def forward(self, inputs):
"""
Forward process
Parameters
----------
inputs: Input tensor
"""
# encoder process
outputs = inputs
encoders = []
for i, filters in enumerate(self.block_filters):
first_blocks = i == 0
outputs = self.build_res_blocks(outputs, filters,
self.block_res_nums[i],
first_blocks)
if i != len(self.block_filters) - 1:
encoders.append(outputs)
outputs = maxpool(outputs, pool_size=2)
# decoder process
encoder_num = len(encoders)
for k, encoder in enumerate(encoders[::-1]):
i = encoder_num - k - 1
first_blocks = i == 0
filters = int(encoder.get_shape()[-1])
outputs = conv_bn_relu(outputs,
filters,
3,
1,
1,
use_bn=self.use_bn,
training=self.training)
outputs = upsample(outputs, 2, self.interpolation_type)
outputs = concat([encoder, outputs])
outputs = self.build_res_blocks(outputs, filters,
self.block_res_nums[i],
first_blocks)
outputs = conv(outputs, self.config.get('output_channel'))
outputs = softmax(outputs)
return outputs
def build_one_dense_block(self, inputs, growth_rate):
"""
Build one dense conv block
Parameters
----------
inputs: Input tensor
growth_rate: number of channel growth for each dense block
"""
if self.config.get('bottleneck'):
shrinked_filters = self.config.get('bottleneck_factor') * \
growth_rate
branch = self.conv_block(inputs, shrinked_filters, ksize=1)
branch = self.conv_block(branch, growth_rate, ksize=3)
else:
branch = self.conv_block(inputs, growth_rate, ksize=3)
if self.config.get('dropout_rate'):
branch = dropout(branch, self.config.get('dropout_rate'))
outputs = concat([inputs, branch])
return outputs