def __init__(self, model_dim: int, feed_forward_dimension: int,
num_heads: list, maximum_position_encoding: int,
conv_filters: int, dropout_rate: float, dense_blocks: int,
kernel_size: int, conv_activation: str, **kwargs):
super(SelfAttentionBlocks, self).__init__(**kwargs)
self.model_dim = model_dim
self.maximum_position_encoding = maximum_position_encoding
self.pos_encoding_scalar = tf.Variable(1.)
self.pos_encoding = positional_encoding(maximum_position_encoding,
model_dim)
self.dropout = tf.keras.layers.Dropout(dropout_rate)
self.encoder_SADB = [
SelfAttentionDenseBlock(model_dim=model_dim,
dropout_rate=dropout_rate,
num_heads=n_heads,
dense_hidden_units=feed_forward_dimension,
name=f'{self.name}_SADB_IN_{i}')
for i, n_heads in enumerate(num_heads[:dense_blocks])
]
self.encoder_SACB = [
SelfAttentionConvBlock(model_dim=model_dim,
dropout_rate=dropout_rate,
num_heads=n_heads,
name=f'{self.name}_SACB_IN_{i}',
kernel_size=kernel_size,
conv_activation=conv_activation,
conv_filters=conv_filters)
for i, n_heads in enumerate(num_heads[dense_blocks:])
]
self.seq_resnet = FFNResNorm(model_dim=model_dim,
dense_hidden_units=model_dim,
dropout_rate=dropout_rate)
def __init__(self,
model_dim: int,
feed_forward_dimension: int,
num_heads: list,
maximum_position_encoding: int,
dropout_rate: float,
dense_blocks: int,
conv_filters: int,
conv_activation: str,
conv_padding: str,
conv_kernel: int,
**kwargs):
super(CrossAttentionBlocks, self).__init__(**kwargs)
self.model_dim = model_dim
self.pos_encoding_scalar = tf.Variable(1.)
self.pos_encoding = positional_encoding(maximum_position_encoding, model_dim)
self.dropout = tf.keras.layers.Dropout(dropout_rate)
self.CADB = [
CrossAttentionDenseBlock(model_dim=model_dim, dropout_rate=dropout_rate, num_heads=n_heads,
dense_hidden_units=feed_forward_dimension, name=f'{self.name}_CADB_{i}')
for i, n_heads in enumerate(num_heads[:dense_blocks])]
self.CACB = [
CrossAttentionConvBlock(model_dim=model_dim, dropout_rate=dropout_rate, num_heads=n_heads,
name=f'{self.name}_CACB_{i}', conv_filters=conv_filters,
conv_activation=conv_activation, conv_padding=conv_padding, kernel_size=conv_kernel)
for i, n_heads in enumerate(num_heads[dense_blocks:])]
def __init__(self,
model_dim: int,
feed_forward_dimension: int,
num_heads: list,
maximum_position_encoding: int,
dropout_rate=0.1,
dense_blocks=1,
**kwargs):
super(SelfAttentionBlocks, self).__init__(**kwargs)
self.model_dim = model_dim
self.pos_encoding_scalar = tf.Variable(1.)
self.pos_encoding = positional_encoding(maximum_position_encoding,
model_dim)
self.dropout = tf.keras.layers.Dropout(dropout_rate)
self.encoder_SADB = [
SelfAttentionDenseBlock(model_dim=model_dim,
dropout_rate=dropout_rate,
num_heads=n_heads,
dense_hidden_units=feed_forward_dimension,
name=f'{self.name}_SADB_{i}')
for i, n_heads in enumerate(num_heads[:dense_blocks])
]
self.encoder_SACB = [
SelfAttentionConvBlock(model_dim=model_dim,
dropout_rate=dropout_rate,
num_heads=n_heads,
name=f'{self.name}_SACB_{i}')
for i, n_heads in enumerate(num_heads[dense_blocks:])
]
def call(self, inputs, training, fb_switch, padding_mask, min_index,
random_padding_mask, drop_n_heads):
shift_pos_encoding = positional_encoding(
self.maximum_position_encoding,
self.model_dim,
start_index=min_index)
seq_len = tf.shape(inputs)[1]
x = inputs * tf.math.sqrt(tf.cast(self.model_dim, tf.float32))
x1 = x + self.pos_encoding_scalar * shift_pos_encoding[:, :seq_len, :]
x2 = x + self.pos_encoding_scalar * self.pos_encoding[:, :seq_len, :]
x1 = self.dropout(x1, training=training)
x2 = self.dropout(x2, training=training)
random_mask = tf.maximum(
tf.cast(padding_mask, tf.float32),
tf.cast(random_padding_mask[:, tf.newaxis, tf.newaxis, :],
tf.float32))
attention_weights = {}
for i, block in enumerate(self.encoder_SACB):
