def _contracting_block(self, block_num, _input, option='sum'):
layer = []
# 52*100*80*8 / 26*50*40*32 / 13*25*20*64
output = _conv3d(_input, kernel_size=3, stride=2, output_feature=self.output_channels[str(block_num)][0], name='conv_1')
layer.append(output)
output = _active_layer(output, name='active_layer_1', is_training=self.is_training)
# 52*100*80*16 / 26*50*40*32 / 13*25*20*64
output = _conv3d(output, kernel_size=3, stride=1, output_feature=self.output_channels[str(block_num)][1], name='conv_2')
if block_num == 1:
# if output channles equal, ues zero padding
big_zero = tf.zeros_like(layer[-1])
print(big_zero.get_shape())
layer[-1] = tf.concat([layer[-1],big_zero],axis=-1)
#output = self.feature_pyramid(output)
# print(output)
if option == 'sum':
output = tf.add(output, layer[-1], name='elemwise_sum')
else:
# 52*100*80*16 / 26*50*40*32 / 13*25*20*64
output = tf.concat(values=(output, layer[-1]), axis=4, name='concat')
output = _active_layer(output, name='active_layer_2', is_training=self.is_training)
#output = self.feature_pyramid(output)
layer.append(output)
return output
def convolution_block(layer_input, n_channels, num_convolutions):
x = layer_input
for i in range(num_convolutions - 1):
with tf.variable_scope('conv_' + str(i + 1)):
#x = convolution_3d(x, [5, 5, 5, n_channels, n_channels], [1, 1, 1, 1, 1])
x = _conv3d(x, output_feature=n_channels, kernel_size=5, stride=1)
#x = v_conv3d(x,kernel_shape=[5, 5, 5, n_channels, n_channels],stride=1)
x = prelu(x)
#x = convolution_3d(x, [5, 5, 5, n_channels, n_channels], [1, 1, 1, 1, 1])
#x = v_conv3d(x, kernel_shape=[5, 5, 5, n_channels, n_channels], stride=1)
x = _conv3d(x, output_feature=n_channels, kernel_size=5, stride=1)
x = x + layer_input
return prelu(x)
def cascade_feature(self, _input, block_num):
atrous_layer = []
out_feature = int(_input.get_shape()[-1])
output = _input
#output_1x1 = _conv3d(_input, kernel_size=1, stride=1, output_feature=out_feature, use_bias=True, name='pyramid_conv_1')
for i in range(1, self.atrou_num + 1):
dilate_rate = int(np.power(2, 4 - block_num) * i)
#print(dilate_rate)
# dilate rate : 24 16 8 / 12 8 4 / 6 4 2
output = atrous_bn_prelu(output,
kernel_size=3,
stride=1,
output_channels=out_feature,
dilation_rate=dilate_rate,
is_training=self.is_training,
name='atrous_conv%d' % i)
atrous_layer.append(output)
output = tf.concat([atrous_layer[0], atrous_layer[1], atrous_layer[2]],
axis=-1)
print('atrous conv shape:', output)
output = _conv3d(output,
kernel_size=1,
stride=1,
output_feature=out_feature,
use_bias=True,
name='pyramid_conv_1x1')
return output
def _expanding_block(self, block_num, _input, layer, option='concat'):
# 13*25*20*32 / 26*50*40*32 / 52*100*80*16
#output = conv_bn_prelu(_input, kernel_size=1, stride=1, output_channels=self.output_channels[str(block_num)][0],
# name='conv_1', is_training=self.is_training)
output = _conv3d(
_input,
kernel_size=1,
stride=1,
output_feature=self.output_channels[str(block_num)][0],
use_bias=True,
name='conv_1')
output = deconv_bn_prelu(
output,
output_channels=self.output_channels[str(block_num)][1],
is_training=self.is_training,
name='deconv')
# 26*50*40*32 / 52*100*80*16 / 104*200*160*8
# 26*50*40*64 / 52*100*80*32 / 104*200*160*16
output = tf.concat(values=(output, layer), axis=4, name='concat')
# 26*50*40*64 / 52*100*80*32 / 104*200*160*16
output = conv_bn_prelu(output,
kernel_size=3,
stride=1,
output_channels=int(output.get_shape()[-1]),
name='conv_2',
is_training=self.is_training)
