forked from hiwonjoon/tf-frrn
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
240 lines (181 loc) · 8.23 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
from __future__ import print_function
from six.moves import xrange
import better_exceptions
import tensorflow as tf
import numpy as np
from ops import *
def _arch_type_a(num_classes):
def _ru(t,conv3_1,bn_1,conv3_2,bn_2):
_t = conv3_1(t)
_t = bn_1(_t)
_t = tf.nn.relu(_t)
_t = conv3_2(_t)
_t = bn_2(_t)
return t + _t
def _frru(y_z,conv3_1,bn_1,conv3_2,bn_2,conv1,scale):
#tf.nn.max_pool(, ksize, strides, padding, data_format='NHWC', name=None)
y,z = y_z
# first concat the the two coming stream (using max pooling for the residual stream
# reducing its size )
_t = tf.concat([y,tf.nn.max_pool(z,[1,scale,scale,1],[1,scale,scale,1],'SAME','NHWC')],axis=3)
# do two unints CONV
_t = conv3_1(_t)
_t = bn_1(_t)
_t = tf.nn.relu(_t)
_t = conv3_2(_t)
_t = bn_2(_t)
y_prime = tf.nn.relu(_t) # y_prime is the next pooling stream input in the network
# y_prime also used as the bias for the computed residual
_t = conv1(y_prime)
# conv1 + bias upscale using an unpooling layers
_t = tf.image.resize_nearest_neighbor(_t, tf.shape(y_prime)[1:3]*scale)
z_prime = _t + z
return y_prime,z_prime
def _divide_stream(t,conv1):
z = conv1(t)
return t,z
def _concat_stream(y_z,conv1):
y,z = y_z
# only unpooling y-----
t = tf.concat([tf.image.resize_bilinear(y, tf.shape(y)[1:3]*2), z],axis=3)
return conv1(t)
from functools import partial
# The First Conv
spec = [Conv2d('conv2d_1',3,48,5,5,1,1,data_format='NHWC'), BatchNorm('conv2d_1_bn',48,axis=3),
lambda t,**kwargs : tf.nn.relu(t)]
# RU Layers
# first using 3 residual units to form the RU layers
for i in range(3):
spec.append(
partial(_ru,
conv3_1 = Conv2d('ru48_%d_1'%i,48,48,3,3,1,1,data_format='NHWC'),
bn_1 = BatchNorm('ru48_%d_1_bn'%i,48,axis=3),
conv3_2=Conv2d('ru48_%d_2'%i,48,48,3,3,1,1,data_format='NHWC'),
bn_2 = BatchNorm('ru48_%d_2_bn'%i,48,axis=3))
)
# Split Streams
spec.append(
partial(_divide_stream,
conv1 = Conv2d('conv32',48,32,1,1,1,1,data_format='NHWC'))
)
# FFRU Layers (Encoding)
prev_ch = 48
# it: num of frru units
# ch: the number of the channels in every unit
# scale:
# prev_ch + 32 means two input for the frru units
#
for it,ch,scale in [(3,96,2),(4,192,4),(2,384,8),(2,384,16)] :
# this layer is used for max pooling before frru units before frru it + 32
spec.append(
lambda y_z : (tf.nn.max_pool(y_z[0],[1,2,2,1],[1,2,2,1],'SAME','NHWC'),y_z[1])
