def __init__(self,
idim,
time_len=8,
mem_len=0,
ext_len=0,
future_len=0,
attention_type="memory",
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):
super(Encoder, self).__init__()
self.idim = idim
self.time_len = time_len
self.future_len = future_len
self.attention_dim = attention_dim
self.attention_heads = attention_heads
self.linear_units = linear_units
self.dropout_rate = dropout_rate
self.input_layer = input_layer
self.normalize_before = normalize_before
self.concat_after = concat_after
self.attention_type = attention_type
self.positional_dropout_rate = positional_dropout_rate
self.pos_enc_class = pos_enc_class
self._generateInputLayer()
if attention_type == "memory":
self.encoders = repeat(
num_blocks, lambda:
EncoderLayerXL(n_head=attention_heads,
d_model=attention_dim,
d_head=attention_dim // attention_heads,
ext_len=ext_len,
mem_len=mem_len,
future_len=future_len,
dropout=dropout_rate,
dropatt=attention_dropout_rate,
pre_lnorm=normalize_before,
pos_ff=PositionwiseFeedForward(
attention_dim, linear_units, dropout_rate)))
elif attention_type == "traditional":
self.encoders = repeat(
num_blocks, lambda: EncoderLayerTD(
attention_dim,
MultiHeadedAttention(attention_heads, attention_dim,
attention_dropout_rate),
PositionwiseFeedForward(attention_dim, linear_units,
dropout_rate), dropout_rate,
normalize_before, concat_after))
else:
ValueError("only memory or traditional can be used")
if self.normalize_before:
self.after_norm = LayerNorm(attention_dim)
def __init__(
self,
odim,
jdim,
attention_dim=512,
attention_heads=4,
linear_units=2048,
num_blocks=6,
dropout_rate=0.1,
positional_dropout_rate=0.0,
attention_dropout_rate=0.0,
input_layer="embed",
pos_enc_class=PositionalEncoding,
blank=0,
):
"""Construct a Decoder object for transformer-transducer models."""
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.decoders = repeat(
num_blocks,
lambda lnum: DecoderLayer(
attention_dim,
MultiHeadedAttention(
attention_heads, attention_dim, attention_dropout_rate
),
PositionwiseFeedForward(attention_dim, linear_units, dropout_rate),
dropout_rate,
),
)
self.after_norm = LayerNorm(attention_dim)
self.lin_enc = torch.nn.Linear(attention_dim, jdim)
self.lin_dec = torch.nn.Linear(attention_dim, jdim, bias=False)
self.lin_out = torch.nn.Linear(jdim, odim)
self.attention_dim = attention_dim
self.odim = odim
self.blank = blank
def __init__(self, idim, args):
super(Encoder, self).__init__()
if args.transformer_input_layer == "linear":
self.input_layer = torch.nn.Sequential(
torch.nn.Linear(idim, args.adim),
torch.nn.LayerNorm(args.adim),
torch.nn.Dropout(args.dropout_rate), torch.nn.ReLU(),
PositionalEncoding(args.adim, args.dropout_rate))
elif args.transformer_input_layer == "conv2d":
self.input_layer = Conv2dSubsampling(idim, args.adim,
args.dropout_rate)
elif args.transformer_input_layer == "embed":
self.input_layer = torch.nn.Sequential(
torch.nn.Embedding(idim, args.adim),
PositionalEncoding(args.adim, args.dropout_rate))
else:
raise ValueError("unknown input_layer: " +
args.transformer_input_layer)
self.encoders = repeat(
args.elayers, lambda: EncoderLayer(
args.adim,
MultiHeadedAttention(args.aheads, args.adim, args.
transformer_attn_dropout_rate),
PositionwiseFeedForward(args.adim, args.eunits, args.
dropout_rate), args.dropout_rate))
self.norm = LayerNorm(args.adim)
def __init__(
self,
vocab_size: int,
encoder_output_size: int,
attention_heads: int = 4,
linear_units: int = 2048,
num_blocks: int = 6,
dropout_rate: float = 0.1,
positional_dropout_rate: float = 0.1,
self_attention_dropout_rate: float = 0.0,
src_attention_dropout_rate: float = 0.0,
input_layer: str = "embed",
use_output_layer: bool = True,
pos_enc_class=PositionalEncoding,
normalize_before: bool = True,
concat_after: bool = False,
):
assert check_argument_types()
super().__init__()
attention_dim = encoder_output_size
if input_layer == "embed":
self.embed = torch.nn.Sequential(
torch.nn.Embedding(vocab_size, attention_dim),
pos_enc_class(attention_dim, positional_dropout_rate),
)
elif input_layer == "linear":
self.embed = torch.nn.Sequential(
torch.nn.Linear(vocab_size, attention_dim),
torch.nn.LayerNorm(attention_dim),
torch.nn.Dropout(dropout_rate),
torch.nn.ReLU(),
pos_enc_class(attention_dim, positional_dropout_rate),
)
else:
raise ValueError(
f"only 'embed' or 'linear' is supported: {input_layer}")
self.normalize_before = normalize_before
self.decoders = repeat(
num_blocks,
lambda: 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,
),
)
if self.normalize_before:
self.after_norm = LayerNorm(attention_dim)
if use_output_layer:
self.output_layer = torch.nn.Linear(attention_dim, vocab_size)
else:
self.output_layer = None
def __init__(self,
idim=83,
d_model=256,
n_heads=4,
d_ffn=2048,
layers=6,
kernel_size=32,
dropout_rate=0.1,
input_layer="conv2d",
padding_idx=-1,
causal=False):
super(Encoder, self).__init__()
if input_layer == "custom":
self.embed = EncoderConv2d(idim, d_model)
elif input_layer == "conv2d":
self.embed = Conv2dSubsampling(idim, d_model, dropout_rate)
else:
raise ValueError("unknown input_layer: " + input_layer)
self.pos_emb = RelativePositionalEncoding(d_model // n_heads, 1000,
False)
self.encoders = repeat(
layers, lambda: EncoderLayer(d_model, n_heads, d_ffn, kernel_size,
dropout_rate, causal))
def __init__(
self,
vocab_size: int,
encoder_output_size: int,
attention_heads: int = 4,
linear_units: int = 2048,
num_blocks: int = 6,
dropout_rate: float = 0.1,
positional_dropout_rate: float = 0.1,
self_attention_dropout_rate: float = 0.0,
src_attention_dropout_rate: float = 0.0,
input_layer: str = "embed",
use_output_layer: bool = True,
pos_enc_class=PositionalEncoding,
normalize_before: bool = True,
concat_after: bool = False,
conv_wshare: int = 4,
conv_kernel_length: Sequence[int] = (11, 11, 11, 11, 11, 11),
conv_usebias: int = False,
):
assert check_argument_types()
if len(conv_kernel_length) != num_blocks:
raise ValueError(
"conv_kernel_length must have equal number of values to num_blocks: "
f"{len(conv_kernel_length)} != {num_blocks}")
super().__init__(
vocab_size=vocab_size,
