def __init__(self,
idim,
selfattention_layer_type="selfattn",
attention_dim=256,
attention_heads=4,
linear_units=2048,
num_blocks=6,
dropout_rate=0.1,
positional_dropout_rate=0.1,
attention_dropout_rate=0.0,
input_layer="conv2d",
pos_enc_class=PositionalEncoding,
normalize_before=True,
concat_after=False,
positionwise_layer_type="linear",
positionwise_conv_kernel_size=1):
super(Encoder, self).__init__()
self.conv_subsampling_factor = 1
self.embed = torch.nn.Sequential(
input_layer,
pos_enc_class(attention_dim, positional_dropout_rate),
)
self.normalize_before = normalize_before
positionwise_layer, positionwise_layer_args = self.get_positionwise_layer(
positionwise_layer_type,
attention_dim,
linear_units,
dropout_rate,
positionwise_conv_kernel_size,
)
self.encoders = repeat(
num_blocks,
lambda lnum: EncoderLayer(
attention_dim,
MultiHeadedAttention(attention_heads, attention_dim,
attention_dropout_rate),
positionwise_layer(*positionwise_layer_args),
dropout_rate,
normalize_before,
concat_after,
),
)
if self.normalize_before:
self.after_norm = LayerNorm(attention_dim)
def __init__(
self,
odim,
selfattention_layer_type="selfattn",
attention_dim=256,
attention_heads=4,
conv_wshare=4,
conv_kernel_length=11,
conv_usebias=False,
linear_units=2048,
num_blocks=6,
dropout_rate=0.1,
positional_dropout_rate=0.1,
self_attention_dropout_rate=0.0,
src_attention_dropout_rate=0.0,
input_layer="embed",
use_output_layer=True,
pos_enc_class=PositionalEncoding,
normalize_before=True,
concat_after=False,
):
"""Construct an Decoder object."""
torch.nn.Module.__init__(self)
if input_layer == "embed":
self.embed = torch.nn.Sequential(
torch.nn.Embedding(odim, attention_dim),
pos_enc_class(attention_dim, positional_dropout_rate),
)
elif input_layer == "linear":
self.embed = torch.nn.Sequential(
torch.nn.Linear(odim, attention_dim),
torch.nn.LayerNorm(attention_dim),
torch.nn.Dropout(dropout_rate),
torch.nn.ReLU(),
pos_enc_class(attention_dim, positional_dropout_rate),
)
elif isinstance(input_layer, torch.nn.Module):
self.embed = torch.nn.Sequential(
input_layer, pos_enc_class(attention_dim, positional_dropout_rate)
)
else:
raise NotImplementedError("only `embed` or torch.nn.Module is supported.")
self.normalize_before = normalize_before
self.decoders = repeat(
num_blocks,
lambda lnum: DecoderLayer(
attention_dim,
MultiHeadedAttention(attention_heads, attention_dim, self_attention_dropout_rate),
MultiHeadedAttention(attention_heads, attention_dim, src_attention_dropout_rate),
PositionwiseFeedForward(attention_dim, linear_units, dropout_rate),
dropout_rate,
normalize_before,
concat_after,
),
)
self.selfattention_layer_type = selfattention_layer_type
if self.normalize_before:
self.after_norm = LayerNorm(attention_dim)
if use_output_layer:
self.output_layer = torch.nn.Linear(attention_dim, odim)
else:
self.output_layer = None
def __init__(self,
idim,
attention_dim=256,
attention_heads=4,
linear_units=2048,
num_blocks=6,
dropout_rate=0.1,
positional_dropout_rate=0.1,
attention_dropout_rate=0.0,
input_layer="conv2d",
normalize_before=True,
concat_after=False,
positionwise_conv_kernel_size=1,
macaron_style=False,
use_cnn_module=False,
cnn_module_kernel=31,
zero_triu=False):
"""Construct a Conformer object."""
super(Conformer, self).__init__()
activation = Swish()
self.conv_subsampling_factor = 1
if isinstance(input_layer, torch.nn.Module):
self.embed = torch.nn.Sequential(
input_layer,
RelPositionalEncoding(attention_dim, positional_dropout_rate),
)
elif input_layer is None:
self.embed = torch.nn.Sequential(
RelPositionalEncoding(attention_dim, positional_dropout_rate))
else:
raise ValueError("unknown input_layer: " + input_layer)
self.normalize_before = normalize_before
# self-attention module definition
encoder_selfattn_layer = RelPositionMultiHeadedAttention
encoder_selfattn_layer_args = (attention_heads, attention_dim,
attention_dropout_rate, zero_triu)
# feed-forward module definition
positionwise_layer = MultiLayeredConv1d
positionwise_layer_args = (
attention_dim,
linear_units,
positionwise_conv_kernel_size,
dropout_rate,
)
# convolution module definition
convolution_layer = ConvolutionModule
convolution_layer_args = (attention_dim, cnn_module_kernel, activation)
self.encoders = repeat(
num_blocks, lambda lnum: EncoderLayer(
attention_dim,
encoder_selfattn_layer(*encoder_selfattn_layer_args),
positionwise_layer(*positionwise_layer_args),
positionwise_layer(*positionwise_layer_args)
if macaron_style else None,
convolution_layer(*convolution_layer_args) if use_cnn_module
else None, dropout_rate, normalize_before, concat_after))
if self.normalize_before:
self.after_norm = LayerNorm(attention_dim)