size, photo2gray=True)
sketch = (sketch[:, np.newaxis, :, :] * 255.0).astype('float64')
vgg16 = VGG16Gray(weight_path=weight_path)
sketch_features = vgg16.get_features(sketch, feature_layers)
np.savez(save_path, *sketch_features)
def generate_target_style(photo, sketch, test_path, base_pool, feature_layers, vgg_weight, compare_size=48, searching_range=6, img_size=(288, 288)):
all_photo_pool, all_sketch_pool = (photo*255).astype('float64'), (sketch*255).astype('float64')
photo_img = cv.imread(test_path)
photo_img = cv.resize(photo_img, img_size).transpose(2, 0, 1).astype('float64')
border_feat_net = VGG16Gray(img_size=(144, 144), weight_path=vgg_weight)
max_find = 15
min_find = 2
total_imgs = all_photo_pool.shape[0]
n_grid_y = 18
n_grid_x = 18
x_step = photo_img.shape[2]/n_grid_x
y_step = photo_img.shape[1]/n_grid_y
compare_shift = (compare_size - x_step)/2
target_feats = [np.ndarray(x.shape[1:]) for x in base_pool]
symb_patch_3 = np.ndarray(shape=(1,3,compare_size,compare_size)).astype('float32')
conv_weights_3 = K.variable(symb_patch_3)
candidate_3 = K.placeholder(shape=(total_imgs,3,compare_size+2*searching_range,compare_size+2*searching_range))
conv_res = K.conv2d(candidate_3,conv_weights_3)
f_conv_3 = K.function([candidate_3],conv_res)
for jj in range(n_grid_y):
for ii in range(n_grid_x):
this_patch = photo_img[ :,
max(0, jj*y_step - compare_shift): min(n_grid_y*y_step, (jj+1)*y_step + compare_shift),
max(0, ii*x_step - compare_shift): min(n_grid_x*x_step, (ii+1)*x_step + compare_shift)]
this_patch = this_patch[np.newaxis, ...]
if ii<min_find or ii>max_find or jj<min_find or jj>max_find:
this_patch_rep = np.repeat(this_patch, total_imgs, 0)
candidate_patch = all_photo_pool[:,:,max(0, jj*y_step - compare_shift): min(n_grid_y*y_step, (jj+1)*y_step + compare_shift),
max(0,ii*x_step - compare_shift): min(n_grid_x*x_step, (ii+1)*x_step + compare_shift)]
diff = this_patch_rep - candidate_patch
sq_diff = np.square(diff)
sq_diff = np.reshape(sq_diff,(sq_diff.shape[0],sq_diff.size/sq_diff.shape[0]))
sum_sq_diff = np.sum(sq_diff,1)
match_idx = np.argmin(sum_sq_diff)
for i in range(len(target_feats)):
x_step_i = x_step / 2**i
y_step_i = y_step / 2**i
target_feats[i][:, jj*y_step_i: (jj+1)*y_step_i, ii*x_step_i:(ii+1)*x_step_i] = base_pool[i][match_idx, :,
jj*y_step_i:(jj+1)*y_step_i, ii*x_step_i:(ii+1)*x_step_i]
else:
candidate_patch = all_photo_pool[ :, :,
max(0, jj*y_step - compare_shift - searching_range): min(n_grid_y*y_step, (jj+1)*y_step + compare_shift + searching_range),
max(0, ii*x_step - compare_shift - searching_range): min(n_grid_x*x_step, (ii+1)*x_step + compare_shift + searching_range)]
diff_photo = compare_patch(this_patch, candidate_patch, x_step, searching_range, compare_size, f_conv_3, conv_weights_3)
total_diff = diff_photo
min_y = 0
max_y = 2*searching_range+1
min_x = 0
max_x = 2*searching_range+1
feat_jj = 4
feat_ii = 4
if jj<5:
feat_jj = jj
min_y = searching_range
if jj>12:
feat_jj = jj-9
max_y = searching_range
if ii<5:
feat_ii = ii
min_x = searching_range
if ii>12:
feat_ii = ii-9
max_x = searching_range
max_diff = np.max(total_diff)
total_diff[:, min_y:max_y, min_x:max_x] = total_diff[:, min_y:max_y, min_x:max_x] - max_diff
best_index = np.argmin(total_diff)
best_index = np.unravel_index(best_index, total_diff.shape)
best_patch = best_index[0]
y_shift = best_index[1] - searching_range
x_shift = best_index[2] - searching_range
start_y = jj*y_step + y_shift-16*feat_jj
start_x = ii*x_step + x_shift-16*feat_ii
target_sketch = all_sketch_pool[best_patch, :, start_y:start_y+144, start_x:start_x+144]
target_sketch = np.expand_dims(target_sketch, 1)
border_patch_feat = border_feat_net.get_features(target_sketch, feature_layers)
for i in range(len(target_feats)):
x_step_i = x_step / 2**i
y_step_i = y_step / 2**i
target_feats[i][:, jj*y_step_i: (jj+1)*y_step_i, ii*x_step_i: (ii+1)*x_step_i] = border_patch_feat[i][
:, :, feat_jj*y_step_i:(feat_jj+1) * y_step_i, feat_ii*x_step_i: (feat_ii+1) * x_step_i]
target_gram = []
for f in target_feats:
f = f.reshape((f.shape[0], f.shape[1]*f.shape[2]))
f_gram = np.dot(f, f.transpose())
target_gram += [f_gram]
# generate nose region gram
nose_mask_pool = get_region_mask()
nose_gram = []
for idx, f in enumerate(target_feats):
f = f * nose_mask_pool[idx]
f = f.reshape((f.shape[0], f.shape[1]*f.shape[2]))
f_gram = np.dot(f, f.transpose())
nose_gram += [f_gram]
return target_gram, nose_gram
if __name__ == '__main__':
photo_path = './Data/photos'
sketch_path = './Data/sketches'
vgg_weight_path = './Weight/vgg16_gray.hdf5'
train_feat_path = './Data/train_sketch_feat.npz'
test_path = './Data/test/1.png'
feature_layers = ['conv1_1', 'conv2_1', 'conv3_1', 'conv4_1', 'conv5_1']
if os.path.exists(train_feat_path):
print 'Train feature data base already exist'
else:
save_train_feat(photo_path, sketch_path, vgg_weight_path, feature_layers, save_path=train_feat_path)
feat = np.load(train_feat_path)
feat_base = [feat[x] for x in sorted(feat.files)]
# for idx, i in enumerate(feature_layers):
# tmp = np.load('../FM_train/sketch_feat%s.npy' % i)
# print tmp.shape, np.sum(tmp), np.sum(feat_base[idx])
# print np.linalg.norm(feat_base[idx] - tmp)
# exit()
# print [[x.shape, np.sum(x)] for x in feat_base]
photo, sketch, _ = generate_train(photo_path, sketch_path, size=(288, 288))
photo = photo.transpose(0, 3, 1, 2)
sketch = sketch[:, np.newaxis, :, :]
start = time()
target_gram, nose_gram = generate_target_style(photo, sketch, test_path, feat_base, feature_layers, vgg_weight_path, compare_size=48, searching_range=6, img_size=(288, 288))
end = time()
print 'Target style time', end - start
# print [x.shape for x in target_gram]
# print [x.shape for x in nose_gram]