output = conv_bn_prelu(output,
kernel_size=3,
stride=1,
output_channels=int(output.get_shape()[-1]),
name='conv_3',
is_training=self.is_training)
print(output)
return output
def feature_pyramid(self,_input):
atrous_layer = []
out_feature = int(_input.get_shape()[-1])
output_1x1 = _conv3d(_input, kernel_size=1, stride=1, output_feature=out_feature, use_bias=True, name='pyramid_conv_1')
for i in range(1, self.atrou_num + 1):
output = _conv3d(_input, kernel_size=3, stride=1, output_feature=out_feature,
dilation_rate=6*i,
name='atrous_conv%d' % i)
atrous_layer.append(output)
output = tf.concat([output_1x1, atrous_layer[0], atrous_layer[1], atrous_layer[2]], axis=-1)
print('atrous conv shape:', output)
output = _conv3d(output, kernel_size=1, stride=1, output_feature=out_feature,
use_bias=True, name='pyramid_conv_1x1')
return output
def _contracting_block(self, _input, option='concat'):
out_feature = _input.get_shape()[-1]
#print('out_feature', out_feature)
layer = []
atrous_layer = []
output = _conv3d(_input,
kernel_size=3,
stride=2,
output_feature=out_feature,
name='conv_1')
#print(output)
layer.append(output)
output = _active_layer(output,
name='active_layer_1',
is_training=self.is_training)
output = _conv3d(output,
kernel_size=3,
stride=1,
output_feature=out_feature,
name='conv_2')
#print(output)
if option == 'sum':
output = tf.add(output, layer[-1], name='elemwise_sum')
#print(output)
else:
output = tf.concat(values=(output, layer[-1]),
axis=4,
name='concat')
output = _active_layer(output,
name='active_layer_2',
is_training=self.is_training)
'''for i in range(self.atrou_num):
output = self._create_conv_layer(_input, atrous=2 ** i)
output = _conv3d(output, kernel_size=3, stride=1, output_feature=out_feature, dilation_rate=2 ** i,
name='atrous_conv%d'%i)
atrous_layer.append(output)'''
#output = tf.concat(atrous_layer,axis=-1)
print(output)
layer.append(output)
return output
def down_convolution(layer_input, in_channels):
with tf.variable_scope('down_convolution'):
#x = convolution_3d(layer_input, [2, 2, 2, in_channels, in_channels * 2], [1, 2, 2, 2, 1])
x = _conv3d(layer_input,
output_feature=in_channels * 2,
kernel_size=2,
stride=2)
#x = v_conv3d(layer_input, kernel_shape=[2, 2, 2, in_channels, in_channels * 2], stride=2)
return prelu(x)
def convolution_block_2(layer_input, fine_grained_features, n_channels,
num_convolutions):
x = tf.concat((layer_input, fine_grained_features), axis=-1)
with tf.variable_scope('conv_' + str(1)):
#x = convolution_3d(x, [5, 5, 5, n_channels * 2, n_channels], [1, 1, 1, 1, 1])
#x = v_conv3d(x,kernel_shape=[5, 5, 5, n_channels * 2, n_channels],stride=1)
x = _conv3d(x, output_feature=n_channels, kernel_size=5, stride=1)
for i in range(1, num_convolutions - 1):
with tf.variable_scope('conv_' + str(i + 1)):
#x = convolution_3d(x, [5, 5, 5, n_channels, n_channels], [1, 1, 1, 1, 1])
x = _conv3d(x, output_feature=n_channels, kernel_size=5, stride=1)
#x = v_conv3d(x, kernel_shape=[5, 5, 5, n_channels, n_channels], stride=1)
x = prelu(x)
#x = convolution_3d(x, [5, 5, 5, n_channels, n_channels], [1, 1, 1, 1, 1])
x = _conv3d(x, output_feature=n_channels, kernel_size=5, stride=1)
#x = v_conv3d(x, kernel_shape=[5, 5, 5, n_channels, n_channels], stride=1)
x = x + layer_input
return prelu(x)
def _expanding_block(self, _input, layer, option='concat'):
out_feature = int(_input.get_shape()[-1]) / 2
output = _conv3d(_input,
kernel_size=1,
stride=1,
output_feature=out_feature,
name='conv_1')
#print(output)