encoder_output_size=encoder_output_size,
dropout_rate=dropout_rate,
positional_dropout_rate=positional_dropout_rate,
input_layer=input_layer,
use_output_layer=use_output_layer,
pos_enc_class=pos_enc_class,
normalize_before=normalize_before,
)
attention_dim = encoder_output_size
self.decoders = repeat(
num_blocks,
lambda lnum: DecoderLayer(
attention_dim,
DynamicConvolution2D(
wshare=conv_wshare,
n_feat=attention_dim,
dropout_rate=self_attention_dropout_rate,
kernel_size=conv_kernel_length[lnum],
use_kernel_mask=True,
use_bias=conv_usebias,
),
MultiHeadedAttention(attention_heads, attention_dim,
src_attention_dropout_rate),
PositionwiseFeedForward(attention_dim, linear_units,
dropout_rate),
dropout_rate,
normalize_before,
concat_after,
),
)
def __init__(self,
odim,
attention_dim=256,
attention_heads=4,
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,
moe_att_mode='linear'):
"""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: HANDecoderLayer(
attention_dim,
MultiHeadedAttention(attention_heads, attention_dim,
self_attention_dropout_rate),
MultiHeadedAttention(attention_heads, attention_dim,
src_attention_dropout_rate),
MultiHeadedAttention(attention_heads, attention_dim,
src_attention_dropout_rate),
PositionwiseFeedForward(attention_dim, linear_units,
dropout_rate),
dropout_rate=dropout_rate,
moe_att_mode=moe_att_mode,
normalize_before=normalize_before,
concat_after=concat_after,
))
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,
vocab_size: int,
encoder_output_size: int,
attention_heads: int = 4,
linear_units: int = 2048,
num_blocks: int = 6,
dropout_rate: float = 0.1,
positional_dropout_rate: float = 0.1,
self_attention_dropout_rate: float = 0.0,
src_attention_dropout_rate: float = 0.0,
input_layer: str = "embed",
use_output_layer: bool = True,
pos_enc_class=PositionalEncoding,
normalize_before: bool = True,
concat_after: bool = False,
):
assert check_argument_types()
super().__init__(
vocab_size=vocab_size,
encoder_output_size=encoder_output_size,
dropout_rate=dropout_rate,
positional_dropout_rate=positional_dropout_rate,
input_layer=input_layer,
use_output_layer=use_output_layer,
pos_enc_class=pos_enc_class,
normalize_before=normalize_before,
)
attention_dim = encoder_output_size
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,
),
)
def __init__(self, odim, args):
super(Decoder, self).__init__()
self.embed = torch.nn.Sequential(
torch.nn.Embedding(odim, args.adim),
PositionalEncoding(args.adim, args.dropout_rate)
)
self.decoders = repeat(
args.dlayers,
lambda: DecoderLayer(
args.adim,
MultiHeadedAttention(args.aheads, args.adim, args.transformer_attn_dropout_rate),
MultiHeadedAttention(args.aheads, args.adim, args.transformer_attn_dropout_rate),
PositionwiseFeedForward(args.adim, args.dunits, args.dropout_rate),
args.dropout_rate
)
)
self.output_norm = LayerNorm(args.adim)
self.output_layer = torch.nn.Linear(args.adim, odim)
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",
pos_enc_class=PositionalEncoding,
normalize_before=True,
concat_after=False):
super(Encoder, self).__init__()
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)
elif input_layer == "embed":
self.embed = torch.nn.Sequential(
torch.nn.Embedding(idim, attention_dim),
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 ValueError("unknown input_layer: " + input_layer)
self.normalize_before = normalize_before
self.encoders = repeat(
num_blocks, lambda: EncoderLayer(
attention_dim,
MultiHeadedAttention(attention_heads, attention_dim,
attention_dropout_rate),
PositionwiseFeedForward(attention_dim, linear_units,
dropout_rate), dropout_rate,
normalize_before, concat_after))
if self.normalize_before:
self.after_norm = LayerNorm(attention_dim)
def __init__(self, idim, time_len,
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="linear",
pos_enc_class=None,
normalize_before=True,
concat_after=False):
super(Encoder, self).__init__()
self.idim = idim
self.attention_dim = attention_dim
self.attention_heads = attention_heads
self.linear_units = linear_units
self.dropout_rate = dropout_rate
self.positional_dropout_rate = positional_dropout_rate
self.input_layer = input_layer
self._generateInputLayer()
self.normalize_before = normalize_before
self.concat_after= concat_after
self.encoders = repeat(
num_blocks,
lambda: EncoderLayer(
n_head = attention_heads,
d_model = attention_dim,
d_head = attention_dim//attention_heads,
d_inner = attention_dim,
tgt_len = time_len,
ext_len = 0,
mem_len = time_len,
dropout = dropout_rate,
dropatt = attention_dropout_rate,
pre_lnorm= normalize_before
)
)
if self.normalize_before:
self.after_norm = LayerNorm(attention_dim)
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",
pos_enc_class=PositionalEncoding,
normalize_before=True,
concat_after=False,
positionwise_layer_type="linear",
positionwise_conv_kernel_size=1,
padding_idx=-1):
super(Encoder, self).__init__()
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 == "custom":
self.embed = EncoderConv2d(idim, attention_dim)
elif input_layer == "conv2d":
self.embed = Conv2dSubsampling(idim, attention_dim, dropout_rate)
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
if positionwise_layer_type == "linear":
positionwise_layer = PositionwiseFeedForward
positionwise_layer_args = (attention_dim, linear_units, dropout_rate)
elif positionwise_layer_type == "conv1d":
positionwise_layer = MultiLayeredConv1d
positionwise_layer_args = (attention_dim, linear_units, positionwise_conv_kernel_size, dropout_rate)
else:
raise NotImplementedError("Support only linear or conv1d.")
self.encoders = repeat(
num_blocks,
lambda: 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,
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,
cnn_module_kernel=31,
padding_idx=-1,
):
"""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
else:
raise ValueError("unknown pos_enc_layer: " + pos_enc_layer_type)
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),
)
elif input_layer == "vgg2l":
self.embed = VGG2L(idim, attention_dim)
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
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.")
