/
ops.py
357 lines (283 loc) · 11.5 KB
/
ops.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
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import tensorflow as tf
import tensorflow.contrib.slim as slim
import numpy as np
def lrelu(x, leak=0.2):
return tf.maximum(x, leak*x)
def conv2d(x, o_dim, data_format='NHWC', name=None, k=4, s=2, act=None):
return slim.conv2d(x, o_dim, k, stride=s, activation_fn=act, scope=name, data_format=data_format)
def deconv2d(x, o_dim, data_format='NHWC', name=None, k=4, s=2, act=None):
return slim.conv2d_transpose(x, o_dim, k, stride=s, activation_fn=act, scope=name, data_format=data_format)
def linear(x, o_dim, name=None, act=None):
return slim.fully_connected(x, o_dim, activation_fn=act, scope=name)
def batch_norm(x, train, data_format='NHWC', name=None, act=lrelu, epsilon=1e-5, momentum=0.9):
return slim.batch_norm(x,
decay=momentum,
updates_collections=None,
epsilon=epsilon,
scale=True,
fused=True,
is_training=train,
activation_fn=act,
data_format=data_format,
scope=name)
def inst_norm(x, train, data_format='NHWC', name=None, affine=False, act=lrelu, epsilon=1e-5):
with tf.variable_scope(name, default_name='Inst', reuse=None) as vs:
if x.get_shape().ndims == 4 and data_format == 'NCHW':
x = nchw_to_nhwc(x)
if x.get_shape().ndims == 4:
mean_dim = [1,2]
else: # 2
mean_dim = [1]
mu, sigma_sq = tf.nn.moments(x, mean_dim, keep_dims=True)
inv = tf.rsqrt(sigma_sq+epsilon)
normalized = (x-mu)*inv
if affine:
var_shape = [x.get_shape()[-1]]
shift = slim.model_variable('shift', shape=var_shape, initializer=tf.zeros_initializer)
scale = slim.model_variable('scale', shape=var_shape, initializer=tf.ones_initializer)
out = scale*normalized + shift
else:
out = normalized
if x.get_shape().ndims == 4 and data_format == 'NCHW':
out = nhwc_to_nchw(out)
if act is None: return out
else: return act(out)
def resize_nearest_neighbor(x, new_size, data_format):
if data_format == 'NCHW':
x = nchw_to_nhwc(x)
x = tf.image.resize_nearest_neighbor(x, new_size)
x = nhwc_to_nchw(x)
else:
x = tf.image.resize_nearest_neighbor(x, new_size)
return x
def upscale(x, scale, data_format):
_, h, w, _ = get_conv_shape(x, data_format)
return resize_nearest_neighbor(x, (h*scale, w*scale), data_format)
def var_on_cpu(name, shape, initializer, dtype=tf.float32):
return slim.model_variable(name, shape, dtype=dtype, initializer=initializer, device='/CPU:0')
def int_shape(tensor):
shape = tensor.get_shape().as_list()
return [num if num is not None else -1 for num in shape]
def get_conv_shape(tensor, data_format):
shape = int_shape(tensor)
# always return [N, H, W, C]
if data_format == 'NCHW':
return [shape[0], shape[2], shape[3], shape[1]]
elif data_format == 'NHWC':
return shape
def nchw_to_nhwc(x):
return tf.transpose(x, [0, 2, 3, 1])
def nhwc_to_nchw(x):
return tf.transpose(x, [0, 3, 1, 2])
def next(loader):
return loader.next()[0].data.numpy()
def to_nhwc(image, data_format):
if data_format == 'NCHW':
new_image = nchw_to_nhwc(image)
else:
new_image = image
return new_image
def to_nchw_numpy(image):
if image.shape[3] in [1,2,3]:
new_image = image.transpose([0, 3, 1, 2])
else:
new_image = image
return new_image
def to_nhwc_numpy(image):
if image.shape[1] in [1,2,3]:
new_image = image.transpose([0, 2, 3, 1])
else:
new_image = image
return new_image