output = _active_layer(output,
name='active_layer_1',
is_training=self.is_training)
#output = _create_dconv_layer(output)
output = deconv3d(output,
output_channels=int(output.get_shape()[-1]),
name='deconv')
#$print(output)
output = _active_layer(output,
name='active_layer_2',
is_training=self.is_training)
if option == 'sum':
output = tf.add(output, layer, name='elemwise_sum')
#print(output)
else:
output = tf.concat(values=(output, layer), axis=4, name='concat')
output = _conv3d(output,
kernel_size=3,
stride=1,
output_feature=int(output.get_shape()[-1]),
name='conv_2')
#print(output)
output = _active_layer(output,
name='active_layer_3',
is_training=self.is_training)
return output
def _expanding_block(self, block_num, _input, layer, option='concat'):
# 13*25*20*32 / 26*50*40*32 / 52*100*80*16
output = _conv3d(_input, kernel_size=1, stride=1, output_feature=self.output_channels[str(block_num)][0], use_bias=True, name='conv_1')
print(output)
output = _active_layer(output, name='active_layer_1', is_training=self.is_training)
# 26*50*40*32 / 52*100*80*16 / 104*200*160*8
output = deconv3d(output, output_channels=self.output_channels[str(block_num)][1], name='deconv')
print(output)
output = _active_layer(output, name='active_layer_2', is_training=self.is_training)
if option == 'sum':
output = tf.add(output, layer, name='elemwise_sum')
# print(output)
else:
# 26*50*40*64 / 52*100*80*32 / 104*200*160*16
output = tf.concat(values=(output, layer), axis=4, name='concat')
# 26*50*40*64 / 52*100*80*32 / 104*200*160*16
output = _conv3d(output, kernel_size=3, stride=1,
output_feature=int(output.get_shape()[-1]), name='conv_2')
print(output)
output = _active_layer(output, name='active_layer_3', is_training=self.is_training)
return output
def _contracting_block(self, block_num, _input):
# xception contracte blcok
# 52*100*80*16 / 26*50*40*32 / 13*25*20*64
output = conv_bn_prelu(
_input,
kernel_size=3,
stride=1,
output_channels=self.output_channels[str(block_num)][0],
name='conv_1',
is_training=self.is_training)
# 52*100*80*8 / 26*50*40*32 / 13*25*20*64
# 52*100*80*16 / 26*50*40*32 / 13*25*20*64
output = conv_bn_prelu(
output,
kernel_size=3,
stride=1,
output_channels=self.output_channels[str(block_num)][1],
name='conv_2',
is_training=self.is_training)
output = conv_bn_prelu(
output,
kernel_size=3,
stride=2,
output_channels=self.output_channels[str(block_num)][2],
name='conv_3',
is_training=self.is_training)
#output = self.feature_pyramid(output)
# do conv 1 x 1 x 1 before sum
input = _conv3d(_input,
kernel_size=1,
stride=2,
output_feature=self.output_channels[str(block_num)][2],
use_bias=True,
name='conv_s2')
#input = conv_bn_prelu(_input, kernel_size=1, stride=2, output_channels=self.output_channels[str(block_num)][2],
# name='conv_s2', is_training=self.is_training)
output = tf.add(output, input, name='elemwise_sum')
output = self.feature_pyramid(output, block_num)
return output
def inference_op(self, _input):
input_channels = 1
output_channels = 2
n_channels = 16
_input = tf.pad(_input,
np.array([[0, 0], [1, 0], [1, 1], [0, 0], [0, 0]]),
name='pad_1')
with tf.variable_scope('contracting_path'):
# if the input has more than 1 channel it has to be expanded because broadcasting only works for 1 input channel
if input_channels == 1:
c0 = tf.tile(_input, [1, 1, 1, 1, n_channels])
else:
with tf.variable_scope('level_0'):
#c0 = convolution_3d(_input, [5, 5, 5, input_channels, n_channels], [1, 1, 1, 1, 1])
c0 = _conv3d(_input,