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 == "rel_selfattn":
assert pos_enc_layer_type == "rel_pos"
encoder_selfattn_layer = RelPositionMultiHeadedAttention
encoder_selfattn_layer_args = (
attention_heads,
attention_dim,
attention_dropout_rate,
)
else:
raise ValueError("unknown encoder_attn_layer: " +
selfattention_layer_type)
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)
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
if selfattention_layer_type == "selfattn":
logging.info("decoder self-attention layer type = self-attention")
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),
PositionwiseFeedForward(attention_dim, linear_units,
dropout_rate),
dropout_rate,
normalize_before,
concat_after,
),
)
elif selfattention_layer_type == "lightconv":
logging.info(
"decoder self-attention layer type = lightweight convolution")
self.decoders = repeat(
num_blocks,
lambda lnum: DecoderLayer(
attention_dim,
LightweightConvolution(
conv_wshare,
attention_dim,
self_attention_dropout_rate,
conv_kernel_length,
lnum,
use_kernel_mask=True,
use_bias=conv_usebias,
),
MultiHeadedAttention(attention_heads, attention_dim,
src_attention_dropout_rate),
PositionwiseFeedForward(attention_dim, linear_units,
dropout_rate),
dropout_rate,
normalize_before,
concat_after,
),
)
elif selfattention_layer_type == "lightconv2d":
logging.info("decoder self-attention layer "
"type = lightweight convolution 2-dimentional")
self.decoders = repeat(
num_blocks,
lambda lnum: DecoderLayer(
attention_dim,
LightweightConvolution2D(
conv_wshare,
attention_dim,
self_attention_dropout_rate,
conv_kernel_length,
lnum,
use_kernel_mask=True,
use_bias=conv_usebias,
),
MultiHeadedAttention(attention_heads, attention_dim,
src_attention_dropout_rate),
PositionwiseFeedForward(attention_dim, linear_units,
dropout_rate),
dropout_rate,
normalize_before,
concat_after,
),
)
elif selfattention_layer_type == "dynamicconv":
logging.info(
"decoder self-attention layer type = dynamic convolution")
self.decoders = repeat(
num_blocks,
lambda lnum: DecoderLayer(
attention_dim,
DynamicConvolution(
conv_wshare,
attention_dim,
self_attention_dropout_rate,
conv_kernel_length,
lnum,
use_kernel_mask=True,
use_bias=conv_usebias,
),
MultiHeadedAttention(attention_heads, attention_dim,
src_attention_dropout_rate),
PositionwiseFeedForward(attention_dim, linear_units,
dropout_rate),
dropout_rate,
normalize_before,
concat_after,
),
)
elif selfattention_layer_type == "dynamicconv2d":
logging.info(
"decoder self-attention layer type = dynamic convolution 2-dimentional"
)
self.decoders = repeat(
num_blocks,
lambda lnum: DecoderLayer(
attention_dim,
DynamicConvolution2D(
conv_wshare,
attention_dim,
self_attention_dropout_rate,
conv_kernel_length,
lnum,
use_kernel_mask=True,
use_bias=conv_usebias,
),
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,
input_size: int,
output_size: int = 256,
attention_heads: int = 4,
linear_units: int = 2048,
num_blocks: int = 6,
dropout_rate: float = 0.1,
positional_dropout_rate: float = 0.1,
attention_dropout_rate: float = 0.0,
input_layer: str = "conv2d",
normalize_before: bool = True,
concat_after: bool = False,
positionwise_layer_type: str = "linear",
positionwise_conv_kernel_size: int = 3,
macaron_style: bool = False,
rel_pos_type: str = "legacy",
pos_enc_layer_type: str = "rel_pos",
selfattention_layer_type: str = "rel_selfattn",
activation_type: str = "swish",
use_cnn_module: bool = True,
zero_triu: bool = False,
cnn_module_kernel: int = 31,
padding_idx: int = -1,
interctc_layer_idx: List[int] = [],
interctc_use_conditioning: bool = False,
stochastic_depth_rate: Union[float, List[float]] = 0.0,
):
assert check_argument_types()
super().__init__()
self._output_size = output_size
if rel_pos_type == "legacy":
if pos_enc_layer_type == "rel_pos":
pos_enc_layer_type = "legacy_rel_pos"
if selfattention_layer_type == "rel_selfattn":
selfattention_layer_type = "legacy_rel_selfattn"
elif rel_pos_type == "latest":
assert selfattention_layer_type != "legacy_rel_selfattn"
assert pos_enc_layer_type != "legacy_rel_pos"
else:
raise ValueError("unknown rel_pos_type: " + rel_pos_type)
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":
assert selfattention_layer_type == "legacy_rel_selfattn"
pos_enc_class = LegacyRelPositionalEncoding
logging.warning(
"Using legacy_rel_pos and it will be deprecated in the future."
)
else:
raise ValueError("unknown pos_enc_layer: " + pos_enc_layer_type)
if input_layer == "linear":
self.embed = torch.nn.Sequential(
torch.nn.Linear(input_size, output_size),
torch.nn.LayerNorm(output_size),
torch.nn.Dropout(dropout_rate),
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "conv2d":
self.embed = Conv2dSubsampling(
input_size,
output_size,
dropout_rate,
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "conv2d2":
self.embed = Conv2dSubsampling2(
input_size,
output_size,
dropout_rate,
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "conv2d6":
self.embed = Conv2dSubsampling6(
input_size,
output_size,
dropout_rate,
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "conv2d8":
self.embed = Conv2dSubsampling8(
input_size,
output_size,
dropout_rate,
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "embed":
self.embed = torch.nn.Sequential(
torch.nn.Embedding(input_size, output_size, padding_idx=padding_idx),
pos_enc_class(output_size, positional_dropout_rate),
)
elif isinstance(input_layer, torch.nn.Module):
self.embed = torch.nn.Sequential(
input_layer,
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer is None:
self.embed = torch.nn.Sequential(
pos_enc_class(output_size, positional_dropout_rate)
)
else:
raise ValueError("unknown input_layer: " + input_layer)
self.normalize_before = normalize_before
if positionwise_layer_type == "linear":
positionwise_layer = PositionwiseFeedForward
positionwise_layer_args = (
output_size,
linear_units,
dropout_rate,
activation,
)
elif positionwise_layer_type == "conv1d":
positionwise_layer = MultiLayeredConv1d
positionwise_layer_args = (
output_size,
linear_units,
positionwise_conv_kernel_size,
dropout_rate,
)
elif positionwise_layer_type == "conv1d-linear":
positionwise_layer = Conv1dLinear
positionwise_layer_args = (
output_size,
linear_units,
positionwise_conv_kernel_size,
dropout_rate,
)
else:
raise NotImplementedError("Support only linear or conv1d.")
if selfattention_layer_type == "selfattn":
encoder_selfattn_layer = MultiHeadedAttention
encoder_selfattn_layer_args = (
attention_heads,
output_size,
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,
output_size,
attention_dropout_rate,
)
logging.warning(
"Using legacy_rel_selfattn and it will be deprecated in the future."
)
elif selfattention_layer_type == "rel_selfattn":
assert pos_enc_layer_type == "rel_pos"
encoder_selfattn_layer = RelPositionMultiHeadedAttention
encoder_selfattn_layer_args = (
attention_heads,
output_size,
attention_dropout_rate,
zero_triu,
)
else:
raise ValueError("unknown encoder_attn_layer: " + selfattention_layer_type)
convolution_layer = ConvolutionModule
convolution_layer_args = (output_size, cnn_module_kernel, activation)
if isinstance(stochastic_depth_rate, float):
stochastic_depth_rate = [stochastic_depth_rate] * num_blocks
if len(stochastic_depth_rate) != num_blocks:
raise ValueError(
f"Length of stochastic_depth_rate ({len(stochastic_depth_rate)}) "
f"should be equal to num_blocks ({num_blocks})"
)
self.encoders = repeat(
num_blocks,
lambda lnum: EncoderLayer(
output_size,
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[lnum],
),
)
if self.normalize_before:
self.after_norm = LayerNorm(output_size)
self.interctc_layer_idx = interctc_layer_idx
if len(interctc_layer_idx) > 0:
assert 0 < min(interctc_layer_idx) and max(interctc_layer_idx) < num_blocks
self.interctc_use_conditioning = interctc_use_conditioning
self.conditioning_layer = None
def __init__(
self,
idim,
pred_into_type,
into_type_num,
reduce_character_embedding,
attention_dim=256,
attention_heads=4,
conv_wshare=4,
conv_kernel_length=11,
conv_usebias=False,
linear_units=2048,
num_blocks=3,
dropout_rate=0.2,
positional_dropout_rate=0.1,
attention_dropout_rate=0.0,
pos_enc_class=PositionalEncoding,
normalize_before=True,
concat_after=False,
positionwise_conv_kernel_size=1,
padding_idx=-1,
elayers=None,
eunits=None,
):
"""Construct an Encoder object."""
