centaur

attention

class parts.centaur.attention.AttentionBlock(hidden_size, attention_dropout, layer_postprocess_dropout, training, cnn_dropout_prob, regularizer=None, conv_params=None, n_heads=1, window_size=None, back_step_size=None, name='attention_block')

Bases: object

Attention block for Centaur model.

__init__(hidden_size, attention_dropout, layer_postprocess_dropout, training, cnn_dropout_prob, regularizer=None, conv_params=None, n_heads=1, window_size=None, back_step_size=None, name='attention_block')

Attention block constructor.

Parameters:

• hidden_size – dimensionality of hidden embeddings.
• attention_dropout – dropout rate for attention layer.
• layer_postprocess_dropout – dropout rate for sublayer.
• training – whether it is training mode.
• cnn_dropout_prob – dropout probabilty for cnn layers.
• regularizer – regularizer for the convolution kernel.
• conv_params – description of convolutional layer.
• n_heads – number of attention heads. Defaults to 1.
• window_size – size of attention window for forcing monotonic attention during the inference. Defaults to None.
• back_step_size – number of steps attention is allowed to go back during the inference. Defaults to 0.
• name – name of the block.

batch_norm

class parts.centaur.batch_norm.BatchNorm1D(*args, **kwargs)

Bases: object

1D batch normalization layer.

conv_block

class parts.centaur.conv_block.ConvBlock(name, conv, norm, activation_fn, dropout, training, is_residual, is_causal)

Bases: object

Convolutional block for Centaur model.

__init__(name, conv, norm, activation_fn, dropout, training, is_residual, is_causal)

Convolutional block constructor.

Parameters:

• name – name of the block.
• conv – convolutional layer.
• norm – normalization layer to use after the convolutional layer.
• activation_fn – activation function to use after the normalization.
• dropout – dropout rate.
• training – whether it is training mode.
• is_residual – whether the block should contain a residual connection.
• is_causal – whether the convolutional layer should be causal.

static create(index, conv_params, regularizer, bn_momentum, bn_epsilon, cnn_dropout_prob, training, is_residual=True, is_causal=False)

prenet

class parts.centaur.prenet.Prenet(n_layers, hidden_size, activation_fn, dropout=0.5, regularizer=None, training=True, dtype=None, name='prenet')

Bases: object

Centaur decoder pre-net.

__init__(n_layers, hidden_size, activation_fn, dropout=0.5, regularizer=None, training=True, dtype=None, name='prenet')

Pre-net constructor.

Parameters:

• n_layers – number of fully-connected layers to use.
• hidden_size – number of units in each pre-net layer.
• activation_fn – activation function to use.
• dropout – dropout rate. Defaults to 0.5.
• regularizer – regularizer for the convolution kernel. Defaults to None.
• training – whether it is training mode. Defaults to None.
• dtype – dtype of the layer’s weights. Defaults to None.
• name – name of the block.