def dense_block(self,
input_x,
nb_layers,
layer_name,
keep_prob=None,
downsample=False):
"""
Creates a dense block connection all same sized filters
:param input_x: The input to this dense block (output of prior downsample operation)
:param nb_layers: how many layers desired
:param layer_name: base name of this block
:param keep_prob: Whether to use dropout
:param trans: whether to include a downsample at the end
:return:
"""
with tf.name_scope(layer_name):
# Array to hold each layer
layers_concat = []
# Append to list
layers_concat.append(input_x)
# The first layer of this block
conv = self.bottleneck_layer(input_x, (layer_name + '_denseN_0'),
keep_prob)
# Concat the first layer
layers_concat.append(conv)
# Loop through the number of layers desired
for z in range(nb_layers):
# Concat all the prior layer into this layer
conv = tf.concat(layers_concat, axis=-1)
# Create a new layer
conv = self.bottleneck_layer(
conv, (layer_name + '_denseN_' + str(z + 1)), keep_prob)
# Append this layer to the running list of dense connected layers
layers_concat.append(conv)
# Combine the layers
conv = tf.concat(layers_concat, axis=-1)
# Downsample if requested
if downsample:
conv = self.transition_layer(conv,
(layer_name + '_Downsample'),
keep_prob)
return conv
def dense_block_3d(self, input_x, nb_layers, layer_name, keep_prob=None):
"""
Creates a dense block
:param input_x: The input to this dense block (output of prior downsample operation)
:param nb_layers: how many layers desired
:param layer_name: base name of this block
:param keep_prob: Whether to use dropout
:return:
"""
with tf.name_scope(layer_name):
# Array to hold each layer
layers_concat = []
# Append to list
layers_concat.append(input_x)
# The first layer of this block
conv = self.bottleneck_layer_3d(input_x,
(layer_name + '_denseN_0'),
keep_prob)
# Concat the first layer
layers_concat.append(conv)
# Loop through the number of layers desired
for z in range(nb_layers):
# Concat all the prior layer into this layer
conv = tf.concat(layers_concat, axis=-1)
# Create a new layer
conv = self.bottleneck_layer_3d(
conv, (layer_name + '_denseN_' + str(z + 1)), keep_prob)
# Append this layer to the running list of dense connected layers
layers_concat.append(conv)
# The concat has to be 2D, first retreive the Z and K size
concat = tf.concat(layers_concat, -1)
Fz, Ch = concat.get_shape().as_list()[1], concat.get_shape(
).as_list()[-1]
# Now create a projection matrix
concat = self.convolution_3d(layer_name + '_projection',
concat, [Fz, 1, 1],
Ch,
1,
'VALID',
self.phase_train,
BN=False)
return conv, tf.squeeze(concat)
def up_transition_3d(self,
scope,
X,
F,
K,
S,
concat_var=None,
padding='VALID'):
"""
Deconvolutions for the DenseUnet
:param scope:
:param X:
:param F:
:param K:
:param S:
:param padding:
:param concat_var:
:param summary:
:return:
"""
with tf.variable_scope(scope) as scope:
# Set channel size based on input depth
C = X.get_shape().as_list()[-1]
# He init
kernel = tf.get_variable(
'Weights',
shape=[F, F, F, K, C],
initializer=tf.contrib.layers.variance_scaling_initializer())
# Define the biases
bias = tf.get_variable('Bias',
shape=[K],
initializer=tf.constant_initializer(0.0))
# Add to the weights collection
tf.add_to_collection('weights', kernel)
tf.add_to_collection('biases', bias)
# Define the output shape
out_shape = X.get_shape().as_list()
out_shape[1] *= 2
out_shape[2] *= 2
out_shape[3] *= 2
out_shape[3] = K
# Perform the deconvolution. output_shape: A 1-D Tensor representing the output shape of the deconvolution op.
conv = tf.nn.conv2d_transpose(X,
kernel,
output_shape=out_shape,
strides=[1, S, S, S, 1],
padding=padding)
# Add in bias
conv = tf.nn.bias_add(conv, bias)
# Concatenate
conv = tf.concat([concat_var, conv], axis=-1)
# Create a histogram summary and summary of sparsity
if self.summary: self._activation_summary(conv)
return conv
def up_transition(self,
scope,
X,
F,
K,
S,
concat_var=None,
padding='SAME',
res=True):
"""
Performs an upsampling procedure
:param scope:
:param X: Inputs
:param F: Filter sizes
:param K: Kernel sizes
:param S: Stride size
:param concat_var: The skip connection
:param padding: SAME or VALID. In general, use VALID for 3D skip connections
:param res: Whether to concatenate or add the skip connection
:return:
"""
with tf.variable_scope(scope) as scope:
# Set channel size based on input depth
C = X.get_shape().as_list()[-1]
# He init
kernel = tf.get_variable(
'Weights',
shape=[F, F, K, C],
initializer=tf.contrib.layers.variance_scaling_initializer())
# Define the biases
bias = tf.get_variable('Bias',
shape=[K],
initializer=tf.constant_initializer(0.0))
# Add to the weights collection
tf.add_to_collection('weights', kernel)
tf.add_to_collection('biases', bias)
# Define the output shape based on shape of skip connection
out_shape = concat_var.get_shape().as_list()
out_shape[-1] = K
# Perform the deconvolution. output_shape: A 1-D Tensor representing the output shape of the deconvolution op.
conv = tf.nn.conv2d_transpose(X,
kernel,
output_shape=out_shape,
strides=[1, S, S, 1],
padding=padding)
# Concatenate
if res: conv = tf.add(conv, concat_var)
else: conv = tf.concat([concat_var, conv], axis=-1)
# Apply the batch normalization. Updates weights during training phase only
conv = self.batch_normalization(conv, self.phase_train, scope)
# Add in bias
conv = tf.nn.bias_add(conv, bias)
# Relu
conv = tf.nn.relu(conv, name=scope.name)
# Create a histogram summary and summary of sparsity
if self.summary: self._activation_summary(conv)
return conv