output_feature=n_channels,
kernel_size=5,
stride=1)
#c0 = v_conv3d(_input, kernel_shape=[5, 5, 5, input_channels, n_channels], stride=1)
c0 = prelu(c0)
print(c0)
with tf.variable_scope('level_1'):
c1 = convolution_block(c0, n_channels, 1)
c12 = down_convolution(c1, n_channels)
with tf.variable_scope('level_2'):
c2 = convolution_block(c12, n_channels * 2, 2)
c22 = down_convolution(c2, n_channels * 2)
with tf.variable_scope('level_3'):
c3 = convolution_block(c22, n_channels * 4, 3)
c32 = down_convolution(c3, n_channels * 4)
with tf.variable_scope('level_4'):
c4 = convolution_block(c32, n_channels * 8, 3)
c42 = down_convolution(c4, n_channels * 8)
with tf.variable_scope('level_5'):
c5 = convolution_block(c42, n_channels * 16, 3)
c52 = up_convolution(c5, tf.shape(c4), n_channels * 16)
with tf.variable_scope('expanding_path'):
with tf.variable_scope('level_4'):
e4 = convolution_block_2(c52, c4, n_channels * 8, 3)
e42 = up_convolution(e4, tf.shape(c3), n_channels * 8)
with tf.variable_scope('level_3'):
e3 = convolution_block_2(e42, c3, n_channels * 4, 3)
e32 = up_convolution(e3, tf.shape(c2), n_channels * 4)
with tf.variable_scope('level_2'):
e2 = convolution_block_2(e32, c2, n_channels * 2, 2)
e22 = up_convolution(e2, tf.shape(c1), n_channels * 2)
with tf.variable_scope('level_1'):
e1 = convolution_block_2(e22, c1, n_channels, 1)
with tf.variable_scope('output_layer'):
#logits = convolution_3d(e1, [1, 1, 1, n_channels, output_channels], [1, 1, 1, 1, 1])
logits = _conv3d(e1,
output_feature=output_channels,
kernel_size=1,
stride=1)
#logits = v_conv3d(e1, kernel_shape=[1, 1, 1, n_channels, output_channels], stride=1)
logits = logits[:, :103, :198, :]
with tf.variable_scope('prediction'):
softmax_prob = tf.nn.softmax(logits=logits, name='softmax_prob')
predicted_label = tf.argmax(input=softmax_prob,
axis=4,
name='predicted_label')
#predicted_label = tf.nn.sigmoid(logits, name='predicted_label')
#print(logits,predicted_label)
return logits, predicted_label
def inference_op(self, _input):
conv_layer = []
dconv_layer = []
# padding output
output = tf.pad(_input,
np.array([[0, 0], [1, 0], [1, 1], [0, 0], [0, 0]]),
name='pad_1')
with tf.device(device_name_or_function=self.device[0]):
output = conv_bn_prelu(output,
output_channels=8,
kernel_size=3,
stride=1,
is_training=self.is_training,
name='conv_1') # 104x200x160 8
input_layer = output
conv_layer.append(output)
for block_num in range(1, self.cont_block_num + 1):
with tf.variable_scope('contract_block_%d' % block_num):
output = self._contracting_block(block_num, output,
input_layer)
conv_layer.append(output)
for block_num in range(4, self.expand_block_num + 4):
with tf.variable_scope('expand_block_%d' % block_num):
output = self._expanding_block(block_num, output,
conv_layer[2 - block_num])
dconv_layer.append(output)
with tf.device(device_name_or_function=self.device[1]):
'''auxiliary prediction'''
# forth level
auxiliary3_prob_4x = _conv3d(inputs=dconv_layer[0],
output_feature=self.output_classes,
kernel_size=1,
stride=1,
use_bias=True,
name='auxiliary3_prob_4x')
auxiliary3_prob_2x = deconv3d(inputs=auxiliary3_prob_4x,
output_channels=self.output_classes,
name='auxiliary3_prob_2x')
auxiliary3_prob_1x = deconv3d(inputs=auxiliary3_prob_2x,
output_channels=self.output_classes,
name='auxiliary3_prob_1x')
# third level
auxiliary2_prob_2x = _conv3d(inputs=dconv_layer[1],
output_feature=self.output_classes,
kernel_size=1,
stride=1,
use_bias=True,
name='auxiliary2_prob_2x')