super(SentenceEncoder, self).__init__()
self.conv_subsampling_factor = 1
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),
)
self.normalize_before = normalize_before
positionwise_layer = PositionwiseFeedForward
positionwise_layer_args = (attention_dim, linear_units, dropout_rate)
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)
# For reduction
self.reduce_character_embedding = reduce_character_embedding
self.query = None # For embedding reduction
if reduce_character_embedding or pred_into_type:
query = torch.nn.Parameter(torch.FloatTensor((attention_dim)),
requires_grad=True)
self.query = torch.nn.init.uniform_(query)
# self.d_k = math.sqrt(eunits)
self.K = torch.nn.Linear(attention_dim, attention_dim)
# self.V = torch.nn.Linear(eunits, eunits)
self.score_dropout = torch.nn.Dropout(p=dropout_rate)
# For prediction
self.pred_prj = None
if pred_into_type:
self.pred_prj = torch.nn.Linear(attention_dim, into_type_num)
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",
pos_enc_class=PositionalEncoding,
normalize_before=True,
concat_after=False,
positionwise_layer_type="linear",
positionwise_conv_kernel_size=1,
padding_idx=-1,
):
"""Construct an Encoder object."""
super(Encoder, self).__init__()
if input_layer == "custom":
self.embed = EncoderConv2d(idim, attention_dim)
elif input_layer == "conv2d":
self.embed = Conv2dSubsampling(idim, attention_dim, dropout_rate)
else:
raise ValueError("unknown input_layer: " + input_layer)
self.normalize_before = normalize_before
if positionwise_layer_type == "linear":
positionwise_layer = PositionwiseFeedForward
positionwise_layer_args = (attention_dim, linear_units,
dropout_rate)
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.")
self.encoders = repeat(
num_blocks,
lambda: 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,
input_size: int,
output_size: int = 256,
attention_heads: int = 4,
linear_units: int = 2048,
num_blocks: int = 6,
dropout_rate: float = 0.1,
positional_dropout_rate: float = 0.1,
attention_dropout_rate: float = 0.0,
input_layer: str = "conv2d",
normalize_before: bool = True,
concat_after: bool = False,
positionwise_layer_type: str = "linear",
positionwise_conv_kernel_size: int = 3,
macaron_style: bool = False,
pos_enc_layer_type: str = "rel_pos",
selfattention_layer_type: str = "rel_selfattn",
activation_type: str = "swish",
use_cnn_module: bool = True,
cnn_module_kernel: int = 31,
padding_idx: int = -1,
):
assert check_argument_types()
super().__init__()
self._output_size = output_size
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
else:
raise ValueError("unknown pos_enc_layer: " + pos_enc_layer_type)
if input_layer == "linear":
self.embed = torch.nn.Sequential(
torch.nn.Linear(input_size, output_size),
torch.nn.LayerNorm(output_size),
torch.nn.Dropout(dropout_rate),
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "conv2d":
self.embed = Conv2dSubsampling(
input_size,
output_size,
dropout_rate,
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "conv2d6":
self.embed = Conv2dSubsampling6(
input_size,
output_size,
dropout_rate,
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "conv2d8":
self.embed = Conv2dSubsampling8(
input_size,
output_size,
dropout_rate,
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "embed":
self.embed = torch.nn.Sequential(
torch.nn.Embedding(input_size,
output_size,
padding_idx=padding_idx),
pos_enc_class(output_size, positional_dropout_rate),
)
elif isinstance(input_layer, torch.nn.Module):
self.embed = torch.nn.Sequential(
input_layer,
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer is None:
self.embed = torch.nn.Sequential(
pos_enc_class(output_size, positional_dropout_rate))
else:
raise ValueError("unknown input_layer: " + input_layer)
self.normalize_before = normalize_before
if positionwise_layer_type == "linear":
positionwise_layer = PositionwiseFeedForward
positionwise_layer_args = (
output_size,
linear_units,
dropout_rate,
activation,
)
elif positionwise_layer_type == "conv1d":
positionwise_layer = MultiLayeredConv1d
positionwise_layer_args = (
output_size,
linear_units,
positionwise_conv_kernel_size,
dropout_rate,
)
elif positionwise_layer_type == "conv1d-linear":
positionwise_layer = Conv1dLinear
positionwise_layer_args = (
output_size,
linear_units,
positionwise_conv_kernel_size,
dropout_rate,
)
else:
raise NotImplementedError("Support only linear or conv1d.")
if selfattention_layer_type == "selfattn":
encoder_selfattn_layer = MultiHeadedAttention
encoder_selfattn_layer_args = (
attention_heads,
output_size,
attention_dropout_rate,
)
elif selfattention_layer_type == "rel_selfattn":
assert pos_enc_layer_type == "rel_pos"
encoder_selfattn_layer = RelPositionMultiHeadedAttention
encoder_selfattn_layer_args = (
attention_heads,
output_size,
attention_dropout_rate,
)
else:
raise ValueError("unknown encoder_attn_layer: " +
selfattention_layer_type)
convolution_layer = ConvolutionModule
convolution_layer_args = (output_size, cnn_module_kernel, activation)
self.encoders = repeat(
num_blocks,
lambda lnum: EncoderLayer(
output_size,
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(output_size)
def __init__(self, idim, odim, args, ignore_id=-1, blank_id=0):
"""Construct an E2E object for transducer model."""
torch.nn.Module.__init__(self)
if "transformer" in args.etype:
if args.enc_block_arch is None:
raise ValueError(
"Transformer-based blocks in transducer mode should be"
"defined individually in the YAML file."
"See egs/vivos/asr1/conf/transducer/* for more info.")
self.subsample = get_subsample(args,
mode="asr",
arch="transformer")
# 2. use transformer to joint feature maps
# transformer without positional encoding
self.clayers = repeat(
2,
lambda lnum: EncoderLayer(
16,
MultiHeadedAttention(4, 16, 0.1),
PositionwiseFeedForward(16, 2048, 0.1),
dropout_rate=0.1,
normalize_before=True,
concat_after=False,
),
)
self.conv = torch.nn.Sequential(
torch.nn.Conv2d(1, 32, kernel_size=(3, 5), stride=(1, 2)),
torch.nn.ReLU(),
torch.nn.Conv2d(32, 32, kernel_size=(3, 7), stride=(2, 2)),
torch.nn.ReLU())
self.encoder = Encoder(
idim,
args.enc_block_arch,
input_layer=args.transformer_enc_input_layer,
repeat_block=args.enc_block_repeat,
self_attn_type=args.transformer_enc_self_attn_type,
positional_encoding_type=args.
transformer_enc_positional_encoding_type,
positionwise_activation_type=args.
transformer_enc_pw_activation_type,
conv_mod_activation_type=args.
transformer_enc_conv_mod_activation_type,
)
encoder_out = self.encoder.enc_out
args.eprojs = self.encoder.enc_out
self.most_dom_list = args.enc_block_arch[:]
else:
self.subsample = get_subsample(args, mode="asr", arch="rnn-t")
self.enc = encoder_for(args, idim, self.subsample)
encoder_out = args.eprojs
if "transformer" in args.dtype:
if args.dec_block_arch is None:
raise ValueError(
"Transformer-based blocks in transducer mode should be"
"defined individually in the YAML file."
