-
Notifications
You must be signed in to change notification settings - Fork 1
/
utils.py
210 lines (188 loc) · 8.13 KB
/
utils.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
import tensorflow as tf
import tensorflow.contrib.slim as slim
import numpy as np
def train_model(loss, global_step, data_num,batch_size, args):
print ('train_model ...')
lr_factor = 0.1
lr_epoch = args.lr_epoch.strip().split(',')
lr_epoch = list(map(int, lr_epoch))
boundaries = [epoch*data_num//batch_size for epoch in lr_epoch]
lr_values = [args.learning_rate*(lr_factor**x) for x in range(0, len(lr_epoch)+1)]
for lr in lr_values:
print('lr: ',lr)
lr_op = tf.train.piecewise_constant(global_step, boundaries, lr_values)
optimizer = tf.train.AdamOptimizer(lr_op)
all_vars = tf.trainable_variables()
var_to_train = all_vars
if len(args.to_train_prefix) > 0 and False:
var_to_train = []
for var in all_vars:
for prefix in args.to_train_prefix:
if prefix == var.name[:len(prefix)]:
var_to_train.append(var)
print("vars to train:")
for var in var_to_train:
print (var.name)
# exit()
# train_op = slim.learning.create_train_op(loss, optimizer, global_step=global_step,variables_to_train=var_to_train)
train_op = slim.learning.create_train_op(loss, optimizer, global_step=global_step)
return train_op, lr_op
def compute_gradients(loss, global_step, data_num,batch_size, args,reuse_optimizer = False):
print ('train_model ...')
lr_factor = 0.1
lr_epoch = args.lr_epoch.strip().split(',')
lr_epoch = list(map(int, lr_epoch))
boundaries = [epoch*data_num//batch_size for epoch in lr_epoch]
lr_values = [args.learning_rate*(lr_factor**x) for x in range(0, len(lr_epoch)+1)]
for lr in lr_values:
print('lr: ',lr)
lr_op = tf.train.piecewise_constant(global_step, boundaries, lr_values)
optimizer = tf.train.AdamOptimizer(lr_op)
for v in tf.trainable_variables():
if 'Adam' in v.name:
print (v.name)
# all_vars = tf.trainable_variables()
# var_to_train = all_vars
# if len(args.to_train_prefix) > 0 and False:
# var_to_train = []
# for var in all_vars:
# for prefix in args.to_train_prefix:
# if prefix == var.name[:len(prefix)]:
# var_to_train.append(var)
# print("vars to train:")
# for var in var_to_train:
# print (var.name)
# exit()
# train_op = slim.learning.create_train_op(loss, optimizer, global_step=global_step,variables_to_train=var_to_train)
# train_op = slim.learning.create_train_op(loss, optimizer, global_step=global_step)
gradients = optimizer.compute_gradients(loss)
return gradients, lr_op, optimizer
def GaussianMaps(sigma):
d = int(3 * sigma + 0.5)
cx, cy = (d, d)
Pixels = np.zeros((2 * d, 2 * d, 2), dtype=np.int32)
value = np.zeros((2 * d, 2 * d, 1), dtype=np.float32)
for x in range(2 * d):
for y in range(2 * d):
D = (x - cx) ** 2 + (y - cy) ** 2
Pixels[y,x] = (x - cx, y - cy)
if D < (3 * sigma) ** 2:
value[y,x] = np.exp(-D / 2.0 / sigma / sigma)
else:
value[y,x] = 0.0
return Pixels, value, (2*d,2*d)
def LandmarkImage(Landmarks, image_size, sigma=None):
if sigma is None:
sigma = tf.to_float(tf.reduce_max(image_size[1:3]))/4
d = tf.to_int32(3 * sigma + 0.5)
xx = tf.tile(tf.expand_dims(tf.range(-d, d, 1), 0), (2 * d, 1))
yy = tf.tile(tf.expand_dims(tf.range(-d, d, 1), 1), (1, 2 * d))
Pixels = tf.concat([tf.expand_dims(yy,-1), tf.expand_dims(xx,-1)], axis=-1)
D = tf.reduce_sum(tf.square(tf.to_float(Pixels)), axis=-1)
zeros = tf.zeros((2 * d, 2 * d), dtype=tf.float32)
Gaussian = tf.exp(-D/(2*sigma*sigma))
values = tf.where(tf.greater(D, (3 * sigma) ** 2), zeros, Gaussian)
shape = tf.to_float(tf.expand_dims(image_size[1:3],axis=0))
def Do(L):