auxiliary2_prob_1x = deconv3d(inputs=auxiliary2_prob_2x,
output_channels=self.output_classes,
name='auxiliary2_prob_2x')
# second level
auxiliary1_prob_1x = _conv3d(inputs=dconv_layer[2],
output_feature=self.output_classes,
kernel_size=1,
stride=1,
use_bias=True,
name='auxiliary1_prob_1x')
# with tf.variable_scope('last_stage'):
# # out_feature = int(output.get_shape()[-1]) / 2
# #print(dconv_layer[0],dconv_layer[1],dconv_layer[2])
# output1 = _conv3d(dconv_layer[0], kernel_size=1, stride=1, output_feature=5, use_bias=True,
# name='block1_conv1x1')
#
# #output1 = deconv3d(output1, output_channels=int(output1.get_shape()[-1]), name='block1_deconv')
#
# output1 = BilinearUpsample3d(output1, up_factor=2)
#
# output2 = _conv3d(dconv_layer[1], kernel_size=1, stride=1, output_feature=5, use_bias=True,
# name='block2_conv1x1')
#
# output2 = tf.add(output1, output2)
#
# #output2 = deconv3d(output2, output_channels=int(output2.get_shape()[-1]), name='block2_deconv')
# output2 = BilinearUpsample3d(output2, up_factor=2)
#
# output3 = _conv3d(dconv_layer[2], kernel_size=1, stride=1, output_feature=5, use_bias=True,
# name='block3_conv1x1')
#
# output3 = tf.add(output2, output3)
#
output = _conv3d(output,
kernel_size=1,
stride=1,
output_feature=self.output_classes,
use_bias=True,
name='fc_layer')
output = output[:, :103, :198, :]
with tf.device(device_name_or_function=self.device[2]):
with tf.variable_scope('prediction'):
softmax_prob = tf.nn.softmax(logits=output,
name='softmax_prob')
predicted_label = tf.argmax(input=softmax_prob,
axis=4,
name='predicted_label')
#predicted_label = tf.nn.sigmoid(output, name='predicted_label')
return output, predicted_label, auxiliary1_prob_1x, auxiliary2_prob_1x, auxiliary3_prob_1x
def inference_op(self, _input):
conv_layer = []
dconv_layer = []
# padding output
output = tf.pad(_input, np.array([[0, 0], [1, 0], [1, 1], [0, 0], [0, 0]]), name='pad_1')
output = conv_bn_prelu(output, output_channels=8, kernel_size=3, stride=1,
is_training=self.is_training, name='conv_1') # 104x200x160 8
conv_layer.append(output)
for block_num in range(1, self.cont_block_num + 1):
with tf.variable_scope('contract_block_%d' % block_num):
output = self._contracting_block(block_num,output)
conv_layer.append(output)
for block_num in range(4, self.expand_block_num + 4):
with tf.variable_scope('expand_block_%d' % block_num):
output = self._expanding_block(block_num, output, conv_layer[2 - block_num])
dconv_layer.append(output)
'''auxiliary prediction'''
# forth level
auxiliary3_prob_4x = _conv3d(inputs=dconv_layer[0], output_feature=1, kernel_size=1,
stride=1, use_bias=True, name='auxiliary3_prob_4x')
auxiliary3_prob_2x = deconv3d(inputs=auxiliary3_prob_4x, output_channels=1,
name='auxiliary3_prob_2x')
auxiliary3_prob_1x = deconv3d(inputs=auxiliary3_prob_2x, output_channels=1,
name='auxiliary3_prob_1x')
# third level
auxiliary2_prob_2x = _conv3d(inputs=dconv_layer[1], output_feature=1, kernel_size=1,
stride=1, use_bias=True, name='auxiliary2_prob_2x')
auxiliary2_prob_1x = deconv3d(inputs=auxiliary2_prob_2x, output_channels=1,
name='auxiliary2_prob_2x')
# second level
auxiliary1_prob_1x = _conv3d(inputs=dconv_layer[2], output_feature=1, kernel_size=1,
stride=1, use_bias=True, name='auxiliary1_prob_1x')
with tf.variable_scope('last_stage'):
# out_feature = int(output.get_shape()[-1]) / 2
#print(dconv_layer[0],dconv_layer[1],dconv_layer[2])
output1 = _conv3d(dconv_layer[0], kernel_size=1, stride=1, output_feature=5, use_bias=True,