"See egs/vivos/asr1/conf/transducer/* for more info.")
self.decoder = DecoderTT(
odim,
encoder_out,
args.joint_dim,
args.dec_block_arch,
input_layer=args.transformer_dec_input_layer,
repeat_block=args.dec_block_repeat,
joint_activation_type=args.joint_activation_type,
positionwise_activation_type=args.
transformer_dec_pw_activation_type,
dropout_rate_embed=args.dropout_rate_embed_decoder,
)
if "transformer" in args.etype:
self.most_dom_list += args.dec_block_arch[:]
else:
self.most_dom_list = args.dec_block_arch[:]
else:
if args.rnnt_mode == "rnnt-att":
self.att = att_for(args)
self.dec = DecoderRNNTAtt(
args.eprojs,
odim,
args.dtype,
args.dlayers,
args.dunits,
blank_id,
self.att,
args.dec_embed_dim,
args.joint_dim,
args.joint_activation_type,
args.dropout_rate_decoder,
args.dropout_rate_embed_decoder,
)
else:
self.dec = DecoderRNNT(
args.eprojs,
odim,
args.dtype,
args.dlayers,
args.dunits,
blank_id,
args.dec_embed_dim,
args.joint_dim,
args.joint_activation_type,
args.dropout_rate_decoder,
args.dropout_rate_embed_decoder,
)
if hasattr(self, "most_dom_list"):
self.most_dom_dim = sorted(
Counter(d["d_hidden"] for d in self.most_dom_list
if "d_hidden" in d).most_common(),
key=lambda x: x[0],
reverse=True,
)[0][0]
self.etype = args.etype
self.dtype = args.dtype
self.rnnt_mode = args.rnnt_mode
self.sos = odim - 1
self.eos = odim - 1
self.blank_id = blank_id
self.ignore_id = ignore_id
self.space = args.sym_space
self.blank = args.sym_blank
self.odim = odim
self.reporter = Reporter()
self.criterion = TransLoss(args.trans_type, self.blank_id)
self.default_parameters(args)
if args.report_cer or args.report_wer:
from espnet.nets.e2e_asr_common import ErrorCalculatorTransducer
if self.dtype == "transformer":
decoder = self.decoder
else:
decoder = self.dec
self.error_calculator = ErrorCalculatorTransducer(
decoder,
args.char_list,
args.sym_space,
args.sym_blank,
args.report_cer,
args.report_wer,
)
else:
self.error_calculator = None
self.loss = None
self.rnnlm = None
def __init__(
self,
input_size: int,
output_size: int = 256,
attention_heads: int = 4,
linear_units: int = 2048,
num_blocks: int = 6,
dropout_rate: float = 0.1,
positional_dropout_rate: float = 0.1,
attention_dropout_rate: float = 0.0,
input_layer: str = "conv2d",
normalize_before: bool = True,
concat_after: bool = False,
positionwise_layer_type: str = "linear",
positionwise_conv_kernel_size: int = 3,
macaron_style: bool = False,
rel_pos_type: str = "legacy",
pos_enc_layer_type: str = "abs_pos",
selfattention_layer_type: str = "lf_selfattn",
activation_type: str = "swish",
use_cnn_module: bool = True,
zero_triu: bool = False,
cnn_module_kernel: int = 31,
padding_idx: int = -1,
interctc_layer_idx: List[int] = [],
interctc_use_conditioning: bool = False,
attention_windows: list = [100, 100, 100, 100, 100, 100],
attention_dilation: list = [1, 1, 1, 1, 1, 1],
attention_mode: str = "sliding_chunks",
):
assert check_argument_types()
super().__init__(input_size)
self._output_size = output_size
activation = get_activation(activation_type)
if pos_enc_layer_type == "abs_pos":
pos_enc_class = PositionalEncoding
else:
raise ValueError("incorrect or unknown pos_enc_layer: " +
pos_enc_layer_type + "Use abs_pos")
if len(attention_dilation) != num_blocks:
raise ValueError(
"incorrect attention_dilation parameter of length" +
str(len(attention_dilation)) + " does not match num_blocks" +
str(num_blocks))
if len(attention_windows) != num_blocks:
raise ValueError(
"incorrect attention_windows parameter of length" +
str(len(attention_windows)) + " does not match num_blocks" +
str(num_blocks))
if attention_mode != "tvm" and max(attention_dilation) != 1:
raise ValueError("incorrect attention mode for dilation: " +
attention_mode +
"Use attention_mode=tvm with Cuda Kernel")
if input_layer == "linear":
self.embed = torch.nn.Sequential(
torch.nn.Linear(input_size, output_size),
torch.nn.LayerNorm(output_size),
torch.nn.Dropout(dropout_rate),
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "conv2d":
self.embed = Conv2dSubsampling(
input_size,
output_size,
dropout_rate,
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "conv2d2":
self.embed = Conv2dSubsampling2(
input_size,
output_size,
dropout_rate,
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "conv2d6":
self.embed = Conv2dSubsampling6(
input_size,
output_size,
dropout_rate,
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "conv2d8":
self.embed = Conv2dSubsampling8(
input_size,
output_size,
dropout_rate,
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "embed":
self.embed = torch.nn.Sequential(
torch.nn.Embedding(input_size,
output_size,
padding_idx=padding_idx),
pos_enc_class(output_size, positional_dropout_rate),
)
elif isinstance(input_layer, torch.nn.Module):
self.embed = torch.nn.Sequential(
input_layer,
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer is None:
self.embed = torch.nn.Sequential(
pos_enc_class(output_size, positional_dropout_rate))
else:
raise ValueError("unknown input_layer: " + input_layer)
self.normalize_before = normalize_before
if positionwise_layer_type == "linear":
positionwise_layer = PositionwiseFeedForward
positionwise_layer_args = (
output_size,
linear_units,
dropout_rate,
activation,
)
elif positionwise_layer_type == "conv1d":
positionwise_layer = MultiLayeredConv1d
positionwise_layer_args = (
output_size,
linear_units,
positionwise_conv_kernel_size,
dropout_rate,
)
elif positionwise_layer_type == "conv1d-linear":
positionwise_layer = Conv1dLinear
positionwise_layer_args = (
output_size,
linear_units,
positionwise_conv_kernel_size,
dropout_rate,
)
else:
raise NotImplementedError("Support only linear or conv1d.")
