def prediction(self):
# Focus network
shape = tf.shape(self.X)
channels = []
for i in range(self.S):
V = tf.gather(self.speaker_focus_vector,
self.I[:, i]) # [B, focus_dim]
V = tf.tile(tf.reshape(V, [-1, 1, self.focus_dim]),
[1, tf.shape(self.X)[-2], 1])
layers = [
BLSTM(self.layer_size, 'BLSTM_' + str(i) + '_' + str(j))
for j in range(self.nb_layers)
]
layers_sp = [
Conv1D([1, self.layer_size, self.embedding_size * self.F]),
Reshape([self.B, shape[1], self.F, self.embedding_size]),
Normalize(3)
]
layers += layers_sp
input_ = tf.concat([self.X, V], -1)
channels.append(f_props(layers, input_))
return channels
def prediction(self):
# L41 network
shape = tf.shape(self.X)
self.true_masks = 1.0 + self.y
X_in = tf.identity(self.X)
layers = [
BLSTM(self.layer_size, 'BLSTM_' + str(i))
for i in range(self.nb_layers)
]
layers_sp = [
Conv1D([1, self.layer_size, self.embedding_size * self.F]),
Reshape([self.B, shape[1], self.F, self.embedding_size])
]
if self.normalize:
layers_sp += [Normalize(3)]
layers += layers_sp
y = f_props(layers, X_in)
return y
def prediction(self):
# L41 network
self.true_masks = 1.0 + self.y
X_in = tf.identity(self.X)
if self.abs_input:
X_in = tf.abs(X_in)
if self.normalize_input == '01':
self.min_ = tf.reduce_min(X_in, axis=[1, 2], keep_dims=True)
self.max_ = tf.reduce_max(X_in, axis=[1, 2], keep_dims=True)
X_in = (X_in - self.min_) / (self.max_ - self.min_)
elif self.normalize_input == 'meanstd':
mean, var = tf.nn.moments(X_in, axes=[1, 2], keep_dims=True)
X_in = (X_in - mean) / tf.sqrt(var)
f = 128
X_in = tf.expand_dims(X_in, 3)
X_in = tf.reshape(X_in, [-1, 80, 256, 1])
y = tf.contrib.layers.conv2d(X_in, f, [1, 7], rate=[1, 1])
y = tf.contrib.layers.conv2d(y, f, [7, 1], rate=[1, 1])
y = tf.contrib.layers.conv2d(y, f, [5, 5], rate=[4, 1])
y = tf.contrib.layers.conv2d(y, f, [5, 5], rate=[8, 1])
y = tf.contrib.layers.conv2d(y, f, [5, 5], rate=[16, 1])
y = tf.contrib.layers.conv2d(y, f, [5, 5], rate=[32, 1])
y = tf.contrib.layers.conv2d(y, f, [5, 5], rate=[1, 1])
y = tf.contrib.layers.conv2d(y, f, [5, 5], rate=[2, 2])
y = tf.contrib.layers.conv2d(y, f, [5, 5], rate=[4, 4])
y = tf.contrib.layers.conv2d(y, f, [5, 5], rate=[8, 8])
y = tf.contrib.layers.conv2d(y, f, [5, 5], rate=[16, 16])
y = tf.contrib.layers.conv2d(y, f, [5, 5], rate=[32, 32])
y = tf.contrib.layers.conv2d(y, 4, [5, 5], rate=[1, 1])
y = tf.reshape(y, [self.B, 80, 256 * 4])
y = BLSTM(400, 'BLSTM_1').f_prop(y)
y = BLSTM(400, 'BLSTM_2').f_prop(y)
y = BLSTM(400, 'BLSTM_3').f_prop(y)
y = Conv1D([1, 400, self.embedding_size * self.F]).f_prop(y)
y = tf.reshape(y, [self.B, 80, self.F, self.embedding_size])
if self.normalize:
y += tf.nn.l2_normalize(y)
return y
def enhance(self):
# [B, S, T, F]
separated = tf.reshape(self.separate, [self.B, self.S, -1, self.F])
# X [B, T, F]
# Tiling the input S time - like [ a, b, c] -> [ a, a, b, b, c, c], not [a, b, c, a, b, c]
X_in = tf.expand_dims(self.X_input, 1)
X_in = tf.tile(X_in, [1, self.S, 1, 1])
X_in = tf.reshape(X_in, [self.B, self.S, -1, self.F])
# Concat the separated input and the actual tiled input
sep_and_in = tf.concat([separated, X_in], axis=3)
sep_and_in = tf.reshape(sep_and_in, [self.B * self.S, -1, 2 * self.F])
if self.args['normalize_enhance']:
mean, var = tf.nn.moments(sep_and_in, axes=[1, 2], keep_dims=True)
sep_and_in = (sep_and_in - mean) / tf.sqrt(var)
layers = [
BLSTM(self.args['layer_size_enhance'],
drop_val=self.args["recurrent_dropout_enhance"],
name='BLSTM_' + str(i))
for i in range(self.args['nb_layers_enhance'])
]
layers += [Conv1D([1, self.args['layer_size_enhance'], self.F])]
y = f_props(layers, sep_and_in)
y = tf.reshape(y, [self.B, self.S, -1]) # [B, S, TF]
tf.summary.image('mask/predicted/enhanced',