name='block1_conv1x1')
#output1 = deconv3d(output1, output_channels=int(output1.get_shape()[-1]), name='block1_deconv')
output1 = BilinearUpsample3d(output1,up_factor=2)
#print('block1_deconv1', output1)
# out_feature = int(output.get_shape()[-1]) / 2
output2 = _conv3d(dconv_layer[1], kernel_size=1, stride=1, output_feature=5, use_bias=True,
name='block2_conv1x1')
output2 = tf.add(output1, output2)
#output2 = deconv3d(output2, output_channels=int(output2.get_shape()[-1]), name='block2_deconv')
output2 = BilinearUpsample3d(output2, up_factor=2)
output3 = _conv3d(dconv_layer[2], kernel_size=1, stride=1, output_feature=5, use_bias=True,
name='block3_conv1x1')
output3 = tf.add(output2, output3)
output = _conv3d(output3, kernel_size=1, stride=1, output_feature=1, use_bias=True, name='fc_layer')
return output, auxiliary1_prob_1x, auxiliary2_prob_1x, auxiliary3_prob_1x
def inference_op(self, _input):
conv_layer = []
dconv_layer = []
# padding output
output = tf.pad(_input,
np.array([[0, 0], [1, 0], [1, 1], [0, 0], [0, 0]]),
name='pad_1')
#print(output)
output = conv_bn_prelu(output,
output_channels=8,
kernel_size=3,
stride=1,
is_training=self.is_training,
name='conv_1')
conv_layer.append(output)
for i in range(self.cont_block_num):
with tf.variable_scope('contract_block_%d' % i):
output = self._contracting_block(output)
#print(output)
conv_layer.append(output)
for i in range(self.expand_block_num):
with tf.variable_scope('expand_block_%d' % i):
output = self._expanding_block(output, conv_layer[2 - i])
dconv_layer.append(output)
'''auxiliary prediction'''
# forth level
auxiliary3_prob_4x = _conv3d(inputs=dconv_layer[0],
output_feature=1,
kernel_size=1,
stride=1,
use_bias=True,
name='auxiliary3_prob_4x')
auxiliary3_prob_2x = deconv3d(inputs=auxiliary3_prob_4x,
output_channels=1,
name='auxiliary3_prob_2x')
auxiliary3_prob_1x = deconv3d(inputs=auxiliary3_prob_2x,
output_channels=1,
name='auxiliary3_prob_1x')
# third level
auxiliary2_prob_2x = _conv3d(inputs=dconv_layer[1],
output_feature=1,
kernel_size=1,
stride=1,
use_bias=True,
name='auxiliary2_prob_2x')
auxiliary2_prob_1x = deconv3d(inputs=auxiliary2_prob_2x,
output_channels=1,
name='auxiliary2_prob_2x')
# second level
auxiliary1_prob_1x = _conv3d(inputs=dconv_layer[2],
output_feature=1,
kernel_size=1,
stride=1,
use_bias=True,
name='auxiliary1_prob_1x')
#print(auxiliary3_prob_1x,'\n',auxiliary2_prob_1x,'\n',auxiliary1_prob_1x)
with tf.variable_scope('last_stage'):
# out_feature = int(output.get_shape()[-1]) / 2
_output = _conv3d(dconv_layer[0],
kernel_size=1,
stride=1,
output_feature=32,
use_bias=True,
name='block1_conv1x1')
_output = deconv3d(_output,
output_channels=int(_output.get_shape()[-1]),
name='block1_deconv')
#print('block1_deconv1', _output)
# out_feature = int(output.get_shape()[-1]) / 2
_output2 = _conv3d(dconv_layer[1],
kernel_size=1,
stride=1,
output_feature=32,
use_bias=True,
name='block2_conv1x1')
_output = tf.add(_output, _output2)
#print('1', _output)
_output = deconv3d(_output,
output_channels=int(_output.get_shape()[-1]),
name='block2_deconv')
# out_feature = int(output.get_shape()[-1]) / 2
_output3 = _conv3d(dconv_layer[2],
kernel_size=1,
stride=1,
output_feature=32,
use_bias=True,
name='block3_conv1x1')
output = tf.add(_output, _output3)
output = _conv3d(output,
kernel_size=1,
stride=1,
output_feature=1,
use_bias=True,
name='fc_layer')
#logits = tf.nn.sigmoid(output)'''
return output, auxiliary1_prob_1x, auxiliary2_prob_1x, auxiliary3_prob_1x
