def __init__(
self,
size,
self_attn,
src_attn,
feed_forward,
dropout_rate,
normalize_before=True,
concat_after=False,
):
"""Construct an DecoderLayer object."""
super(DecoderLayer, self).__init__()
self.size = size
self.self_attn = self_attn
self.src_attn = src_attn
self.feed_forward = feed_forward
self.norm1 = LayerNorm(size)
self.norm2 = LayerNorm(size)
self.norm3 = LayerNorm(size)
self.dropout = nn.Dropout(dropout_rate)
self.normalize_before = normalize_before
self.concat_after = concat_after
if self.concat_after:
self.concat_linear1 = nn.Linear(size + size, size)
self.concat_linear2 = nn.Linear(size + size, size)
def __init__(
self,
size,
self_attn,
feed_forward,
dropout_rate,
total_layer_num,
normalize_before=True,
concat_after=False,
):
"""Construct an EncoderLayer object."""
super(ContextualBlockEncoderLayer, self).__init__()
self.self_attn = self_attn
self.feed_forward = feed_forward
self.norm1 = LayerNorm(size)
self.norm2 = LayerNorm(size)
self.dropout = nn.Dropout(dropout_rate)
self.size = size
self.normalize_before = normalize_before
self.concat_after = concat_after
self.total_layer_num = total_layer_num
if self.concat_after:
self.concat_linear = nn.Linear(size + size, size)
def __init__(
self,
size,
self_attn,
feed_forward,
feed_forward_macaron,
conv_module,
dropout_rate,
normalize_before=True,
concat_after=False,
stochastic_depth_rate=0.0,
):
"""Construct an EncoderLayer object."""
super(EncoderLayer, self).__init__()
self.self_attn = self_attn
self.feed_forward = feed_forward
self.feed_forward_macaron = feed_forward_macaron
self.conv_module = conv_module
self.norm_ff = LayerNorm(size) # for the FNN module
self.norm_mha = LayerNorm(size) # for the MHA module
if feed_forward_macaron is not None:
self.norm_ff_macaron = LayerNorm(size)
self.ff_scale = 0.5
else:
self.ff_scale = 1.0
if self.conv_module is not None:
self.norm_conv = LayerNorm(size) # for the CNN module
self.norm_final = LayerNorm(
size) # for the final output of the block
self.dropout = nn.Dropout(dropout_rate)
self.size = size
self.normalize_before = normalize_before
self.concat_after = concat_after
if self.concat_after:
self.concat_linear = nn.Linear(size + size, size)
self.stochastic_depth_rate = stochastic_depth_rate
def __init__(self,
idim,
n_layers=2,
n_chans=384,
kernel_size=3,
dropout_rate=0.1,
offset=1.0):
"""Initilize duration predictor module.
Args:
idim (int): Input dimension.
n_layers (int, optional): Number of convolutional layers.
n_chans (int, optional): Number of channels of convolutional layers.
kernel_size (int, optional): Kernel size of convolutional layers.
dropout_rate (float, optional): Dropout rate.
offset (float, optional): Offset value to avoid nan in log domain.
"""
super(DurationPredictor, self).__init__()
self.offset = offset
self.conv = torch.nn.ModuleList()
for idx in range(n_layers):
in_chans = idim if idx == 0 else n_chans
self.conv += [
torch.nn.Sequential(
torch.nn.Conv1d(
in_chans,
n_chans,
kernel_size,
stride=1,
padding=(kernel_size - 1) // 2,
),
torch.nn.ReLU(),
LayerNorm(n_chans, dim=1),
torch.nn.Dropout(dropout_rate),
)
]
self.linear = torch.nn.Linear(n_chans, 1)
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_layer_type="linear",
positionwise_conv_kernel_size=1,
macaron_style=False,
pos_enc_layer_type="abs_pos",
selfattention_layer_type="selfattn",
activation_type="swish",
use_cnn_module=False,
zero_triu=False,
cnn_module_kernel=31,
padding_idx=-1,
stochastic_depth_rate=0.0,
intermediate_layers=None,
):
"""Construct an Encoder object."""