self.selfattention_layer_type = selfattention_layer_type
if selfattention_layer_type == "lf_selfattn":
assert pos_enc_layer_type == "abs_pos"
from longformer.longformer import LongformerConfig
from espnet.nets.pytorch_backend.transformer.longformer_attention import (
LongformerAttention, )
encoder_selfattn_layer = LongformerAttention
config = LongformerConfig(
attention_window=attention_windows,
attention_dilation=attention_dilation,
autoregressive=False,
num_attention_heads=attention_heads,
hidden_size=output_size,
attention_probs_dropout_prob=dropout_rate,
attention_mode=attention_mode,
)
encoder_selfattn_layer_args = (config, )
else:
raise ValueError("incompatible or unknown encoder_attn_layer: " +
selfattention_layer_type + " Use lf_selfattn")
convolution_layer = ConvolutionModule
convolution_layer_args = (output_size, cnn_module_kernel, activation)
self.encoders = repeat(
num_blocks,
lambda layer_id: EncoderLayer(
output_size,
encoder_selfattn_layer(*(encoder_selfattn_layer_args +
(layer_id, ))),
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(output_size)
self.interctc_layer_idx = interctc_layer_idx
if len(interctc_layer_idx) > 0:
assert 0 < min(interctc_layer_idx) and max(
interctc_layer_idx) < num_blocks
self.interctc_use_conditioning = interctc_use_conditioning
self.conditioning_layer = None
def __init__(self,
odim,
attention_dim=256,
attention_heads=4,
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,
cross_operator=None,
cross_shared=False,
cross_weight_learnable=False,
cross_weight=0.0):
"""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
cross_self_attn = None
cross_src_attn = None
if cross_operator:
if 'src_' in cross_operator:
# cross_src_attn = MultiHeadedAttention(attention_heads, attention_dim, self_attention_dropout_rate)
cross_src_attn = True
if 'self_' in cross_operator:
if cross_shared and cross_src_attn is not None:
# cross_self_attn = cross_src_attn
cross_self_attn = True # TODO: backward compatibility for shared self and source
else:
# cross_self_attn = MultiHeadedAttention(attention_heads, attention_dim, self_attention_dropout_rate)
cross_self_attn = True
if 'concat' in cross_operator:
cross_operator = 'concat'
elif 'sum' in cross_operator:
cross_operator = 'sum'
else:
raise NotImplementedError
self.decoders = repeat(
num_blocks, lambda: 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,
cross_self_attn=MultiHeadedAttention(
attention_heads, attention_dim, self_attention_dropout_rate
) if cross_self_attn else None,
cross_src_attn=MultiHeadedAttention(
attention_heads, attention_dim, self_attention_dropout_rate
) if cross_src_attn else None,
cross_operator=cross_operator,
cross_shared=cross_shared,
cross_weight_learnable=cross_weight_learnable,
cross_weight=cross_weight))
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_sd=4,
num_blocks_rec=8,
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,
padding_idx=-1,
num_spkrs=2,
):
"""Construct an Encoder object."""
super(EncoderMix, self).__init__(
idim=idim,
selfattention_layer_type="selfattn",
attention_dim=attention_dim,
attention_heads=attention_heads,
linear_units=linear_units,
num_blocks=num_blocks_rec,
dropout_rate=dropout_rate,
positional_dropout_rate=positional_dropout_rate,
attention_dropout_rate=attention_dropout_rate,
input_layer=input_layer,
pos_enc_class=pos_enc_class,
normalize_before=normalize_before,
concat_after=concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
padding_idx=padding_idx,
)
positionwise_layer, positionwise_layer_args = self.get_positionwise_layer(
positionwise_layer_type,
attention_dim,
linear_units,
dropout_rate,
positionwise_conv_kernel_size,
)
self.num_spkrs = num_spkrs
self.encoders_sd = torch.nn.ModuleList([
repeat(
num_blocks_sd,
lambda lnum: EncoderLayer(
attention_dim,
MultiHeadedAttention(attention_heads, attention_dim,
attention_dropout_rate),
positionwise_layer(*positionwise_layer_args),
dropout_rate,
normalize_before,
concat_after,
),
) for i in range(num_spkrs)
])
def __init__(
self,
input_size: int,
output_size: int = 256,
attention_heads: int = 4,
linear_units: int = 2048,
num_blocks: int = 6,
dropout_rate: float = 0.1,
positional_dropout_rate: float = 0.1,
attention_dropout_rate: float = 0.0,
input_layer: Optional[str] = "conv2d",
pos_enc_class=PositionalEncoding,
normalize_before: bool = True,
concat_after: bool = False,
positionwise_layer_type: str = "linear",
positionwise_conv_kernel_size: int = 1,
padding_idx: int = -1,
):
assert check_argument_types()
super().__init__()
self._output_size = output_size
if input_layer == "linear":
self.embed = torch.nn.Sequential(
torch.nn.Linear(input_size, output_size),
torch.nn.LayerNorm(output_size),
torch.nn.Dropout(dropout_rate),
torch.nn.ReLU(),
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "conv2d":
self.embed = Conv2dSubsampling(input_size, output_size, dropout_rate)
elif input_layer == "conv2d6":
self.embed = Conv2dSubsampling6(input_size, output_size, dropout_rate)
elif input_layer == "conv2d8":
self.embed = Conv2dSubsampling8(input_size, output_size, dropout_rate)
elif input_layer == "embed":
self.embed = torch.nn.Sequential(
torch.nn.Embedding(input_size, output_size, padding_idx=padding_idx),
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer is None:
self.embed = torch.nn.Sequential(
pos_enc_class(output_size, positional_dropout_rate)
)
else:
raise ValueError("unknown input_layer: " + input_layer)
self.normalize_before = normalize_before
if positionwise_layer_type == "linear":
positionwise_layer = PositionwiseFeedForward
positionwise_layer_args = (
output_size,
linear_units,
dropout_rate,
)
elif positionwise_layer_type == "conv1d":
positionwise_layer = MultiLayeredConv1d
positionwise_layer_args = (
output_size,
linear_units,
positionwise_conv_kernel_size,
dropout_rate,
)
elif positionwise_layer_type == "conv1d-linear":
positionwise_layer = Conv1dLinear
positionwise_layer_args = (
output_size,
linear_units,
positionwise_conv_kernel_size,
dropout_rate,
)
else:
raise NotImplementedError("Support only linear or conv1d.")
self.encoders = repeat(
num_blocks,
lambda lnum: EncoderLayer(
output_size,
MultiHeadedAttention(
attention_heads, output_size, attention_dropout_rate
),
positionwise_layer(*positionwise_layer_args),
dropout_rate,
normalize_before,
concat_after,
),
)
if self.normalize_before:
self.after_norm = LayerNorm(output_size)
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,
stochastic_depth_rate=0.0,
intermediate_layers=None,
):
"""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-dimensional")
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-dimensional"
)
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,
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,
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,
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,
padding_idx=-1,
):
"""Construct an Encoder object."""
