def upsample(self, node, current_level, is_training):
if self.data_format == 'channels_last':
# suppose that 'channels_last' means CPU
# resize_trilinear is much faster on CPU
return resize_trilinear(node,
factors=[2, 2, 2],
name='upsample' + str(current_level),
data_format=self.data_format)
else:
# suppose that 'channels_first' means GPU
# upsample3d_linear is much faster on GPU
return upsample3d_linear(node,
factors=[2, 2, 2],
name='upsample' + str(current_level),
data_format=self.data_format,
padding='valid_cropped')
def upsample(self, node, current_level, is_training):
return upsample3d_linear(node, [2, 2, 2],
name='upsample' + str(current_level),
data_format=self.data_format)
def network_scn(input, num_labels, is_training, data_format='channels_first'):
downsampling_factor = 4
kernel_initializer = he_initializer
activation = tf.nn.relu
local_kernel_initializer = he_initializer
local_activation = tf.nn.tanh
spatial_kernel_initializer = he_initializer
spatial_activation = None
padding = 'reflect'
with tf.variable_scope('unet'):
unet = UnetClassicAvgLinear3d(64,
4,
data_format=data_format,
double_filters_per_level=True,
kernel_initializer=kernel_initializer,
activation=activation,
padding=padding)
local_prediction = unet(input, is_training=is_training)
local_prediction = conv3d(local_prediction,
num_labels, [1, 1, 1],
name='local_prediction',
padding=padding,
kernel_initializer=local_kernel_initializer,
activation=local_activation,
is_training=is_training)
with tf.variable_scope('spatial_configuration'):
local_prediction_pool = avg_pool3d(local_prediction,
[downsampling_factor] * 3,
name='local_prediction_pool')
scconv = conv3d(local_prediction_pool,
64, [5, 5, 5],
name='scconv0',
padding=padding,
kernel_initializer=kernel_initializer,
activation=activation,
is_training=is_training)
scconv = conv3d(scconv,
64, [5, 5, 5],
name='scconv1',
padding=padding,
kernel_initializer=kernel_initializer,
activation=activation,
is_training=is_training)
scconv = conv3d(scconv,
64, [5, 5, 5],
name='scconv2',
padding=padding,
kernel_initializer=kernel_initializer,
activation=activation,
is_training=is_training)
spatial_prediction_pool = conv3d(
scconv,
num_labels, [5, 5, 5],
name='spatial_prediction_pool',
padding=padding,
kernel_initializer=spatial_kernel_initializer,
activation=spatial_activation,
is_training=is_training)
spatial_prediction = upsample3d_linear(spatial_prediction_pool,
[downsampling_factor] * 3,
name='spatial_prediction',
padding='valid_cropped')
with tf.variable_scope('combination'):
prediction = local_prediction * spatial_prediction
return prediction, local_prediction, spatial_prediction