def DoIn(Point):
intPoint = tf.to_int32(Point)
locations = Pixels + intPoint
img = tf.scatter_nd(locations, values, shape=(image_size[1], image_size[2]))
return img
L = tf.reverse(tf.reshape(L, [-1, 2]), [-1])*shape
L = tf.map_fn(DoIn, L)
L = tf.reshape(tf.reduce_max(L, axis=0), (image_size[1], image_size[2]))
return L
Landmarks = tf.clip_by_value(Landmarks, 0, 1)
return tf.map_fn(Do, Landmarks)
def LandmarkImage_98(Landmarks, image_size, sigma=None):
if sigma is None:
sigma = tf.to_float(tf.reduce_max(image_size[1:3]))/4
d = tf.to_int32(3 * sigma + 0.5)
xx = tf.tile(tf.expand_dims(tf.range(-d, d, 1), 0), (2 * d, 1))
yy = tf.tile(tf.expand_dims(tf.range(-d, d, 1), 1), (1, 2 * d))
Pixels = tf.concat([tf.expand_dims(yy,-1), tf.expand_dims(xx,-1)], axis=-1)
D = tf.reduce_sum(tf.square(tf.to_float(Pixels)), axis=-1)
zeros = tf.zeros((2 * d, 2 * d), dtype=tf.float32)
Gaussian = tf.exp(-D/(2*sigma*sigma))
values = tf.where(tf.greater(D, (3 * sigma) ** 2), zeros, Gaussian)
shape = tf.to_float(tf.expand_dims(image_size[1:3],axis=0))
#print('debug 0 ::','shape value :',shape) #(1,2)
def Do(L):
def DoIn(Point):
intPoint = tf.to_int32(Point)
locations = Pixels + intPoint
#print('debug 2','intpoint shape:{} Pixels shape{} locations shape{}'.format(intPoint.shape,Pixels.shape,locations.shape))
#(2,) (,,2) (,,2)
img = tf.scatter_nd(locations, values, shape=(image_size[1], image_size[2]))
#print('debug 3','img shape:',img.shape)
#(56,56)
return img
L = tf.reverse(tf.reshape(L, [-1, 2]), [-1])*shape
#print('debug 5',L.shape)
#(98,2)
L = tf.map_fn(DoIn, L)
# L = tf.reshape(tf.reduce_max(L, axis=0), (image_size[1], image_size[2]))
L = tf.transpose(L,[1,2,0])
#print('debug 4:','L shape:',L.shape)
#(56,56)
return L
Landmarks = tf.clip_by_value(Landmarks, 0, 1)
#print('debug 1 ::','Landmarks shape',Landmarks.shape) #(n,196)
return tf.map_fn(Do, Landmarks)
if __name__ == "__main__":
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import cv2
landmarks_placeholder = tf.placeholder(tf.float32, shape=(None, 18), name='landmarks')
landmarks = np.asarray([[0, 0, 0, 5, 0, 10,\
5, 0, 5, 5, 5, 10,\
10,0, 10,5, 10,10]], dtype=np.float32)/16
img = LandmarkImage(landmarks_placeholder, (1,16,16,1))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
out = sess.run(img,feed_dict={landmarks_placeholder:landmarks})
# print(out)
print(out.shape)
# for img in out:
# plt.xticks([])
# plt.yticks([])
# plt.imshow(img,cmap=plt.cm.hot)
# plt.show()
# plt.clf()
image=out[0]*255
for i in range(image.shape[2]):
print(image[:,:,i].shape)
# cv2.imwrite('./result{}.png'.format(i),image[:,:,i])
def avg_grads(tower_grads):
average_grads = []
for grad_and_vars in zip(*tower_grads):
print(grad_and_vars)
# Note that each grad_and_vars looks like the following:
# ((grad0_gpu0, var0_gpu0), ... , (grad0_gpuN, var0_gpuN))
grads = []
# g_is_None = False
print(grad_and_vars)
for g, _ in grad_and_vars:
print(g)
if g is None:
# g_is_None = True
continue
# Add 0 dimension to the gradients to represent the tower.
expanded_g = tf.expand_dims(g, 0)
# Append on a 'tower' dimension which we will average over below.
grads.append(expanded_g)
if len(grads) == 0:
continue
# Average over the 'tower' dimension.
grad = tf.concat(grads, 0)
grad = tf.reduce_mean(grad, 0)
# Keep in mind that the Variables are redundant because they are shared
# across towers. So .. we will just return the first tower's pointer to
# the Variable.
v = grad_and_vars[0][1]
grad_and_var = (grad, v)
average_grads.append(grad_and_var)
# print(average_grads)
return average_grads