-
Notifications
You must be signed in to change notification settings - Fork 20
/
infer.py
215 lines (177 loc) · 7.99 KB
/
infer.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
import glob
import os
import time
import argparse
import numpy as np
from PIL import Image
from libtiff import TIFF
import tensorflow as tf
import cv2
import keras
from keras.optimizers import Adam
from keras.models import Model
import keras.backend as K
from src.model import coarse_generator,fine_generator
global g_local_model
global g_global_model
def normalize_pred(img,mask):
img = np.reshape(img,[1,128,128,3])
mask = np.reshape(mask,[1,128,128,1])
img_coarse = tf.image.resize(img, (64,64), method=tf.image.ResizeMethod.LANCZOS3)
img_coarse = (img_coarse - 127.5) / 127.5
img_coarse = np.array(img_coarse)
mask_coarse = tf.image.resize(mask, (64,64), method=tf.image.ResizeMethod.LANCZOS3)
mask_coarse = (mask_coarse - 127.5) / 127.5
mask_coarse = np.array(mask_coarse)
start_time = time.time()
X_fakeB_coarse,x_global = g_global_model.predict([img_coarse,mask_coarse])
X_fakeB_coarse = (X_fakeB_coarse + 1) /2.0
X_fakeB_coarse = cv2.normalize(X_fakeB_coarse, None, alpha = 0, beta = 1, norm_type = cv2.NORM_MINMAX, dtype = cv2.CV_32F)
X_fakeB_coarse = tf.image.resize(X_fakeB_coarse, (128,128), method=tf.image.ResizeMethod.LANCZOS3)
pred_img_coarse = X_fakeB_coarse[:,:,:,0]
img = (img - 127.5) / 127.5
mask = (mask - 127.5) / 127.5
X_fakeB = g_local_model.predict([img,mask,x_global])
end_time = time.time()
print(end_time-start_time)
X_fakeB = (X_fakeB + 1) /2.0
X_fakeB = cv2.normalize(X_fakeB, None, alpha = 0, beta = 1, norm_type = cv2.NORM_MINMAX, dtype = cv2.CV_32F)
pred_img = X_fakeB[:,:,:,0]
return np.asarray(pred_img,dtype=np.float32),np.asarray(pred_img_coarse,dtype=np.float32)
def strided_crop(img, mask, img_h,img_w,height, width,stride=1):
full_prob = np.zeros((img_h, img_w),dtype=np.float32)
full_prob_coarse = np.zeros((img_h, img_w),dtype=np.float32)
full_sum = np.ones((img_h, img_w),dtype=np.float32)
full_sum_coarse = np.ones((img_h, img_w),dtype=np.float32)
max_x = int(((img.shape[0]-height)/stride)+1)
max_y = int(((img.shape[1]-width)/stride)+1)
max_crops = (max_x)*(max_y)
i = 0
for h in range(max_x):
for w in range(max_y):
crop_img_arr = img[h * stride:(h * stride) + height,w * stride:(w * stride) + width]
crop_mask_arr = mask[h * stride:(h * stride) + height,w * stride:(w * stride) + width]
pred,pred_coarse = normalize_pred(crop_img_arr,crop_mask_arr)
full_prob[h * stride:(h * stride) + height,w * stride:(w * stride) + width] += pred[0]
full_sum[h * stride:(h * stride) + height,w * stride:(w * stride) + width] += 1
full_prob_coarse[h * stride:(h * stride) + height,w * stride:(w * stride) + width] += pred_coarse[0]
full_sum_coarse[h * stride:(h * stride) + height,w * stride:(w * stride) + width] += 1
i = i + 1
out_img = full_prob / full_sum
out_img_coarse = full_prob_coarse / full_sum_coarse
return out_img, out_img_coarse
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--test_data', type=str, default='DRIVE', required=True, choices=['DRIVE','CHASE','STARE'])
parser.add_argument('--out_dir', type=str, default='pred', required=False)
parser.add_argument('--weight_name_global',type=str, help='path/to/global/weight/.h5 file', required=True)
parser.add_argument('--weight_name_local',type=str, help='path/to/local/weight/.h5 file', required=True)
parser.add_argument('--stride', type=int, default=3, help='For faster inference use stride 16/32, for better result use stride 3.')
