def front(self):
# Front-End
# x : [ Btot , 1, L , 1]
# Equivalent to B_tot batches of image of height = 1, width = L and 1 channel -> for Conv1D with Conv2D
input_front = tf.reshape(self.x, [self.B_tot, 1, self.L, 1])
# Filter [filter_height, filter_width, input_channels, output_channels] = [1, W, 1, N]
# self.window_filter = get_scope_variable('window', 'w', shape=[self.window], initializer=tf.contrib.layers.xavier_initializer_conv2d())
# self.bases = get_scope_variable('bases', 'bases', shape=[self.window, self.N], initializer=tf.contrib.layers.xavier_initializer_conv2d())
# self.conv_filter = tf.reshape(tf.expand_dims(self.window_filter,1)*self.bases , [1, self.window, 1, self.N])
self.conv_filter = get_scope_variable('filters_front','filters_front', shape=[1, self.window, 1, self.N], initializer=tf.contrib.layers.xavier_initializer_conv2d())
# 1 Dimensional convolution along T axis with a window length = self.window
# And N = 256 filters -> Create a [Btot, 1, T, N]
self.X = tf.nn.conv2d(input_front, self.conv_filter, strides=[1, 1, 1, 1], padding="SAME", name='Conv_STFT')
# Reshape to Btot batches of T x N images with 1 channel
self.X = tf.reshape(self.X, [self.B_tot, -1, self.N, 1])
self.T = tf.shape(self.X)[1]
# Max Pooling with argmax for unpooling later in the back-end layer
# Along the T axis (time)
self.y, argmax = tf.nn.max_pool_with_argmax(self.X, (1, self.max_pool_value, 1, 1),
strides=[1, self.max_pool_value, 1, 1], padding="SAME", name='output')
y_shape = tf.shape(self.y)
y = tf.reshape(self.y, [self.B_tot, y_shape[1]*y_shape[2]])
self.p_hat = tf.reduce_mean(tf.abs(y), 0)
self.sparse_constraint = tf.reduce_sum(kl_div(self.p, self.p_hat))
return self.y, argmax
def front(self):
# Front-End
# x : [ Btot , 1, L , 1]
# Equivalent to B_tot batches of image of height = 1, width = L and 1 channel -> for Conv1D with Conv2D
input_front = tf.reshape(self.x, [self.B_tot, 1, self.L, 1])
# Filter [filter_height, filter_width, input_channels, output_channels] = [1, W, 1, N]
self.window_filter = get_scope_variable(
'window',
'w',
shape=[self.window],
initializer=tf.contrib.layers.xavier_initializer_conv2d())
self.bases = get_scope_variable(
'bases',
'bases',
shape=[self.window, self.N],
initializer=tf.contrib.layers.xavier_initializer_conv2d())
self.conv_filter = tf.reshape(
tf.abs(tf.expand_dims(self.window_filter, 1)) * self.bases,
[1, self.window, 1, self.N])
# self.conv_filter = get_scope_variable('filters_front','filters_front', shape=[1, self.window, 1, self.N])
variable_summaries(self.conv_filter)
# 1 Dimensional convolution along T axis with a window length = self.window
# And N = 256 filters -> Create a [Btot, 1, T, N]
self.T = tf.shape(input_front)[2]
if self.with_max_pool:
self.X = tf.nn.conv2d(input_front,
self.conv_filter,
strides=[1, 1, 1, 1],
padding="SAME",
name='Conv_STFT')
self.y, self.argmax = tf.nn.max_pool_with_argmax(
self.X, [1, 1, self.max_pool_value, 1],
strides=[1, 1, self.max_pool_value, 1],
padding="SAME",
name='output')
print self.argmax
elif self.with_average_pool:
self.X = tf.nn.conv2d(input_front,
self.conv_filter,
strides=[1, 1, 1, 1],
padding="SAME",
name='Conv_STFT')
# self.y = tf.nn.avg_pool(self.X, [1, 1, self.max_pool_value, 1], [1, 1, self.max_pool_value, 1], padding="SAME")
self.y = tf.layers.average_pooling2d(self.X,
(1, self.max_pool_value),
strides=(1,
self.max_pool_value),
name='output')
else:
self.y = tf.nn.conv2d(input_front,
self.conv_filter,
strides=[1, 1, self.max_pool_value, 1],
padding="SAME",
name='Conv_STFT')
# Reshape to Btot batches of T x N images with 1 channel
# [Btot, 1, T_pool, N] -> [Btot, T-pool, N, 1]
self.y = tf.transpose(self.y, [0, 2, 3, 1], name='output')
tf.summary.image('front/output', self.y, max_outputs=3)
y_shape = tf.shape(self.y)
y = tf.reshape(self.y, [self.B_tot, y_shape[1] * y_shape[2]])
self.p_hat = tf.reduce_sum(tf.abs(y), 0)
self.sparse_constraint = tf.reduce_sum(kl_div(self.p, self.p_hat))
return self.y
def back(self):
input_tensor = self.separator
if self.with_max_pool:
argmax_mix = self.argmax[:self.B]
repeats = [self.S, 1, 1, 1]
shape = tf.shape(argmax_mix)
argmax_mix = tf.expand_dims(argmax_mix, 1)
argmax_mix = tf.tile(argmax_mix, [1, self.S, 1, 1, 1])
argmax_mix = tf.reshape(argmax_mix, shape * repeats)
output_shape = [self.B * self.S, 1, self.T, self.N]
output = unpool(input_tensor,
argmax_mix,
ksize=[1, 1, self.max_pool_value, 1],
output_shape=output_shape,
scope='unpool')
output = tf.reshape(output, [self.B * self.S, 1, self.T, self.N])
elif self.with_average_pool:
input_tensor = tf.reshape(
input_tensor, [self.B * self.S, 1, self.T_max_pooled, self.N])
output = tf.keras.layers.UpSampling2D(
(1, self.max_pool_value))(input_tensor)
output = tf.reshape(output, [self.B * self.S, 1, self.T, self.N])
else:
output = tf.reshape(
input_tensor, [self.B * self.S, 1, self.T_max_pooled, self.N])
self.window_filter_2 = get_scope_variable(
'window',
'value',
shape=[self.window],
initializer=tf.contrib.layers.xavier_initializer_conv2d())
self.bases_2 = get_scope_variable(
'bases',
'value',
shape=[self.window, self.N],
initializer=tf.contrib.layers.xavier_initializer_conv2d())
self.conv_filter_2 = tf.reshape(
tf.abs(tf.expand_dims(self.window_filter_2, 1)) * self.bases_2,
[1, self.window, 1, self.N],
name='filters'
) # self.conv_filter_2 = tf.Variable(self.conv_filter.initialized_value(), name="filters_back")
if not (self.with_max_pool or self.with_average_pool):
strides = [1, 1, self.max_pool_value, 1]
else:
strides = [1, 1, 1, 1]
output = tf.nn.conv2d_transpose(
output,
filter=self.conv_filter_2,
output_shape=[self.B * self.S, 1, self.L, 1],
strides=strides,
padding='SAME')
output = tf.reshape(output, [self.B, self.S, self.L],
name='back_output')
self.output = output
return output