def add_channels(x, num_ch=1, data_format='NHWC'):
b, h, w, c = get_conv_shape(x, data_format)
if data_format == 'NCHW':
x = tf.concat([x, tf.zeros([b, num_ch, h, w])], axis=1)
else:
x = tf.concat([x, tf.zeros([b, h, w, num_ch])], axis=-1)
return x
def remove_channels(x, data_format='NHWC'):
b, h, w, c = get_conv_shape(x, data_format)
if data_format == 'NCHW':
x, _ = tf.split(x, [3, -1], axis=1)
else:
x, _ = tf.split(x, [3, -1], axis=3)
return x
def denorm_img(norm, data_format):
_, _, _, c = get_conv_shape(norm, data_format)
if c == 2:
norm = add_channels(norm, num_ch=1, data_format=data_format)
elif c > 3:
norm = remove_channels(norm, data_format=data_format)
return tf.clip_by_value(to_nhwc(norm*255, data_format), 0, 255)
def reshape(x, h, w, c, data_format):
if data_format == 'NCHW':
x = tf.reshape(x, [-1, c, h, w])
else:
x = tf.reshape(x, [-1, h, w, c])
return x
def show_all_variables():
model_vars = tf.trainable_variables()
slim.model_analyzer.analyze_vars(model_vars, print_info=True)
# https://stackoverflow.com/questions/39051451/ssim-ms-ssim-for-tensorflow
# https://github.com/tensorflow/models/blob/master/compression/image_encoder/msssim.py
def fspecial_gauss(size, sigma, channels):
"""
Function to mimic the 'fspecial' gaussian MATLAB function
"""
radius = size // 2
offset = 0.0
start, stop = -radius, radius + 1
if size % 2 == 0:
offset = 0.5
stop -= 1
x, y = np.mgrid[offset + start:stop, offset + start:stop]
assert len(x) == size
x = x.reshape(x.shape+(1,1))
x = np.repeat(x, channels, axis=2)
x = np.repeat(x, channels, axis=3)
y = y.reshape(y.shape+(1,1))
y = np.repeat(y, channels, axis=2)
y = np.repeat(y, channels, axis=3)
x = tf.constant(x, dtype=tf.float32)
y = tf.constant(y, dtype=tf.float32)
g = tf.exp(-((x**2 + y**2)/(2.0*sigma**2)))
return g / tf.reduce_sum(g)
def ssim(img1, img2, mean_metric=True,
filter_size=11, filter_sigma=1.5, k1=0.01, k2=0.03,
min_val=-1.0, max_val=1.0):
# input should be rescaled to [-1,1]
img_shape = img1.get_shape()
height = img_shape[1].value
width = img_shape[2].value
channels = img_shape[3].value
# print(img_shape)
# Filter size can't be larger than height or width of images.
size = min(filter_size, height, width)
# print(size)
# Scale down sigma if a smaller filter size is used.
sigma = filter_sigma * size / filter_size if filter_size else 0
# print(sigma)
# ! normalize image to [0,1]
img1 = (img1 - min_val) / (max_val - min_val)
img2 = (img2 - min_val) / (max_val - min_val)
if filter_size:
window = fspecial_gauss(size, sigma, channels) # window shape [size, size]
mu1 = tf.nn.conv2d(img1, window, strides=[1,1,1,1], padding='VALID')
mu2 = tf.nn.conv2d(img2, window, strides=[1,1,1,1], padding='VALID')
sigma11 = tf.nn.conv2d(img1*img1, window, strides=[1,1,1,1], padding='VALID')
sigma22 = tf.nn.conv2d(img2*img2, window, strides=[1,1,1,1], padding='VALID')
sigma12 = tf.nn.conv2d(img1*img2, window, strides=[1,1,1,1], padding='VALID')
else:
mu1 = img1, mu2 = img2
sigma11 = img1*img1
sigma22 = img2*img2
sigma12 = img1*img2
mu11 = mu1*mu1
mu22 = mu2*mu2
mu12 = mu1*mu2
sigma11 -= mu11
sigma22 -= mu22
sigma12 -= mu12
L = 1.0 # max scale, already normalized to 1
c1 = (k1*L)**2
c2 = (k2*L)**2
v1 = 2.0 * sigma12 + c2
v2 = sigma11 + sigma22 + c2
value = ((2.0 * mu12 + c1) * v1) / ((mu11 + mu22 + c1) * v2)
if mean_metric: return tf.reduce_mean(value)
