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 build_conformer_block(
block: Dict[str, Any],
self_attn_class: str,
pw_layer_type: str,
pw_activation_type: str,
conv_mod_activation_type: str,
) -> ConformerEncoderLayer:
"""Build function for conformer block.
Args:
block: Conformer block parameters.
self_attn_type: Self-attention module type.
pw_layer_type: Positionwise layer type.
pw_activation_type: Positionwise activation type.
conv_mod_activation_type: Convolutional module activation type.
Returns:
: Function to create conformer (encoder) block.
"""
d_hidden = block["d_hidden"]
d_ff = block["d_ff"]
dropout_rate = block.get("dropout-rate", 0.0)
pos_dropout_rate = block.get("pos-dropout-rate", 0.0)
att_dropout_rate = block.get("att-dropout-rate", 0.0)
if pw_layer_type == "linear":
pw_layer = PositionwiseFeedForward(
d_hidden,
d_ff,
pos_dropout_rate,
get_activation(pw_activation_type),
)
else:
raise NotImplementedError("Conformer block only supports linear yet.")
macaron_net = (PositionwiseFeedForward(
d_hidden,
d_ff,
pos_dropout_rate,
get_activation(pw_activation_type),
) if block["macaron_style"] else None)
conv_mod = (ConvolutionModule(
d_hidden,
block["conv_mod_kernel"],
get_activation(conv_mod_activation_type),
) if block["use_conv_mod"] else None)
return lambda: ConformerEncoderLayer(
d_hidden,
self_attn_class(block["heads"], d_hidden, att_dropout_rate),
pw_layer,
macaron_net,
conv_mod,
dropout_rate,
)
def _generateEncoderLayer(self):
subsize = 1
if self.input_layer == "conv2d":
subsize = 4
subsample = None
else:
if self.subtype == "max":
subsample = PoolSubsampling()
elif self.subtype == "average":
subsample = PoolSubsampling(type="average")
else:
subsample = NormalSubsamplingPos()
for i in range(self.num_blocks):
if self.input_layer != "conv2d":
for sp in self.subpos:
if sp == i:
self.encoders.append(subsample)
subsize *= 2
self.encoders.append(
EncoderLayer(n_head=self.attention_heads,
d_model=self.attention_dim,
d_head=self.attention_dim // self.attention_heads,
att_type=self.att_type,
ext_len=self.hop_len // subsize,
mem_len=self.mem_len // subsize,
tgt_len=self.center_len,
future_len=self.right_len,
rel_pos=self.rel_pos,
dropout=self.dropout_rate,
dropatt=self.attention_dropout_rate,
pre_lnorm=self.normalize_before,
pos_ff=PositionwiseFeedForward(
self.attention_dim, self.linear_units,
self.dropout_rate)))
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,
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,
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,
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,
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 build_transformer_block(
net_part: str,
block: Dict[str, Any],
pw_layer_type: str,
pw_activation_type: str,
) -> Union[EncoderLayer, TransformerDecoderLayer]:
"""Build function for transformer block.
Args:
net_part: Network part, either 'encoder' or 'decoder'.
block: Transformer block parameters.
pw_layer_type: Positionwise layer type.
pw_activation_type: Positionwise activation type.
Returns:
: Function to create transformer (encoder or decoder) block.
"""
d_hidden = block["d_hidden"]
dropout_rate = block.get("dropout-rate", 0.0)
pos_dropout_rate = block.get("pos-dropout-rate", 0.0)
att_dropout_rate = block.get("att-dropout-rate", 0.0)
if pw_layer_type != "linear":
raise NotImplementedError(
"Transformer block only supports linear pointwise layer.")
if net_part == "encoder":
transformer_layer_class = EncoderLayer
elif net_part == "decoder":
transformer_layer_class = TransformerDecoderLayer
return lambda: transformer_layer_class(
d_hidden,
MultiHeadedAttention(block["heads"], d_hidden, att_dropout_rate),
PositionwiseFeedForward(
d_hidden,
block["d_ff"],
pos_dropout_rate,
get_activation(pw_activation_type),
),
dropout_rate,
)
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,
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,
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 build_conformer_block(
configuration: List[Dict[str, Any]],
main_params: Dict[str, Any],
) -> Conformer:
"""Build Conformer block.
Args:
configuration: Conformer block configuration.
main_params: Encoder main parameters.
Returns:
: Conformer block function.
