def construct_up_block_2_sub_pixel(self,
inputs,
down_inputs,
name,
first=False,
final=False):
print("--------")
print("Inputs:")
print(inputs.shape)
if first:
# producing r^2 times more feature maps
num_outputs = inputs.shape[self.channel_axis].value
inputs = tf.contrib.layers.conv2d(inputs,
num_outputs * self.conf.ratio**2,
self.conv_size,
scope=name + '/conv_sub_pixel',
data_format='NHWC',
activation_fn=None,
biases_initializer=None)
print("Conv before subpixel:")
print(inputs.shape)
sub_pixel_conv = self.deconv_func()(
inputs=inputs,
scope=name + '/subpixel',
r=self.conf.ratio,
debug=self.conf.debug,
activation=self.conf.activation_function)
print("Sub Pixel:")
print(sub_pixel_conv.shape)
concat = tf.concat([sub_pixel_conv, down_inputs],
self.channel_axis,
name=name + '/concat')
print("Sub Pixel + down inputs:")
print(concat.shape)
num_outputs = concat.shape[self.channel_axis].value
num_outputs = num_outputs / 2 if final else num_outputs
conv2 = ops.conv2d(concat,
num_outputs,
self.conv_size,
name + '/conv2',
activation=self.conf.activation_function)
print("Conv after concat:")
print(conv2.shape)
num_outputs = self.conf.class_num if final else num_outputs
conv3 = ops.conv2d(conv2,
num_outputs,
self.conv_size,
name + '/conv3',
activation=self.conf.activation_function)
print("Conv:")
print(conv3.shape)
return conv3
def construct_up_block(self, inputs, down_inputs, name, final=False):
num_outputs = inputs.shape[self.channel_axis].value
conv1 = self.deconv_func()(inputs, num_outputs, self.conv_size,
name + '/conv1')
print("--------")
print("Inputs:")
print(inputs.shape)
print("Deconv:")
print(conv1.shape)
print("Down inputs:")
print(down_inputs.shape)
conv1 = tf.concat([conv1, down_inputs],
self.channel_axis,
name=name + '/concat')
print("After concat:")
print(conv1.shape)
conv2 = self.conv_func()(conv1, num_outputs, self.conv_size,
name + '/conv2')
num_outputs = self.conf.class_num if final else num_outputs / 2
print("Conv2:")
print(conv2.shape)
conv3 = ops.conv2d(conv2, num_outputs, self.conv_size, name + '/conv3')
print("Conv3:")
print(conv3.shape)
print("--------")
return conv3
def construct_down_block(self, inputs, name, down_outputs, first=False):
num_outputs = self.conf.start_channel_num if first else 2 * inputs.shape[
self.channel_axis].value
conv1 = ops.conv2d(inputs,
num_outputs,
self.conv_size,
name + '/conv1',
activation=self.conf.activation_function)
conv2 = ops.conv2d(conv1,
num_outputs,
self.conv_size,
name + '/conv2',
activation=self.conf.activation_function)
down_outputs.append(conv2)
pool = ops.pool2d(conv2, self.pool_size, name + '/pool')
return pool
def construct_bottom_block(self, inputs, name):
print("--------")
print("Bottom layer:")
print("Inputs:")
print(inputs.shape)
num_outputs = inputs.shape[self.channel_axis].value
conv1 = ops.conv2d(inputs,
2 * num_outputs,
self.conv_size,
name + '/conv1',
activation=self.conf.activation_function)
print("Conv1:")
print(conv1.shape)
if self.conf.up_architecture == 1 or self.conf.up_architecture == 2:
conv2 = ops.conv2d(conv1,
num_outputs,
self.conv_size,
name + '/conv2',
activation=self.conf.activation_function)
print("Conv2:")
print(conv2.shape)
return conv2
else:
return conv1
def construct_up_block_5_sub_pixel(self,
inputs,
down_inputs,
name,
first=False,
final=False):
print("--------")
print("Inputs:")
print(inputs.shape)
if first:
# producing r^2 times more feature maps (in this architecture the first layer only
# needs to produce 2x more feature maps)
num_outputs = inputs.shape[self.channel_axis].value
inputs = tf.contrib.layers.conv2d(inputs,
num_outputs * 2,
self.conv_size,
scope=name + '/conv_sub_pixel',
data_format='NHWC',
activation_fn=None,
biases_initializer=None)
print("Conv before subpixel:")
print(inputs.shape)
sub_pixel_conv = self.deconv_func()(
inputs=inputs,
scope=name + '/subpixel',
r=self.conf.ratio,
debug=self.conf.debug,
activation=self.conf.activation_function)
print("Sub Pixel:")
print(sub_pixel_conv.shape)
outputs = tf.concat([sub_pixel_conv, down_inputs],
self.channel_axis,
name=name + '/concat')
print("Sub Pixel + down inputs:")
print(outputs.shape)
num_outputs = outputs.shape[self.channel_axis].value
if final:
outputs = ops.conv2d(outputs,
num_outputs,
self.conv_size,
name + '/unique_conv',
activation=self.conf.activation_function)
print("Conv unique:")
print(outputs.shape)
num_outputs = num_outputs / 2 if final else num_outputs
outputs = ops.conv2d(outputs,
num_outputs,
self.conv_size,
name + '/conv2',
activation=self.conf.activation_function)
print("Conv after concat:")
print(outputs.shape)
if final:
print("Its the final one:")
num_outputs = self.conf.class_num
outputs = tf.contrib.layers.conv2d(outputs,
num_outputs,
kernel_size=(1, 1),
stride=1,
scope=name + '/conv_final',
data_format='NHWC',
activation_fn=None,
biases_initializer=None,
padding='VALID')
else:
outputs = ops.conv2d(outputs,
num_outputs,
self.conv_size,
name + '/conv3',
activation=self.conf.activation_function)
print("Conv:")
print(outputs.shape)
return outputs