def __init__(self, inputs, n_layers, n_neurons, activation, output_size, name='decoder'):
with tf.name_scope(name):
self.input = inputs
# Slicer--Second slicer to multiply PE outcome
self.slice_output = RToC.get_c(self.input)
self.slice_output = tf.slice(self.slice_output, [0, l1], [-1, l2 - l1 + 1])
#print("output from slice_output: ", self.slice_output.shape)
# FE
self._fe = util.build_neural_net(input=self.input, n_layers=1, n_neurons=n_neurons,
activation=activation, n_outputs=output_size)
self._fe = tf.contrib.layers.fully_connected(self._fe, num_outputs=8, activation_fn=None)
self._fe = RToC.get_c(self._fe)
#print("output from FE: ", self._fe.shape)
# PE
self._pe = util.build_neural_net(input=self.input, n_layers=1, n_neurons=n_neurons,
activation=activation, n_outputs=output_size)
self._pe = tf.contrib.layers.fully_connected(self._pe, num_outputs=2, activation_fn=None)
self._pe = RToC.get_c(self._pe)
# Mulitiply
self._mul = tf.multiply(self.slice_output, self._pe)
# Concatenate
self._concat = tf.concat([self._mul, self._fe], 1)
self._concat = CToR.get_r(self._concat)
# RX
self._decoding = util.build_neural_net(input=self._concat, n_layers=n_layers, n_neurons=n_neurons,
activation=activation, n_outputs=output_size)
#print("self._decoding shape is: ", self._decoding.shape)
# reshape input back to (batch_size, 1, 256)
self._decoding = tf.reshape(self._decoding, [-1, 1, 256])
def __init__(self,
inputs,
n_layers,
n_neurons,
activation,
latent_size,
name='encoder'):
with tf.name_scope(name):
self._input = inputs
self._w = util.build_neural_net(input=inputs,
n_layers=n_layers,
n_neurons=n_neurons,
activation=activation,
n_outputs=latent_size)
self._encoding = tf.contrib.layers.fully_connected(
self._w, num_outputs=latent_size, activation_fn=tf.nn.relu)
#Normalization
axis = list(range(len(self._encoding.get_shape()) - 1))
mean, variance = tf.nn.moments(self._encoding, axis)
self._encoding = tf.nn.batch_normalization(self._encoding,
mean,
variance,
offset=None,
scale=0.7071,
variance_epsilon=1e-8)
#print("self._encoding shape is: ", self._encoding.shape)
# Reshape-combine 13 messages
self._encoding = tf.reshape(self._encoding,
[-1, 1, group_num * latent_size])
def __init__(self,
inputs,
n_layers,
n_neurons,
activation,
latent_size,
name='encoder'):
with tf.name_scope(name):
self._input = inputs
self._w = util.build_neural_net(input=self._input,
n_layers=n_layers,
n_neurons=n_neurons,
activation=activation,
n_outputs=latent_size)
self._encoding = tf.contrib.layers.fully_connected(
self._w, num_outputs=latent_size,
activation_fn=tf.nn.relu) # shape (?, 13, 8)
# RtoC
self._real = tf.slice(self._encoding, [0, 0, 0], [-1, -1, 4])
self._image = tf.slice(self._encoding, [0, 0, 4], [-1, -1, 4])
self._complex = tf.complex(self._real, self._image)
#print("output is: ", self._complex.shape)
# Normalization
mean, variance = tf.nn.moments(self._complex, 2, keep_dims=True)
variance = variance + 1e-8
minus_mean = tf.math.subtract(self._complex, mean)
self._encoding = tf.math.divide(minus_mean, tf.math.sqrt(variance))
#print("self._encoding shape is: ", self._encoding.shape) # shape of the output from normalization is (?, 13, 4)
# CtoR
self._real = tf.math.real(self._encoding)
self._real = tf.reshape(self._real, [-1, 1, 52])
self._image = tf.math.imag(self._encoding)
self._image = tf.reshape(self._image, [-1, 1, 52])
self._encoding = tf.concat([self._real, self._image], 2)
# self.normalized = tf.math.l2_normalize(self._encoding, axis=2)
# lenth = tf.dtypes.cast(self.normalized.get_shape()[2], tf.float32)
# sqrt_len = tf.math.sqrt(lenth)
# self._encoding = tf.multiply(sqrt_len/2, self.normalized)
# axis = list(range(len(self._encoding.get_shape()) - 1))
# #mean, variance = tf.nn.moments(self._encoding, axis)
# print("axis is: ", axis)
# self._encoding = tf.compat.v1.layers.batch_normalization(self._encoding, axis=2, center=False, scale=False, trainable=False)
#tf.nn.batch_normalization, tf.contrib.layers.batch_norm
# Reshape-combine 13 messages
#self._encoding = tf.concat([self._real, self._image], 2)
print("self._encoding shape is: ",
self._encoding.shape) # shape of the output is (?, 1, 104)
def __init__(self,
inputs,
n_layers,
n_neurons,
activation,
output_size,
name='decision'):
with tf.name_scope(name):
self._decision = util.build_neural_net(input=inputs,
n_layers=n_layers,
n_neurons=n_neurons,
activation=activation,
n_outputs=output_size)
def __init__(self,
inputs,
n_layers,
n_neurons,
activation,
latent_size,
name='encoder'):
with tf.name_scope(name):
self._encoding = util.build_neural_net(input=inputs,
n_layers=n_layers,
n_neurons=n_neurons,
activation=activation,
n_outputs=latent_size)