x1, x2, attn_weights1, attn_weights2 = block(
x1,
x2,
training=training,
padding_mask=padding_mask,
random_mask=random_mask,
drop_n_heads=drop_n_heads)
attention_weights[
f'{self.name}_ConvBlock{i + 1}_SelfAttention1'] = attn_weights1
attention_weights[
f'{self.name}_ConvBlock{i + 1}_SelfAttention2'] = attn_weights2
for i, block in enumerate(self.encoder_SADB):
x1, x2, attn_weights1, attn_weights2 = block(
x1,
x2,
training=training,
padding_mask=padding_mask,
random_mask=random_mask,
drop_n_heads=drop_n_heads)
attention_weights[
f'{self.name}_DenseBlock{i + 1}_SelfAttention1'] = attn_weights1
attention_weights[
f'{self.name}_DenseBlock{i + 1}_SelfAttention2'] = attn_weights2
x1 = self.spk_resnet(x1)
if (fb_switch < 0.5):
x1 = tf.reverse(x1, axis=[-1])
x1 = self.spk_rnn(x1)
x1 = tf.nn.l2_normalize(x1, 1)
x2 = self.seq_resnet(x2)
return x1, x2, attention_weights
def __init__(self, num_heads, d_model, dff, output_vocab_size, maximum_position_encoding, layernorm, num_layers = 4, rate= 0.1, **kwargs):
super(Decoder, self).__init__(**kwargs)
self.d_model = d_model
self.num_layers = num_layers
self.num_heads = num_heads
self._layernorm= layernorm
self.embedding = tf.keras.layers.Embedding(output_vocab_size, d_model)
self.pos_encoding = positional_encoding(maximum_position_encoding, self.d_model)
self.layernorm = tf.keras.layers.LayerNormalization(epsilon= 1e-6)
self.dec_layers = [DecoderLayer(num_heads, dff, d_model, rate) for _ in range(num_layers)]
self.dropout = tf.keras.layers.Dropout(rate)
def __init__(self, model_dim: int, feed_forward_dimension: int,
num_heads: list, maximum_position_encoding: int,
dropout_rate: float, **kwargs):
super(CrossAttentionBlocks, self).__init__(**kwargs)
self.model_dim = model_dim
self.pos_encoding_scalar = tf.Variable(1.)
self.pos_encoding = positional_encoding(maximum_position_encoding,
model_dim)
self.dropout = tf.keras.layers.Dropout(dropout_rate)
self.CADB = [
CrossAttentionDenseBlock(model_dim=model_dim,
dropout_rate=dropout_rate,
num_heads=n_heads,
dense_hidden_units=feed_forward_dimension,
name=f'{self.name}_CADB_{i}')
for i, n_heads in enumerate(num_heads[:-1])
]
self.last_CADB = CrossAttentionDenseBlock(
model_dim=model_dim,
dropout_rate=dropout_rate,
num_heads=num_heads[-1],
dense_hidden_units=feed_forward_dimension,
name=f'{self.name}_CADB_last')
def __init__(self,
model_dim: int,
feed_forward_dimension: int,
num_heads: list,
maximum_position_encoding: int,
conv_filters: int,
dropout_rate: float,
dense_blocks: int,
kernel_size: int,
conv_activation: str,
**kwargs):
super(SelfAttentionBlocks, self).__init__(**kwargs) # initialize the parent class (=keras.layers.Layer)
self.model_dim = model_dim # initialize model dimension
self.pos_encoding_scalar = tf.Variable(1.) # position of the last element
self.pos_encoding = positional_encoding(maximum_position_encoding, model_dim) # positional encoding (triangular encoding)
self.dropout = tf.keras.layers.Dropout(dropout_rate) # droput layer for regularization
# encoder using dense-type attention mechanism
self.encoder_SADB = [
SelfAttentionDenseBlock(model_dim=model_dim, dropout_rate=dropout_rate, num_heads=n_heads,
dense_hidden_units=feed_forward_dimension, name=f'{self.name}_SADB_{i}')
for i, n_heads in enumerate(num_heads[:dense_blocks])] # some heads will be dedicated to dense-type attention: num_heads[:dense_blocks]
# encoder using convolutional-type attention mechanism
self.encoder_SACB = [
SelfAttentionConvBlock(model_dim=model_dim, dropout_rate=dropout_rate, num_heads=n_heads,
name=f'{self.name}_SACB_{i}', kernel_size=kernel_size,
conv_activation=conv_activation, conv_filters=conv_filters)
for i, n_heads in enumerate(num_heads[dense_blocks:])] # some heads will be dedicated to conv-type attention: num_heads[dense_blocks:]