parser.add_argument('--out_dir')
args = parser.parse_args()
## Input dimensions
image_shape_fine = (128,128,3)
mask_shape_fine = (128,128,1)
label_shape_fine = (128,128,1)
image_shape_x_coarse = (64,64,128)
image_shape_coarse = (64,64,3)
mask_shape_coarse = (64,64,1)
label_shape_coarse = (64,64,1)
img_shape_g = (64,64,3)
ndf=64
ncf=128
nff=128
## Load models
K.clear_session()
opt = Adam()
g_local_model = fine_generator(x_coarse_shape=image_shape_x_coarse,input_shape=image_shape_fine,mask_shape=mask_shape_fine,nff=nff)
g_local_model.load_weights(args.weight_name_local)
g_local_model.compile(loss='mse', optimizer=opt)
g_global_model = coarse_generator(img_shape=image_shape_coarse,mask_shape=mask_shape_coarse,ncf=ncf)
g_global_model.load_weights(args.weight_name_global)
g_global_model.compile(loss='mse',optimizer=opt)
## Create Output Directory
out_path = args.out_dir
directories = [out_path,out_path+'/Coarse',out_path+'/Fine']
for directory in directories:
if not os.path.exists(directory):
os.makedirs(directory)
## Find file numbers,paths or names
if args.test_data == 'DRIVE':
limit = 20
elif args.test_data == 'CHASE':
filenames = glob.glob("CHASE/test/images/*.jpg")
limit = len(filenames)
elif args.test_data == 'STARE':
arr = ["im0001","im0082","im0236","im0319"]
limit = 4
## Iterating for each image
for i in range(0,limit):
if args.test_data == 'DRIVE':
if i<9:
ind = str(0)
else:
ind = str("")
label_name = "DRIVE/test/1st_manual/"+ind+str(i+1)+"_manual1.gif"
label = Image.open(label_name)
label_arr = np.asarray(label,dtype=np.float32)
img_name = "DRIVE/test/images/"+ind+str(i+1)+"_test.tif"
tif = TIFF.open(img_name)
img_arr = tif.read_image(tif)
img_arr = np.asarray(img_arr,dtype=np.float32)
mask_name = "DRIVE/test/mask/"+ind+str(i+1)+"_test_mask.gif"
mask = Image.open(mask_name)
mask_arr = np.asarray(mask,dtype=np.float32)
elif args.test_data == 'CHASE':
k = filenames[i].split('/')
k = k[-1].split('.')[0]
label_name = "CHASE/test/labels/"+k+"_1stHO.png"
label = Image.open(label_name)
label_arr = np.asarray(label,dtype=np.float32)
img_name = "CHASE/test/images/"+k+".jpg"
img = Image.open(img_name)
img_arr = np.asarray(img,dtype=np.float32)
mask_name = "CHASE/test/mask/"+k+"_mask.png"
mask = Image.open(mask_name)
mask_arr = np.asarray(mask,dtype=np.float32)
elif args.test_data == 'STARE':
label_name = "STARE/test/labels-ah/"+arr[i]+".ah.ppm"
label = Image.open(label_name)
label_arr = np.asarray(label,dtype=np.float32)
img_name = "STARE/test/stare-original-images/"+arr[i]+".ppm"
img = Image.open(img_name)
img_arr = np.asarray(img,dtype=np.float32)
mask_name = "STARE/test/mask/"+arr[i]+"_mask.png"
mask = Image.open(mask_name)
mask_arr = np.asarray(mask,dtype=np.float32)
## Get the output predictions as array
## Stride =3 (best result), Stride = 32 (faster result).
out_img,out_img_coarse = strided_crop(img_arr, mask_arr, mask_arr.shape[0], mask_arr.shape[1], 128, 128,args.stride)
out_img[mask_arr==0] = 0
out_img[out_img>=0.5] = 1
out_img[out_img<0.5] = 0
save_im = out_img.astype(np.uint8)
save_im[save_im==1] = 255
save_im = Image.fromarray(save_im)
out_img_coarse[mask_arr==0] = 0
out_img_coarse[out_img_coarse>=0.5] = 1
out_img_coarse[out_img_coarse<0.5] = 0
save_im_coarse = out_img_coarse.astype(np.uint8)
save_im_coarse[save_im_coarse==1] = 255
save_im_coarse = Image.fromarray(save_im_coarse)
## Save files
if args.test_data == 'DRIVE':
pred_name = directories[2]+"/"+ind+str(i+1)+".png"
pred_name_coarse = directories[1]+"/"+ind+str(i+1)+"coarse.png"
save_im.save(pred_name)
save_im_coarse.save(pred_name_coarse)
elif args.test_data == 'CHASE':
pred_name = directories[2]+"/"+k+".png"
pred_name_coarse = directories[1]+"/"+k+"coarse.png"
save_im.save(pred_name)
save_im_coarse.save(pred_name_coarse)
elif args.test_data == 'STARE':
pred_name = directories[2]+"/"+arr[i]+".png"
pred_name_coarse = directories[1]+"/"+arr[i]+"coarse.png"
save_im.save(pred_name)
save_im_coarse.save(pred_name_coarse)