super(Encoder, self).__init__()
activation = get_activation(activation_type)
if pos_enc_layer_type == "abs_pos":
pos_enc_class = PositionalEncoding
elif pos_enc_layer_type == "scaled_abs_pos":
pos_enc_class = ScaledPositionalEncoding
elif pos_enc_layer_type == "rel_pos":
assert selfattention_layer_type == "rel_selfattn"
pos_enc_class = RelPositionalEncoding
elif pos_enc_layer_type == "legacy_rel_pos":
pos_enc_class = LegacyRelPositionalEncoding
assert selfattention_layer_type == "legacy_rel_selfattn"
else:
raise ValueError("unknown pos_enc_layer: " + pos_enc_layer_type)
self.conv_subsampling_factor = 1
if input_layer == "linear":
self.embed = torch.nn.Sequential(
torch.nn.Linear(idim, attention_dim),
torch.nn.LayerNorm(attention_dim),
torch.nn.Dropout(dropout_rate),
pos_enc_class(attention_dim, positional_dropout_rate),
)
elif input_layer == "conv2d":
self.embed = Conv2dSubsampling(
idim,
attention_dim,
dropout_rate,
pos_enc_class(attention_dim, positional_dropout_rate),
)
self.conv_subsampling_factor = 4
elif input_layer == "vgg2l":
self.embed = VGG2L(idim, attention_dim)
self.conv_subsampling_factor = 4
elif input_layer == "embed":
self.embed = torch.nn.Sequential(
torch.nn.Embedding(idim, attention_dim, padding_idx=padding_idx),
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),
)
elif input_layer is None:
self.embed = torch.nn.Sequential(
pos_enc_class(attention_dim, positional_dropout_rate)
)
else:
raise ValueError("unknown input_layer: " + input_layer)
self.normalize_before = normalize_before
# self-attention module definition
if selfattention_layer_type == "selfattn":
logging.info("encoder self-attention layer type = self-attention")
encoder_selfattn_layer = MultiHeadedAttention
encoder_selfattn_layer_args = (
attention_heads,
attention_dim,
attention_dropout_rate,
)
elif selfattention_layer_type == "legacy_rel_selfattn":
assert pos_enc_layer_type == "legacy_rel_pos"
encoder_selfattn_layer = LegacyRelPositionMultiHeadedAttention
encoder_selfattn_layer_args = (
attention_heads,
attention_dim,
attention_dropout_rate,
)
elif selfattention_layer_type == "rel_selfattn":
logging.info("encoder self-attention layer type = relative self-attention")
assert pos_enc_layer_type == "rel_pos"
encoder_selfattn_layer = RelPositionMultiHeadedAttention
encoder_selfattn_layer_args = (
attention_heads,
attention_dim,
attention_dropout_rate,
zero_triu,
)
else:
raise ValueError("unknown encoder_attn_layer: " + selfattention_layer_type)
# feed-forward module definition
if positionwise_layer_type == "linear":
positionwise_layer = PositionwiseFeedForward
positionwise_layer_args = (
attention_dim,
linear_units,
dropout_rate,
activation,
)
elif positionwise_layer_type == "conv1d":
positionwise_layer = MultiLayeredConv1d
positionwise_layer_args = (
attention_dim,
linear_units,
positionwise_conv_kernel_size,
dropout_rate,
)
elif positionwise_layer_type == "conv1d-linear":
positionwise_layer = Conv1dLinear
positionwise_layer_args = (
attention_dim,
linear_units,
positionwise_conv_kernel_size,
dropout_rate,
)
else:
raise NotImplementedError("Support only linear or conv1d.")
# 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,
stochastic_depth_rate * float(1 + lnum) / num_blocks,
),
)
if self.normalize_before:
self.after_norm = LayerNorm(attention_dim)
self.intermediate_layers = intermediate_layers
def __init__(
self,
idim,
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,
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,
selfattention_layer_type="selfattn",
padding_idx=-1,
):
"""Construct an Encoder object."""
super(Encoder, self).__init__()
self._register_load_state_dict_pre_hook(_pre_hook)
self.conv_subsampling_factor = 1
if input_layer == "linear":
self.embed = torch.nn.Sequential(
torch.nn.Linear(idim, attention_dim),
torch.nn.LayerNorm(attention_dim),
torch.nn.Dropout(dropout_rate),
torch.nn.ReLU(),
pos_enc_class(attention_dim, positional_dropout_rate),
)
elif input_layer == "conv2d":
self.embed = Conv2dSubsampling(idim, attention_dim, dropout_rate)
self.conv_subsampling_factor = 4
elif input_layer == "conv2d-scaled-pos-enc":
self.embed = Conv2dSubsampling(
idim,
attention_dim,
dropout_rate,
pos_enc_class(attention_dim, positional_dropout_rate),
)
self.conv_subsampling_factor = 4
elif input_layer == "conv2d6":
self.embed = Conv2dSubsampling6(idim, attention_dim, dropout_rate)
self.conv_subsampling_factor = 6
elif input_layer == "conv2d8":
self.embed = Conv2dSubsampling8(idim, attention_dim, dropout_rate)
self.conv_subsampling_factor = 8
elif input_layer == "vgg2l":