tf.reshape(y, [self.B * self.S, -1, self.F, 1]))
y = tf.transpose(y, [0, 2, 1]) # [B, TF, S]
if self.args['nonlinearity'] == 'softmax':
y = tf.nn.softmax(y)
elif self.args['nonlinearity'] == 'tanh':
y = tf.nn.tanh(y)
self.enhanced_masks = tf.identity(y, name='enhanced_masks')
tf.summary.image(
'mask/predicted/enhanced_soft',
tf.reshape(tf.transpose(y, [0, 2, 1]),
[self.B * self.S, -1, self.F, 1]))
y = y * tf.reshape(self.X_input, [
self.B, -1, 1
]) # Apply enhanced filters # [B, TF, S] -> [BS, T, F, 1]
self.cost_in = y
y = tf.transpose(y, [0, 2, 1])
self.separated = y
return tf.reshape(y, [self.B * self.S, -1, self.F, 1])
def prediction(self): # DPCL network shape = tf.shape(self.X) layers = [BLSTM(self.layer_size, 'BLSTM_'+str(i)) for i in range(self.nb_layers)] layers_sp = [ Conv1D([1, self.layer_size, self.embedding_size*self.F]), Reshape([self.B, shape[1], self.F, self.embedding_size]), Normalize(3) ] layers += layers_sp y = f_props(layers, self.X) return y
def enhance(self):
# [B, S, T, F]
separated = tf.reshape(self.separate, [self.B, self.S, -1, self.F])
min_ = tf.reduce_min(separated, axis=[1, 2], keep_dims=True)
max_ = tf.reduce_max(separated, axis=[1, 2], keep_dims=True)
separated = (separated - min_) / (max_ - min_)
# X [B, T, F]
# Tiling the input S time - like [ a, b, c] -> [ a, a, b, b, c, c], not [a, b, c, a, b, c]
X_in = tf.expand_dims(self.X, 1)
X_in = tf.tile(X_in, [1, self.S, 1, 1])
X_in = tf.reshape(X_in, [self.B, self.S, -1, self.F])
# Concat the binary separated input and the actual tiled input
sep_and_in = tf.concat([separated, X_in], axis=3)
sep_and_in = tf.reshape(sep_and_in, [self.B * self.S, -1, 2 * self.F])
layers = [
BLSTM(self.args['layer_size_enhance'], 'BLSTM_' + str(i))
for i in range(self.args['nb_layers_enhance'])
]
y = f_props(layers, sep_and_in)
y = tf.layers.dense(y, self.F)
y = tf.reshape(y, [self.B, self.S, -1]) # [B, S, TF]
y = tf.transpose(y, [0, 2, 1]) # [B, TF, S]
if self.args['nonlinearity'] == 'softmax':
y = tf.nn.softmax(y) * tf.reshape(self.X, [
self.B, -1, 1
]) # Apply enhanced filters # [B, TF, S] -> [BS, T, F, 1]
elif self.args['nonlinearity'] == 'tanh':
y = tf.nn.tanh(y) * tf.reshape(self.X, [
self.B, -1, 1
]) # Apply enhanced filters # [B, TF, S] -> [BS, T, F, 1]
# y = y * tf.reshape(self.X, [self.B, -1, 1]) # Apply enhanced filters # [B, TF, S] -> [BS, T, F, 1]
self.cost_in = y
y = tf.transpose(y, [0, 2, 1])
return tf.reshape(y, [self.B * self.S, -1, self.F, 1])
def prediction(self):
# L41 network
shape = tf.shape(self.X)
self.true_masks = 1.0 + self.y
X_in = tf.identity(self.X)
layers = [
BLSTM(self.layer_size,
name='BLSTM_' + str(i),
drop_val=self.rdropout) for i in range(self.nb_layers)
]
layers_sp = [
Conv1D([1, self.layer_size, self.embedding_size * self.F]),
Reshape([self.B, shape[1], self.F, self.embedding_size]),
]
layers += layers_sp
y = f_props(layers, X_in)
return y
def prediction(self):
# L41 network
shape = tf.shape(self.X)
self.true_masks = 1.0 + self.y
X_in = tf.identity(self.X)
if self.abs_input:
X_in = tf.abs(X_in)
if self.normalize_input == '01':
self.min_ = tf.reduce_min(X_in, axis=[1, 2], keep_dims=True)
self.max_ = tf.reduce_max(X_in, axis=[1, 2], keep_dims=True)
X_in = (X_in - self.min_) / (self.max_ - self.min_)
elif self.normalize_input == 'meanstd':
mean, var = tf.nn.moments(X_in, axes=[1, 2], keep_dims=True)
X_in = (X_in - mean) / tf.sqrt(var)
layers = [
BLSTM(self.layer_size, 'BLSTM_' + str(i))
for i in range(self.nb_layers)
]
layers_sp = [
Conv1D([1, self.layer_size, self.embedding_size * self.F]),
Reshape([self.B, shape[1], self.F, self.embedding_size])
]
if self.normalize:
layers_sp += [Normalize(3)]
layers += layers_sp
y = f_props(layers, X_in)
return y