def composite_function(_input, outSize, width):
with tf.variable_scope("composite_function"):
output = _conv1d(_input, outSize, width=width)
output = ln(output)
output = tf.nn.relu(output)
return output
def conv1d(x, outputsize, width):
inputSize = x.get_shape()[-1]
filter_ = tf.get_variable("conv_filter",
shape=[width, inputSize, outputsize])
bias = tf.get_variable("conv_bias", shape=[outputsize])
convolved = tf.nn.conv1d(x, filters=filter_, stride=2, padding="SAME")
convolved += bias
convolved = ln(convolved)
return convolved
def transition_to_vector(_input):
'''
Transforms the last block into a single vector by avg_pooling
'''
output = ln(_input)
output = tf.nn.relu(output)
last_pool_kernel = int(output.get_shape()[-2])
output = avg_pool1d(output, last_pool_kernel)
output = tf.squeeze(output, axis=1)
return output
def bottleneck(_input, growthRate):
'''
Per the paper, each bottlneck outputs 4k feature size where k is the growth rate of the network.
:return:
'''
outSize = growthRate * 4
output = _conv1d(_input, outSize, width=1)
output = ln(output)
output = tf.nn.relu(output)
return output
def conv1d_transpose(x, targetHidden, width, growth_rate):
length, inputHidden = x.get_shape()[-2:]
outputShape = [
FLAGS.batch_size, 1, length.value * growth_rate, targetHidden
]
while len(x.get_shape()) < 4:
x = tf.expand_dims(x, axis=1)
bias = tf.get_variable("deconv_filter", shape=[targetHidden])
filter_ = tf.get_variable("deconv_bias",
shape=[1, width, targetHidden, inputHidden])
conv_trans = tf.nn.conv2d_transpose(x,
filter=filter_,
output_shape=outputShape,
strides=[1, 1, growth_rate, 1])
conv_trans += bias
conv_trans = ln(conv_trans)
conv_trans = tf.nn.relu(conv_trans)
conv_trans = tf.squeeze(conv_trans, 1)
return conv_trans
def nonlin_ln(x):
return ln(
tf.nn.dropout(tf.nn.softsign(x), keep_prob=FLAGS.dropout_keep_prob))
def layer_norm(input_tensor):
"""Run layer normalization on the last dimension of the tensor."""
return ln(inputs=input_tensor, begin_norm_axis=-1, begin_params_axis=-1)