# maxpooling y only.
)
# 3,4,2,2 are the units number
for i in range(it):
spec.append(
partial(_frru,
conv3_1=Conv2d('encode_frru%d_%d_%d_1'%(ch,scale,i),prev_ch+32,ch,3,3,1,1,data_format='NHWC'),
bn_1 = BatchNorm('encode_frru%d_%d_%d_1_bn'%(ch,scale,i),ch,axis=3),
conv3_2=Conv2d('encode_frru%d_%d_%d_2'%(ch,scale,i),ch,ch,3,3,1,1,data_format='NHWC'),
bn_2 = BatchNorm('encode_frru%d_%d_%d_2_bn'%(ch,scale,i),ch,axis=3),
conv1 = Conv2d('encode_frru%d_%d_%d_3'%(ch,scale,i),ch,32,1,1,1,1,data_format='NHWC'),
scale=scale)
)
prev_ch = ch
# FRRU Layers (Decoding)
for it,ch,scale in [(2,192,8),(2,192,4),(2,96,2)] :
# this layer useing bilinear interpolation is for the unpooling layer
spec.append(
lambda y_z : (tf.image.resize_bilinear(y_z[0], tf.shape(y_z[0])[1:3]*2), y_z[1])
)
# frru decoding part
for i in range(it):
spec.append(
partial(_frru,
conv3_1=Conv2d('decode_frru%d_%d_%d_1'%(ch,scale,i),prev_ch+32,ch,3,3,1,1,data_format='NHWC'),
bn_1 = BatchNorm('decode_frru%d_%d_%d_1_bn'%(ch,scale,i),ch,axis=3),
conv3_2=Conv2d('decode_frru%d_%d_%d_2'%(ch,scale,i),ch,ch,3,3,1,1,data_format='NHWC'),
bn_2 = BatchNorm('decode_frru%d_%d_%d_2_bn'%(ch,scale,i),ch,axis=3),
conv1 = Conv2d('decode_frru%d_%d_%d_3'%(ch,scale,i),ch,32,1,1,1,1,data_format='NHWC'),
scale=scale)
)
prev_ch = ch
# Concat Streams
spec.append(
partial(_concat_stream,
conv1 = Conv2d('conv48',prev_ch+32,48,1,1,1,1,data_format='NHWC')))
# RU Layers
for i in range(3,6):
spec.append(
partial(_ru,
conv3_1=Conv2d('ru48_%d_1'%i,48,48,3,3,1,1,data_format='NHWC'),
bn_1 = BatchNorm('ru48_%d_1_bn'%i,48,axis=3),
conv3_2=Conv2d('ru48_%d_2'%i,48,48,3,3,1,1,data_format='NHWC'),
bn_2 = BatchNorm('ru48_%d_2_bn'%i,48,axis=3))
)
# Final Classification Layer
spec.append(
Conv2d('conv_c',48,num_classes,1,1,1,1,data_format='NHWC'))
return spec
class FRRN():
def __init__(self,lr,global_step,K,
im,gt,arch_fn,
param_scope,is_training=False):
with tf.variable_scope(param_scope):
net_spec = arch_fn()
with tf.variable_scope('forward'):
_t = im
for block in net_spec:
print(_t)
_t = block(_t)
self.logits = _t
#Tensor("train/forward/BiasAdd_66:0", shape=(?, 256, 512, 20), dtype=float32)
self.preds = tf.argmax(self.logits,axis=3)
#Tensor("train/forward/ArgMax:0", shape=(?, 256, 512), dtype=int64)
# Loss
naive_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=self.logits,labels=gt)
#Tensor("train/forward/SparseSoftmaxCrossEntropyWithLogits/Reshape_2:0", shape=(?, 256, 512), dtype=float32)
# TODO: ignore pixels labed as void? is it requried?
# mask = tf.logical_not(tf.equal(gt,0))
# naive_loss = naive_loss * mask
boot_loss,_ = tf.nn.top_k(tf.reshape(naive_loss,[tf.shape(im)[0],tf.shape(im)[1]*tf.shape(im)[2]]),k=K,sorted=False)
self.loss = tf.reduce_mean(tf.reduce_sum(boot_loss,axis=1))
if( is_training ):
with tf.variable_scope('backward'):
optimizer = tf.train.AdamOptimizer(lr)
self.train_op= optimizer.minimize(self.loss,global_step=global_step)
# save_vars
save_vars = {('train/'+'/'.join(var.name.split('/')[1:])).split(':')[0] : var for var in
tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,param_scope.name) }
#for name,var in save_vars.items():
# print(name,var)
self.saver = tf.train.Saver(var_list=save_vars,max_to_keep = 3)
def save(self,sess,dir,step=None):
if(step is not None):
self.saver.save(sess,dir+'/model.ckpt',global_step=step)
else :
self.saver.save(sess,dir+'/last.ckpt')
def load(self,sess,model):
self.saver.restore(sess,model)
if __name__ == "__main__":
with tf.variable_scope('params') as params:
pass
im = tf.placeholder(tf.float32,[None,256,512,3])
gt = tf.placeholder(tf.int32,[None,256,512]) #19 + unlabeled area(plus ignored labels)
global_step = tf.Variable(0, trainable=False)
from functools import partial
net = FRRN(0.1,global_step,512*64,im,gt,partial(_arch_type_a,20),params,True)
print(net.logits)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.graph.finalize()
sess.run(init_op)
for _ in range(30):
_t,preds,_ = (sess.run([net.logits,net.preds,net.train_op],
feed_dict={im:np.random.random((1,256,512,3)),
gt:np.zeros((1,256,512))}))
print(preds.shape)