-
Notifications
You must be signed in to change notification settings - Fork 27
/
model.py
277 lines (222 loc) · 8.56 KB
/
model.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
from __future__ import print_function
import tensorflow as tf
import numpy as np
from copy import deepcopy
import matplotlib.pyplot as plt
from IPython import display
import scipy.io as scio
import time
from datetime import datetime
from scipy.misc import imsave
import scipy.ndimage
from skimage import img_as_ubyte
from Net import Generator
from LOSS import SSIM_LOSS, Fro_LOSS #, L1_LOSS, Fro_LOSS
from collections import Iterable
from VGGnet.vgg16 import Vgg16
WEIGHT_INIT_STDDEV = 0.05
eps = 1e-8
class Model(object):
def __init__(self, BATCH_SIZE, INPUT_H, INPUT_W, is_training):
self.batchsize = BATCH_SIZE
self.G = Generator('Generator')
self.var_list = []
self.step = 0
if not hasattr(self, "ewc_loss"):
self.Add_loss = 0
self.SOURCE1 = tf.placeholder(tf.float32, shape = (BATCH_SIZE, INPUT_H, INPUT_W, 1), name = 'SOURCE1')
self.SOURCE2 = tf.placeholder(tf.float32, shape = (BATCH_SIZE, INPUT_H, INPUT_W, 1), name = 'SOURCE2')
self.c = tf.placeholder(tf.float32, shape = (), name = 'c')
print('source shape:', self.SOURCE1.shape)
self.generated_img = self.G.transform(I1 = self.SOURCE1, I2 = self.SOURCE2, is_training = is_training, reuse=False)
self.var_list.extend(tf.trainable_variables())
# for i in self.var_list:
# print(i.name)
if is_training:
''' SSIM loss'''
SSIM1 = 1 - SSIM_LOSS(self.SOURCE1, self.generated_img)
SSIM2 = 1 - SSIM_LOSS(self.SOURCE2, self.generated_img)
mse1 = Fro_LOSS(self.generated_img-self.SOURCE1)
mse2 = Fro_LOSS(self.generated_img-self.SOURCE2)
with tf.device('/gpu:1'):
self.S1_VGG_in = tf.image.resize_nearest_neighbor(self.SOURCE1, size = [224, 224])
self.S1_VGG_in = tf.concat((self.S1_VGG_in, self.S1_VGG_in, self.S1_VGG_in), axis = -1)
self.S2_VGG_in = tf.image.resize_nearest_neighbor(self.SOURCE2, size = [224, 224])
self.S2_VGG_in = tf.concat((self.S2_VGG_in, self.S2_VGG_in, self.S2_VGG_in), axis = -1)
vgg1 = Vgg16()
with tf.name_scope("vgg1"):
self.S1_FEAS = vgg1.build(self.S1_VGG_in)
vgg2 = Vgg16()
with tf.name_scope("vgg2"):
self.S2_FEAS = vgg2.build(self.S2_VGG_in)
for i in range(len(self.S1_FEAS)):
self.m1 = tf.reduce_mean(tf.square(features_grad(self.S1_FEAS[i])), axis = [1, 2, 3])
self.m2 = tf.reduce_mean(tf.square(features_grad(self.S2_FEAS[i])), axis = [1, 2, 3])
if i == 0:
self.ws1 = tf.expand_dims(self.m1, axis = -1)
self.ws2 = tf.expand_dims(self.m2, axis = -1)
else:
self.ws1 = tf.concat([self.ws1, tf.expand_dims(self.m1, axis = -1)], axis = -1)
self.ws2 = tf.concat([self.ws2, tf.expand_dims(self.m2, axis = -1)], axis = -1)
self.s1 = tf.reduce_mean(self.ws1, axis = -1) / self.c
self.s2 = tf.reduce_mean(self.ws2, axis = -1) / self.c
self.s = tf.nn.softmax(
tf.concat([tf.expand_dims(self.s1, axis = -1), tf.expand_dims(self.s2, axis = -1)], axis = -1))
self.ssim_loss = tf.reduce_mean(self.s[:, 0] * SSIM1 + self.s[:, 1] * SSIM2)
self.mse_loss = tf.reduce_mean(self.s[:, 0] * mse1 + self.s[:, 1] * mse2)
self.content_loss = self.ssim_loss + 20 * self.mse_loss
def compute_fisher(self, imgset, c, sess, num_samples = 200):
# computer Fisher information for each parameter
# initialize Fisher information for most recent task
self.F_accum = []
print("val_list length:", len(self.var_list))
# for i in self.var_list:
# print(i.name)
for v in range(len(self.var_list)):
self.F_accum.append(np.zeros(self.var_list[v].get_shape().as_list()))
start_time_cf = datetime.now()
for i in range(num_samples):
for k in range(len(imgset)):
# select random input image
set=imgset[k]
im_ind = np.random.randint(set.shape[0] - self.batchsize)
# compute first-order derivatives
s1_index = np.random.choice([0, 1], 1)
s1 = np.expand_dims(set[im_ind:im_ind + self.batchsize, :, :, s1_index[0]], -1)
s2 = np.expand_dims(set[im_ind:im_ind + self.batchsize, :, :, 1 - s1_index[0]], -1)
ders = sess.run(tf.gradients(-self.content_loss, self.var_list),
feed_dict = {self.SOURCE1: s1, self.SOURCE2: s2, self.c: c[k]})
# square the derivatives and add to total
for v in range(len(self.F_accum)):
self.F_accum[v] += np.square(ders[v])
elapsed_time_cf = datetime.now() - start_time_cf
