def build_single_channel(var,
input_size,
output_size,
layer_sizes,
layer_types,
weight_init=lasagne.init.GlorotUniform(),
bias_init=lasagne.init.Constant(0.),
name='',
dropouts_init=None):
"""
Build a single channel containing layers of sizes according to layer_sizes array with initialization given
:param var: tensor variable for the input
:param input_size: input dimensionality
:param output_size: ouput dimensionality
:param layer_sizes: array of layer sizes
:param layer_types: array of layer types
:param weight_init: initialization function for the weights
:param bias_init: initialization function for the biases
:param name: name of the network
:param dropouts_init: initialization of the tied dropout, if none samples a new drop matrix
:return: the model containing the channels with it's weights, biases, drop matrices and batch normalization hooks
"""
model = []
weights = []
biases = []
hidden = []
dropouts = []
hooks = {}
if isinstance(weight_init, lasagne.init.Initializer):
weight_init = [weight_init for i in range(len(layer_sizes) + 1)]
if isinstance(bias_init, lasagne.init.Initializer):
bias_init = [bias_init for i in range(len(layer_sizes) + 1)]
if dropouts_init is None:
dropouts_init = [dropouts_init for i in range(len(layer_sizes) + 1)]
# Add Input Layer
model.append(
lasagne.layers.InputLayer((None, input_size), var,
'input_layer_{0}'.format(name)))
# Add hidden layers
for index, layer_size in enumerate(layer_sizes):
model.append(layer_types[index](
incoming=model[-1],
num_units=layer_size,
W=weight_init[index],
b=bias_init[index],
nonlinearity=lasagne.nonlinearities.LeakyRectify(Params.LEAKINESS)
if not Params.BN_ACTIVATION else lasagne.nonlinearities.identity,
cell_num=Params.LOCALLY_DENSE_M))
weights.append(model[-1].W)
biases.append(model[-1].b)
if Params.BN:
model.append(
BatchNormalizationLayer(
model[-1],
nonlinearity=lasagne.nonlinearities.LeakyRectify(
Params.LEAKINESS) if Params.BN_ACTIVATION else
lasagne.nonlinearities.identity))
model.append(
Params.NOISE_LAYER(model[-1],
p=Params.DROP_PROBABILITY,
noise_layer=dropouts_init[-(index + 1)]))
dropouts.append(model[-1])
hidden.append(model[-1])
# Add output layer
model.append(layer_types[-1](model[-1],
num_units=output_size,
W=weight_init[-1],
b=bias_init[-1],
nonlinearity=lasagne.nonlinearities.identity,
cell_num=Params.LOCALLY_DENSE_M))
weights.append(model[-1].W)
biases.append(model[-1].b)
prediction = lasagne.layers.get_output(model[-1], moving_avg_hooks=hooks)
return model, hidden, weights, biases, prediction, hooks, dropouts