result = {'ssim_map': value, 'cs_map': v1/v2, 'g': window}
return result
def ms_ssim(img1, img2, mean_metric=True, min_val=-1.0, max_val=1.0):
# input should be rescaled to [-1,1]
weight = [0.0448, 0.2856, 0.3001, 0.2363, 0.1333]
mssim = []
mcs = []
for w in weight:
result = ssim(img1, img2, mean_metric=False, min_val=min_val, max_val=max_val)
mssim.append(tf.reduce_mean(result['ssim_map']))
mcs.append(tf.reduce_mean(result['cs_map']))
filtered_im1 = tf.nn.avg_pool(img1, [1,2,2,1], [1,2,2,1], padding='SAME')
filtered_im2 = tf.nn.avg_pool(img2, [1,2,2,1], [1,2,2,1], padding='SAME')
img1 = filtered_im1
img2 = filtered_im2
# ! doesn't work
# filter_sigmas = [0.5, 1, 2, 4, 8]
# cs_map0 = None
# for i, filter_sigma in enumerate(filter_sigmas):
# result = ssim(img1, img2, filter_sigma=filter_sigma,
# min_val=min_val, mean_metric=False)
# if i == 0: cs_map0 = result['cs_map']
# mssim.append(tf.reduce_mean(result['ssim_map']))
# mcs.append(tf.reduce_mean(tf.nn.conv2d(cs_map0, result['g'], strides=[1,1,1,1], padding='VALID')))
# list to tensor of dim D+1
mssim = tf.stack(mssim, axis=0)
mcs = tf.stack(mcs, axis=0)
level = len(weight)
value = (tf.reduce_prod(mcs[0:level-1]**weight[0:level-1])*
(mssim[level-1]**weight[level-1]))
if mean_metric: value = tf.reduce_mean(value)
return value
def main(_):
from skimage import data, transform, img_as_float
import matplotlib.pyplot as plt
color = False
if color:
image = data.astronaut()
else: # [h,w] -> [h,w,1]
image = data.camera()
image = np.expand_dims(image, axis=-1)
# image = transform.resize(image, output_shape=[128, 128])
img = img_as_float(image)
print(img.shape)
rows, cols, channels = img.shape
noise = np.ones_like(img) * 0.2 * (img.max() - img.min())
noise[np.random.random(size=noise.shape) > 0.5] *= -1
img_noise = img + noise
img_noise = np.clip(img_noise, a_min=0, a_max=1)
plt.figure()
plt.subplot(121)
if color:
plt.imshow(img)
plt.subplot(122)
plt.imshow(img_noise)
else:
plt.imshow(img[:,:,0], cmap='gray')
plt.subplot(122)
plt.imshow(img_noise[:,:,0], cmap='gray')
plt.show()
## TF CALC START
image1 = tf.placeholder(tf.float32, shape=[rows, cols, channels])
image2 = tf.placeholder(tf.float32, shape=[rows, cols, channels])
def image_to_4d(image):
image = tf.expand_dims(image, 0)
return image
image4d_1 = image_to_4d(image1)
image4d_2 = image_to_4d(image2)
print(img.min(), img.max(), img_noise.min(), img_noise.max())
ssim_index = ssim(image4d_1, image4d_2) #, min_val=0.0, max_val=1.0)
msssim_index = ms_ssim(image4d_1, image4d_2) #, min_val=0.0, max_val=1.0)
# img *= 255
# img_noise *= 255
# ssim_index = ssim(image4d_1, image4d_2, min_val=0.0, max_val=255.0)
# msssim_index = ms_ssim(image4d_1, image4d_2, min_val=0.0, max_val=255.0)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
tf_ssim_none = sess.run(ssim_index,
feed_dict={image1: img, image2: img})
tf_ssim_noise = sess.run(ssim_index,
feed_dict={image1: img, image2: img_noise})
tf_msssim_none = sess.run(msssim_index,
feed_dict={image1: img, image2: img})
tf_msssim_noise = sess.run(msssim_index,
feed_dict={image1: img, image2: img_noise})
###TF CALC END
print('tf_ssim_none', tf_ssim_none)
print('tf_ssim_noise', tf_ssim_noise)
print('tf_msssim_none', tf_msssim_none)
print('tf_msssim_noise', tf_msssim_noise)
if __name__ == '__main__':
tf.app.run()