print("compute fisher: %s/%s, elapsed_time: %s" % (i + 1, num_samples, elapsed_time_cf))
# divide totals by number of samples
for v in range(len(self.F_accum)):
self.F_accum[v] /= (num_samples*len(imgset))
# scio.savemat('F1.mat', {'F': self.F_accum})
# f = open("F.txt", "w")
# str = '\n'
# f.writelines(str.join(self.F_accum))
# f.close()
# F_save = flatten(self.F_accum)
# for v in range(len(F_save)):
# print(F_save,'\n')
def star(self):
# used for saving optimal weights after most recent task training
self.star_vars = []
for v in range(len(self.var_list)):
self.star_vars.append(self.var_list[v].eval())
def restore(self, sess):
# reassign optimal weights for latest task
if hasattr(self, "star_vars"):
for v in range(len(self.var_list)):
sess.run(self.var_list[v].assign(self.star_vars[v]))
def update_ewc_loss(self, lam):
# elastic weight consolidation
# lam is weighting for previous task(s) constraints
if not hasattr(self, "ewc_loss"):
self.ewc_loss = self.content_loss
for v in range(len(self.var_list)):
self.Add_loss += tf.reduce_sum(
tf.multiply(self.F_accum[v].astype(np.float32), tf.square(self.var_list[v] - self.star_vars[v])))
self.ewc_loss += (lam / 2) * self.Add_loss
# def EN(inputs):
# len = inputs.shape[0]
# entropies = [] # tf.Variable(tf.zeros(shape = (len, 1)))
# grey_level = 256
# counter = tf.Variable(tf.zeros(shape = (grey_level, 1), dtype = tf.int32))
# one = tf.constant(1)
#
# for i in range(len):
# input_uint8 = tf.cast(inputs[i, :, :, 0] * 255, dtype = tf.int32)
# input_uint8 = input_uint8 + 1
# W = inputs.shape[1]
# H = inputs.shape[2]
# for m in range(int(W)):
# for n in range(int(H)):
# indexx = input_uint8[m, n]
# # print("counter[indexx]", counter[indexx])
# counter[indexx] = tf.add(counter[indexx], one)
# counter = tf.cast(counter, tf.float32)
# total = tf.reduce_sum(counter)
# p = counter / total
# for k in range(grey_level):
# entropies.append(- p[k] * (tf.log(p[k] + eps) / tf.log(2.0)))
# return entropies
#
#
# def tf_EN(inputs):
# len = inputs.shape[0]
# W = int(inputs.shape[1])
# H = int(inputs.shape[2])
# total = W * H * 1.0
# grey_level = 256
#
# for i in range(len):
# input0 = tf.cast(inputs[i, :, :, 0] * 255, dtype = tf.int32)
# input1 = tf.cast(input0, dtype = tf.float32)
# input2 = input1 + 1
#
# for k in range(grey_level):
# if k == 0:
# p = tf.expand_dims(num(input2, k) / total, axis = 0)
# else:
# p = tf.concat([p, tf.expand_dims(num(input2, k) / total, axis = 0)], axis = 0)
# ep = - tf.multiply(p, tf.log(p + eps * eps) / tf.Variable(tf.log(2.0)))
#
# if i == 0:
# entropies = tf.expand_dims(tf.reduce_sum(ep), axis = 0)
# else:
# entropies = tf.concat([entropies, tf.expand_dims(tf.reduce_sum(ep), axis = 0)], axis = 0)
# return entropies
def num(input, k):
a = binary(input - tf.Variable(k + 0.5))
b = binary(tf.Variable(k + 1.5) - input)
return tf.reduce_sum(tf.multiply(a, b))
def grad(img):
kernel = tf.constant([[1 / 8, 1 / 8, 1 / 8], [1 / 8, -1, 1 / 8], [1 / 8, 1 / 8, 1 / 8]])
kernel = tf.expand_dims(kernel, axis = -1)
kernel = tf.expand_dims(kernel, axis = -1)
g = tf.nn.conv2d(img, kernel, strides = [1, 1, 1, 1], padding = 'SAME')
return g
def features_grad(features):
kernel = tf.constant([[1 / 8, 1 / 8, 1 / 8], [1 / 8, -1, 1 / 8], [1 / 8, 1 / 8, 1 / 8]])
kernel = tf.expand_dims(kernel, axis = -1)
kernel = tf.expand_dims(kernel, axis = -1)
_, _, _, c = features.shape
c = int(c)
for i in range(c):
fg = tf.nn.conv2d(tf.expand_dims(features[:, :, :, i], axis = -1), kernel, strides = [1, 1, 1, 1],
padding = 'SAME')
if i == 0:
fgs = fg
else:
fgs = tf.concat([fgs, fg], axis = -1)
return fgs
def flatten(input_list):
output_list = []
while True:
if input_list == []:
break
for index, i in enumerate(input_list):
if type(i) == list:
input_list = i + input_list[index + 1:]
break
else:
output_list.append(i)
input_list.pop(index)
break
return output_list
@tf.RegisterGradient("QuantizeGrad")
def sign_grad(op, grad):
input = op.inputs[0]
cond = (input >= -1) & (input <= 1)
zeros = tf.zeros_like(grad)
return tf.where(cond, grad, zeros)
def binary(input):
x = input
with tf.get_default_graph().gradient_override_map({"Sign": 'QuantizeGrad'}):
x = tf.sign(x)
x = x / 2 + 0.5
return x