super(Encoder, self).__init__()
self._register_load_state_dict_pre_hook(_pre_hook)
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)
elif input_layer == "conv2d-scaled-pos-enc":
self.embed = Conv2dSubsampling(
idim,
attention_dim,
dropout_rate,
pos_enc_class(attention_dim, positional_dropout_rate),
)
elif input_layer == "conv2d6":
self.embed = Conv2dSubsampling6(idim, attention_dim, dropout_rate)
elif input_layer == "conv2d8":
self.embed = Conv2dSubsampling8(idim, attention_dim, dropout_rate)
elif input_layer == "vgg2l":
self.embed = VGG2L(idim, attention_dim)
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 == "selfattn":
logging.info("encoder self-attention layer type = self-attention")
self.encoders = repeat(
num_blocks,
lambda lnum: EncoderLayer(
attention_dim,
MultiHeadedAttention_wordscale(
attention_heads, attention_dim, attention_dropout_rate
),
positionwise_layer(*positionwise_layer_args),
dropout_rate,
normalize_before,
concat_after,
),
)
elif selfattention_layer_type == "lightconv":
logging.info(
"encoder self-attention layer type = lightweight convolution")
self.encoders = repeat(
num_blocks,
lambda lnum: EncoderLayer(
attention_dim,
LightweightConvolution(
conv_wshare,
attention_dim,
attention_dropout_rate,
conv_kernel_length,
lnum,
use_bias=conv_usebias,
),
positionwise_layer(*positionwise_layer_args),
dropout_rate,
normalize_before,
concat_after,
),
)
elif selfattention_layer_type == "lightconv2d":
logging.info("encoder self-attention layer "
"type = lightweight convolution 2-dimentional")
self.encoders = repeat(
num_blocks,
lambda lnum: EncoderLayer(
attention_dim,
LightweightConvolution2D(
conv_wshare,
attention_dim,
attention_dropout_rate,
conv_kernel_length,
lnum,
use_bias=conv_usebias,
),
positionwise_layer(*positionwise_layer_args),
dropout_rate,
normalize_before,
concat_after,
),
)
elif selfattention_layer_type == "dynamicconv":
logging.info(
"encoder self-attention layer type = dynamic convolution")
self.encoders = repeat(
num_blocks,
lambda lnum: EncoderLayer(
attention_dim,
DynamicConvolution(
conv_wshare,
attention_dim,
attention_dropout_rate,
conv_kernel_length,
lnum,
use_bias=conv_usebias,
),
positionwise_layer(*positionwise_layer_args),
dropout_rate,
normalize_before,
concat_after,
),
)
elif selfattention_layer_type == "dynamicconv2d":
logging.info(
"encoder self-attention layer type = dynamic convolution 2-dimentional"
)
self.encoders = repeat(
num_blocks,
lambda lnum: EncoderLayer(
attention_dim,
DynamicConvolution2D(
conv_wshare,
attention_dim,
attention_dropout_rate,
conv_kernel_length,
lnum,
use_bias=conv_usebias,
),
positionwise_layer(*positionwise_layer_args),
dropout_rate,
normalize_before,
concat_after,
),
)
if self.normalize_before:
self.after_norm = LayerNorm(attention_dim)
def __init__(
self,
languages,
idim,
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,
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,
padding_idx=-1,
sim_adapter=False,
shared_adapter=None,
use_adapters=True,
fusion_languages=None,
):
super().__init__(idim, selfattention_layer_type, attention_dim,
attention_heads, conv_wshare, conv_kernel_length,
conv_usebias, linear_units, num_blocks, dropout_rate,
positional_dropout_rate, attention_dropout_rate,
input_layer, pos_enc_class, normalize_before,
concat_after, positionwise_layer_type,
positionwise_conv_kernel_size, padding_idx)
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 == "selfattn":
logging.info("encoder self-attention layer type = self-attention")
self.encoders = repeat(
num_blocks,
lambda lnum: AdaptiveEncoderLayer(
languages,
attention_dim,
MultiHeadedAttention(attention_heads, attention_dim,
attention_dropout_rate),
positionwise_layer(*positionwise_layer_args),
dropout_rate,
normalize_before,
concat_after,
torch.nn.ModuleDict({
"_".join(sorted(fusion_languages)):
SimAdapter(attention_dim, attention_dropout_rate,
fusion_languages)
}) if sim_adapter else None,
shared_adapter,
use_adapters,
),
)
else:
raise NotImplementedError(
"Only support self-attention encoder layer")
def __init__(
self,
input_size: int,
output_size: int = 256,
attention_heads: int = 4,
linear_units: int = 2048,
num_blocks: int = 6,
dropout_rate: float = 0.1,
positional_dropout_rate: float = 0.1,
attention_dropout_rate: float = 0.0,
input_layer: Optional[str] = "conv2d",
pos_enc_class=StreamPositionalEncoding,
normalize_before: bool = True,
concat_after: bool = False,
positionwise_layer_type: str = "linear",
positionwise_conv_kernel_size: int = 1,
padding_idx: int = -1,
block_size: int = 40,
hop_size: int = 16,
look_ahead: int = 16,
init_average: bool = True,
ctx_pos_enc: bool = True,
):
assert check_argument_types()
super().__init__()
self._output_size = output_size
self.pos_enc = pos_enc_class(output_size, positional_dropout_rate)
if input_layer == "linear":
self.embed = torch.nn.Sequential(
torch.nn.Linear(input_size, output_size),
torch.nn.LayerNorm(output_size),
torch.nn.Dropout(dropout_rate),
torch.nn.ReLU(),
)
self.subsample = 1
elif input_layer == "conv2d":
self.embed = Conv2dSubsamplingWOPosEnc(input_size,
output_size,
dropout_rate,
kernels=[3, 3],
strides=[2, 2])
self.subsample = 4
elif input_layer == "conv2d6":
self.embed = Conv2dSubsamplingWOPosEnc(input_size,
output_size,
dropout_rate,
kernels=[3, 5],
strides=[2, 3])
self.subsample = 6
elif input_layer == "conv2d8":
self.embed = Conv2dSubsamplingWOPosEnc(
input_size,
output_size,
dropout_rate,
kernels=[3, 3, 3],
strides=[2, 2, 2],
)
self.subsample = 8
elif input_layer == "embed":
self.embed = torch.nn.Sequential(
torch.nn.Embedding(input_size,
output_size,
padding_idx=padding_idx), )
self.subsample = 1
elif input_layer is None:
self.embed = None
self.subsample = 1
else:
raise ValueError("unknown input_layer: " + input_layer)
self.normalize_before = normalize_before
if positionwise_layer_type == "linear":
positionwise_layer = PositionwiseFeedForward
positionwise_layer_args = (
output_size,
linear_units,
dropout_rate,
)
elif positionwise_layer_type == "conv1d":
positionwise_layer = MultiLayeredConv1d
positionwise_layer_args = (
output_size,
linear_units,
positionwise_conv_kernel_size,
dropout_rate,
)
elif positionwise_layer_type == "conv1d-linear":
positionwise_layer = Conv1dLinear
positionwise_layer_args = (
output_size,
linear_units,
positionwise_conv_kernel_size,
dropout_rate,
)
else:
raise NotImplementedError("Support only linear or conv1d.")
