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)
#dilate_rate = int(np.power(2,4-block_num)) # 8 8 8 // 4 4 4 // 2 2 2
dilate_rate = int(np.power(2, i)) # 2 4 8 // 2 4 8 // 2 4 8
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)
# output = self.Squeeze_excitation_layer(output, int(output.get_shape()[-1]), self.reduction_ratio,
# layer_name='SE%d' % block_num)
output = _conv2d(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'):
output = _conv2d(
_input,
kernel_size=1,
stride=1,
output_feature=self.output_channels[str(block_num)][0],
use_bias=True,
name='conv_1')
#print('ex1',output)
output = deconv_bn_prelu(
output,
output_channels=self.output_channels[str(block_num)][1],
is_training=self.is_training,
name='deconv')
#print('de1',output)
if option == 'sum':
output = tf.add(output, layer, name='elemwise_sum')
else:
output = tf.concat(values=(output, layer), axis=-1, 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)
#print('out1',output)
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('out2',output)
return output
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 = _conv2d(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 = _conv2d(
_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=-1, 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 inference_op(self, _input):
conv_layer = []
dconv_layer = []
with tf.device(device_name_or_function=self.device[0]):
output = conv_bn_prelu(_input,
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)
self.class_layer.append(conv_layer[-1])
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 = _conv2d(inputs=dconv_layer[0],
output_feature=self.output_classes,
kernel_size=1,
stride=1,
use_bias=True,
name='auxiliary3_prob_4x')
auxiliary3_prob_2x = deconv2d(inputs=auxiliary3_prob_4x,
output_channels=self.output_classes,
name='auxiliary3_prob_2x')
auxiliary3_prob_1x = deconv2d(inputs=auxiliary3_prob_2x,
output_channels=self.output_classes,
name='auxiliary3_prob_1x')
# third level
auxiliary2_prob_2x = _conv2d(inputs=dconv_layer[1],
output_feature=self.output_classes,
kernel_size=1,
stride=1,
use_bias=True,
name='auxiliary2_prob_2x')
auxiliary2_prob_1x = deconv2d(inputs=auxiliary2_prob_2x,
output_channels=self.output_classes,
name='auxiliary2_prob_2x')
# second level
auxiliary1_prob_1x = _conv2d(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 = _conv2d(dconv_layer[0],
kernel_size=1,
stride=1,
output_feature=5,
use_bias=True,
name='block1_conv1x1')
output1 = BilinearUpsample2d(output1, up_factor=2)
output2 = _conv2d(dconv_layer[1],
kernel_size=1,
stride=1,
output_feature=5,
use_bias=True,
name='block2_conv1x1')
output2 = tf.add(output1, output2)
output2 = BilinearUpsample2d(output2, up_factor=2)
output3 = _conv2d(dconv_layer[2],
kernel_size=1,
stride=1,
output_feature=5,
use_bias=True,
name='block3_conv1x1')
output3 = tf.add(output2, output3)
output = _conv2d(output3,
kernel_size=1,
stride=1,
output_feature=self.output_classes,
use_bias=True,
name='fc_layer')
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=-1,
name='predicted_label')
if cfg.joint_train:
with tf.variable_scope('class_layer'):
cls_input = output
cls_outputs = _conv2d(cls_input,
kernel_size=3,
stride=1,
output_feature=int(
cls_input.get_shape()[-1]),
name='cls_conv1')
# average pooling
last_pool_kernel = int(cls_outputs.get_shape()[-2])
k = last_pool_kernel
cls_outputs = tf.nn.avg_pool(cls_outputs,
ksize=[1, k, k, 1],
strides=[1, k, k, 1],
padding='VALID')
# print(k,outputs)
# FC
features_total = int(cls_outputs.get_shape()[-1])
cls_outputs = tf.reshape(cls_outputs, [-1, features_total])
W = tf.get_variable(
shape=[features_total, 2],
initializer=tf.contrib.layers.xavier_initializer(),
name='W')
bias = tf.get_variable(initializer=tf.constant(0.0, shape=[2]),
name='bias')
cls_outputs = tf.matmul(cls_outputs, W) + bias
with tf.variable_scope('class_prediction'):
cls_softmax = tf.nn.softmax(logits=cls_outputs,
name='softmax_prob')
cls_predict = tf.argmax(input=cls_softmax,
axis=-1,
name='predicted_label')
return output, predicted_label, auxiliary1_prob_1x, auxiliary2_prob_1x, auxiliary3_prob_1x, cls_outputs, cls_predict
else:
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(_input,
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 = _conv2d(inputs=dconv_layer[0],
output_feature=self.output_labels,
kernel_size=1,
stride=1,
use_bias=True,
name='auxiliary3_prob_4x')
auxiliary3_prob_2x = deconv2d(inputs=auxiliary3_prob_4x,
output_channels=self.output_labels,
name='auxiliary3_prob_2x')
auxiliary3_prob_1x = deconv2d(inputs=auxiliary3_prob_2x,
output_channels=self.output_labels,
name='auxiliary3_prob_1x')
# third level
auxiliary2_prob_2x = _conv2d(inputs=dconv_layer[1],
output_feature=self.output_labels,
kernel_size=1,
stride=1,
use_bias=True,
name='auxiliary2_prob_2x')
auxiliary2_prob_1x = deconv2d(inputs=auxiliary2_prob_2x,
output_channels=self.output_labels,
name='auxiliary2_prob_2x')
# second level
auxiliary1_prob_1x = _conv2d(inputs=dconv_layer[2],
output_feature=self.output_labels,
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 = _conv2d(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 = BilinearUpsample2d(output1, up_factor=2)
output2 = _conv2d(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 = BilinearUpsample2d(output2, up_factor=2)
output3 = _conv2d(dconv_layer[2],
kernel_size=1,
stride=1,
output_feature=5,
use_bias=True,
name='block3_conv1x1')
output3 = tf.add(output2, output3)
output = _conv2d(output3,
kernel_size=1,
stride=1,
output_feature=self.output_labels,
use_bias=True,
name='fc_layer')
#output = output[:, :103, :198, :]
with tf.device(device_name_or_function='/cpu:0'):
softmax_prob = tf.nn.softmax(logits=output, name='softmax_prob')
predicted_label = tf.argmax(input=softmax_prob,
axis=-1,
name='predicted_label')
#print(output,predicted_label,auxiliary3_prob_1x,auxiliary1_prob_1x,auxiliary2_prob_1x)
return output, predicted_label, auxiliary1_prob_1x, auxiliary2_prob_1x, auxiliary3_prob_1x