def cheb_fc_fn(x, L, output_units, filter_order, num_filters):
"""
Neural network consisting on 1 convolutional chebyshev layer and one dense layer
:param x: input signal
:param L: graph laplacian
:param output_units: number of output units
:param filter_order: order of convolution
:param num_filters: number of parallel filters
:return: computational graph
"""
with tf.name_scope("chebyshev_conv"):
with tf.name_scope("weights"):
Wcheb = _weight_variable([filter_order, num_filters])
_variable_summaries(Wcheb)
with tf.name_scope("biases"):
bcheb = _bias_variable([num_filters])
_variable_summaries(bcheb)
graph_conv = chebyshev_convolution(x, L, Wcheb, bcheb)
graph_conv = tf.nn.relu(graph_conv)
with tf.name_scope("dropout"):
keep_prob = tf.placeholder(tf.float32)
dropout = tf.nn.dropout(graph_conv, keep_prob=keep_prob)
with tf.name_scope("fc"):
fc_input = tf.layers.flatten(dropout)
fc = tf.layers.dense(inputs=fc_input,
units=output_units,
activation=None,
use_bias=True)
return fc, keep_prob
def _cheb_conv_layer(x, L, vertex_filter_order, num_filters):
_, _, _, num_channels = x.get_shape()
num_channels = int(num_channels)
with tf.name_scope("fir_tv"):
with tf.name_scope("weights"):
hfir = _weight_variable([vertex_filter_order, num_channels, num_filters])
_variable_summaries(hfir)
graph_conv = chebyshev_convolution(x, L, hfir, None)
return graph_conv