self.encoders = repeat(
num_blocks,
lambda lnum: ContextualBlockEncoderLayer(
output_size,
MultiHeadedAttention(attention_heads, output_size,
attention_dropout_rate),
positionwise_layer(*positionwise_layer_args),
dropout_rate,
num_blocks,
normalize_before,
concat_after,
),
)
if self.normalize_before:
self.after_norm = LayerNorm(output_size)
# for block processing
self.block_size = block_size
self.hop_size = hop_size
self.look_ahead = look_ahead
self.init_average = init_average
self.ctx_pos_enc = ctx_pos_enc
def __init__(
self,
languages,
odim_dict,
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,
sim_adapter=False,
shared_adapter=False,
use_adapters=True,
fusion_languages=None,
):
super().__init__(1, selfattention_layer_type, attention_dim,
attention_heads, conv_wshare, conv_kernel_length,
conv_usebias, linear_units, num_blocks, dropout_rate,
positional_dropout_rate, self_attention_dropout_rate,
src_attention_dropout_rate, input_layer,
use_output_layer, pos_enc_class, normalize_before,
concat_after)
if input_layer == "embed":
self.embed = torch.nn.ModuleDict()
for lang in odim_dict.keys():
self.embed[lang] = torch.nn.Sequential(
torch.nn.Embedding(odim_dict[lang], attention_dim),
pos_enc_class(attention_dim, positional_dropout_rate),
)
else:
raise NotImplementedError("only support embed embedding layer")
assert self_attention_dropout_rate == src_attention_dropout_rate
if selfattention_layer_type == "selfattn":
logging.info("decoder self-attention layer type = self-attention")
self.decoders = repeat(
num_blocks,
lambda lnum: AdaptiveDecoderLayer(
languages,
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,
torch.nn.ModuleDict({
"_".join(sorted(fusion_languages)):
SimAdapter(attention_dim, self_attention_dropout_rate,
fusion_languages)
}) if sim_adapter else None,
shared_adapter,
use_adapters,
),
)
else:
raise NotImplementedError(
"Only support self-attention decoder layer")
if use_output_layer:
self.output_layer = torch.nn.ModuleDict()
for lang in odim_dict.keys():
self.output_layer[lang] = torch.nn.Linear(
attention_dim, odim_dict[lang])
else:
self.output_layer = None
def __init__(self,
idim,
center_len=8,
left_len=0,
hop_len=0,
right_len=0,
abs_pos=1,
rel_pos=0,
use_mem=1,
att_type="mta",
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):
super(Encoder, self).__init__()
self.idim = idim
self.center_len = center_len
self.use_mem = use_mem != 0
self.left_len = left_len
if self.use_mem:
self.mem_len = left_len
else:
self.mem_len = 0
self.hop_len = hop_len
self.right_len = right_len
self.abs_pos = abs_pos != 0
self.rel_pos = rel_pos != 0
self.attention_dim = attention_dim
self.attention_heads = attention_heads
self.linear_units = linear_units
self.dropout_rate = dropout_rate
self.input_layer = input_layer
self.normalize_before = normalize_before
self.concat_after = concat_after
self.positional_dropout_rate = positional_dropout_rate
self.pos_enc_class = pos_enc_class
self._generateInputLayer()
self.encoders = repeat(
num_blocks, lambda:
EncoderLayer(n_head=attention_heads,
d_model=attention_dim,
d_head=attention_dim // attention_heads,
att_type=att_type,
ext_len=hop_len // 4,
mem_len=self.mem_len // 4,
tgt_len=center_len,
future_len=right_len,
rel_pos=rel_pos,
dropout=dropout_rate,
dropatt=attention_dropout_rate,
pre_lnorm=normalize_before,
pos_ff=PositionwiseFeedForward(
attention_dim, linear_units, dropout_rate)))
if self.normalize_before:
self.after_norm = LayerNorm(attention_dim)
def __init__(
self,
input_size: int,
output_size: int = 256,
attention_heads: int = 4,
linear_units: int = 2048,
num_blocks: int = 6,
dropout_rate: float = 0.1,
positional_dropout_rate: float = 0.1,
attention_dropout_rate: float = 0.0,
input_layer: Optional[str] = None,
normalize_before: bool = True,
concat_after: bool = False,
positionwise_layer_type: str = "linear",
positionwise_conv_kernel_size: int = 1,
pos_enc_layer_type: str = "rel_pos",
selfattention_layer_type: str = "rel_selfattn",
activation_type='relu',
padding_idx: int = -1,
):
assert check_argument_types()
super().__init__()
self._output_size = output_size
# todo: my change, from conformer/encoder_layer.py
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
else:
raise ValueError("unknown pos_enc_layer: " + pos_enc_layer_type)
# input layer
if input_layer == "linear":
self.embed = torch.nn.Sequential(
torch.nn.Linear(input_size, output_size),
torch.nn.LayerNorm(output_size),
torch.nn.Dropout(dropout_rate),
torch.nn.ReLU(),
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "conv2d":
self.embed = Conv2dSubsampling(input_size, output_size, dropout_rate)
elif input_layer == "conv2d6":
self.embed = Conv2dSubsampling6(input_size, output_size, dropout_rate)
elif input_layer == "conv2d8":
self.embed = Conv2dSubsampling8(input_size, output_size, dropout_rate)
elif input_layer == "embed":
self.embed = torch.nn.Sequential(
torch.nn.Embedding(input_size, output_size, padding_idx=padding_idx),
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer is None:
self.embed = torch.nn.Sequential(
pos_enc_class(output_size, positional_dropout_rate)
)
else:
raise ValueError("unknown input_layer: " + input_layer)
self.normalize_before = normalize_before
# position-wise layer
activation = get_activation(activation_type)
if positionwise_layer_type == "linear":
positionwise_layer = PositionwiseFeedForward
positionwise_layer_args = (
output_size,
linear_units,
dropout_rate,
activation,
)
elif positionwise_layer_type == "conv1d":
positionwise_layer = MultiLayeredConv1d
positionwise_layer_args = (
output_size,
linear_units,
positionwise_conv_kernel_size,
dropout_rate,
)
elif positionwise_layer_type == "conv1d-linear":
positionwise_layer = Conv1dLinear
positionwise_layer_args = (
output_size,
linear_units,
positionwise_conv_kernel_size,
dropout_rate,
)
else:
raise NotImplementedError("Support only linear or conv1d.")
# encoders type and args
if selfattention_layer_type == "selfattn":
encoder_selfattn_layer = MultiHeadedAttention
encoder_selfattn_layer_args = (
attention_heads,
output_size,
attention_dropout_rate,
)
elif selfattention_layer_type == "rel_selfattn":
assert pos_enc_layer_type == "rel_pos"
encoder_selfattn_layer = RelPositionMultiHeadedAttention
encoder_selfattn_layer_args = (
attention_heads,
output_size,
attention_dropout_rate,
)
else:
raise ValueError("unknown encoder_attn_layer: " + selfattention_layer_type)
# encoders
self.encoders = repeat(
num_blocks,
lambda lnum: EncoderLayer(
output_size,
encoder_selfattn_layer(*encoder_selfattn_layer_args),
positionwise_layer(*positionwise_layer_args),
dropout_rate,
normalize_before,
concat_after,
),
)
if self.normalize_before:
self.after_norm = LayerNorm(output_size)