self.embed = VGG2L(idim, attention_dim)
self.conv_subsampling_factor = 4
elif input_layer == "embed":
self.embed = torch.nn.Sequential(
torch.nn.Embedding(idim, attention_dim, padding_idx=padding_idx),
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),
)
elif input_layer is None:
self.embed = torch.nn.Sequential(
pos_enc_class(attention_dim, positional_dropout_rate)
)
else:
raise ValueError("unknown input_layer: " + input_layer)
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,
)
if selfattention_layer_type in [
"selfattn",
"rel_selfattn",
"legacy_rel_selfattn",
]:
logging.info("encoder self-attention layer type = self-attention")
encoder_selfattn_layer = MultiHeadedAttention
encoder_selfattn_layer_args = [
(
attention_heads,
attention_dim,
attention_dropout_rate,
)
] * num_blocks
elif selfattention_layer_type == "lightconv":
logging.info("encoder self-attention layer type = lightweight convolution")
encoder_selfattn_layer = LightweightConvolution
encoder_selfattn_layer_args = [
(
conv_wshare,
attention_dim,
attention_dropout_rate,
int(conv_kernel_length.split("_")[lnum]),
False,
conv_usebias,
)
for lnum in range(num_blocks)
]
elif selfattention_layer_type == "lightconv2d":
logging.info(
"encoder self-attention layer "
"type = lightweight convolution 2-dimentional"
)
encoder_selfattn_layer = LightweightConvolution2D
encoder_selfattn_layer_args = [
(
conv_wshare,
attention_dim,
attention_dropout_rate,
int(conv_kernel_length.split("_")[lnum]),
False,
conv_usebias,
)
for lnum in range(num_blocks)
]
elif selfattention_layer_type == "dynamicconv":
logging.info("encoder self-attention layer type = dynamic convolution")
encoder_selfattn_layer = DynamicConvolution
encoder_selfattn_layer_args = [
(
conv_wshare,
attention_dim,
attention_dropout_rate,
int(conv_kernel_length.split("_")[lnum]),
False,
conv_usebias,
)
for lnum in range(num_blocks)
]
elif selfattention_layer_type == "dynamicconv2d":
logging.info(
"encoder self-attention layer type = dynamic convolution 2-dimentional"
)
encoder_selfattn_layer = DynamicConvolution2D
encoder_selfattn_layer_args = [
(
conv_wshare,
attention_dim,
attention_dropout_rate,
int(conv_kernel_length.split("_")[lnum]),
False,
conv_usebias,
)
for lnum in range(num_blocks)
]
else:
raise NotImplementedError(selfattention_layer_type)
self.encoders = repeat(
num_blocks,
lambda lnum: EncoderLayer(
attention_dim,
encoder_selfattn_layer(*encoder_selfattn_layer_args[lnum]),
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)
self._register_load_state_dict_pre_hook(_pre_hook)
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-attention module definition
if selfattention_layer_type == "selfattn":
logging.info("decoder self-attention layer type = self-attention")
decoder_selfattn_layer = MultiHeadedAttention
decoder_selfattn_layer_args = [
(
attention_heads,
attention_dim,
self_attention_dropout_rate,
)
] * num_blocks
elif selfattention_layer_type == "lightconv":
logging.info("decoder self-attention layer type = lightweight convolution")
decoder_selfattn_layer = LightweightConvolution
decoder_selfattn_layer_args = [
(
conv_wshare,
attention_dim,
self_attention_dropout_rate,
int(conv_kernel_length.split("_")[lnum]),
True,
conv_usebias,
)
for lnum in range(num_blocks)
]
elif selfattention_layer_type == "lightconv2d":
logging.info(
"decoder self-attention layer "
"type = lightweight convolution 2-dimensional"
)
decoder_selfattn_layer = LightweightConvolution2D
decoder_selfattn_layer_args = [
(
conv_wshare,
attention_dim,
self_attention_dropout_rate,
int(conv_kernel_length.split("_")[lnum]),
True,
conv_usebias,
)
for lnum in range(num_blocks)
]
elif selfattention_layer_type == "dynamicconv":
logging.info("decoder self-attention layer type = dynamic convolution")
decoder_selfattn_layer = DynamicConvolution
decoder_selfattn_layer_args = [
(
conv_wshare,
attention_dim,
self_attention_dropout_rate,
int(conv_kernel_length.split("_")[lnum]),
True,
conv_usebias,
)
for lnum in range(num_blocks)
]
elif selfattention_layer_type == "dynamicconv2d":
logging.info(
"decoder self-attention layer type = dynamic convolution 2-dimensional"
)
decoder_selfattn_layer = DynamicConvolution2D
decoder_selfattn_layer_args = [
(
conv_wshare,
attention_dim,
self_attention_dropout_rate,
int(conv_kernel_length.split("_")[lnum]),
True,
conv_usebias,
)
for lnum in range(num_blocks)
]
self.decoders = repeat(
num_blocks,
lambda lnum: DecoderLayer(
attention_dim,
decoder_selfattn_layer(*decoder_selfattn_layer_args[lnum]),
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