"""
hidden_size = configuration["hidden_size"]
linear_size = configuration["linear_size"]
pos_wise_args = (
hidden_size,
linear_size,
configuration.get("pos_wise_dropout_rate", 0.0),
main_params["pos_wise_act"],
)
conv_mod_norm_class, conv_mod_norm_args = get_normalization(
main_params["conv_mod_norm_type"],
eps=configuration.get("conv_mod_norm_eps"),
momentum=configuration.get("conv_mod_norm_momentum"),
partial=configuration.get("conv_mod_norm_partial"),
)
conv_mod_args = (
hidden_size,
configuration["conv_mod_kernel_size"],
main_params["conv_mod_act"],
conv_mod_norm_class,
conv_mod_norm_args,
main_params["dynamic_chunk_training"],
)
mult_att_args = (
configuration.get("heads", 4),
hidden_size,
configuration.get("att_dropout_rate", 0.0),
main_params["simplified_att_score"],
)
norm_class, norm_args = get_normalization(
main_params["norm_type"],
eps=configuration.get("norm_eps"),
partial=configuration.get("norm_partial"),
)
return lambda: Conformer(
hidden_size,
RelPositionMultiHeadedAttention(*mult_att_args),
PositionwiseFeedForward(*pos_wise_args),
PositionwiseFeedForward(*pos_wise_args),
ConformerConvolution(*conv_mod_args),
norm_class=norm_class,
norm_args=norm_args,
dropout_rate=configuration.get("dropout_rate", 0.0),
)
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, odim, args, ignore_id=-1):
"""Construct an E2E object.
:param int idim: dimension of inputs
:param int odim: dimension of outputs
:param Namespace args: argument Namespace containing options
"""
torch.nn.Module.__init__(self)
if args.transformer_attn_dropout_rate is None:
args.transformer_attn_dropout_rate = args.dropout_rate
self.cn_encoder = Encoder(
idim=idim,
attention_dim=args.adim,
attention_heads=args.aheads,
linear_units=args.eunits,
num_blocks=args.elayers,
input_layer=args.transformer_input_layer,
dropout_rate=args.dropout_rate,
positional_dropout_rate=args.dropout_rate,
attention_dropout_rate=args.transformer_attn_dropout_rate)
self.en_encoder = Encoder(
idim=idim,
attention_dim=args.adim,
attention_heads=args.aheads,
linear_units=args.eunits,
num_blocks=args.elayers,
input_layer=args.transformer_input_layer,
dropout_rate=args.dropout_rate,
positional_dropout_rate=args.dropout_rate,
attention_dropout_rate=args.transformer_attn_dropout_rate)
# gated add module
self.vectorize_lambda = args.vectorize_lambda
lambda_dim = args.adim if self.vectorize_lambda else 1
self.aggregation_module = torch.nn.Sequential(
torch.nn.Linear(2 * args.adim, lambda_dim), torch.nn.Sigmoid())
self.additional_encoder_layer = EncoderLayer(
args.adim,
MultiHeadedAttention(args.aheads, args.adim,
args.transformer_attn_dropout_rate),
PositionwiseFeedForward(args.adim, args.eunits, args.dropout_rate),
args.dropout_rate,
normalize_before=True,
concat_after=False)
self.additional_after_norm = LayerNorm(args.adim)
self.decoder = Decoder(
odim=odim,
attention_dim=args.adim,
attention_heads=args.aheads,
linear_units=args.dunits,
num_blocks=args.dlayers,
dropout_rate=args.dropout_rate,
positional_dropout_rate=args.dropout_rate,
self_attention_dropout_rate=args.transformer_attn_dropout_rate,
src_attention_dropout_rate=args.transformer_attn_dropout_rate)
self.sos = odim - 1
self.eos = odim - 1
self.odim = odim
self.ignore_id = ignore_id
self.subsample = [1]
self.reporter = Reporter()
# self.lsm_weight = a
self.criterion = LabelSmoothingLoss(
self.odim, self.ignore_id, args.lsm_weight,
args.transformer_length_normalized_loss)
# self.verbose = args.verbose
self.adim = args.adim
self.mtlalpha = args.mtlalpha
if args.mtlalpha > 0.0:
self.ctc = CTC(odim,
args.adim,
args.dropout_rate,
ctc_type=args.ctc_type,
reduce=True)
else:
self.ctc = None
if args.report_cer or args.report_wer:
from espnet.nets.e2e_asr_common import ErrorCalculator
self.error_calculator = ErrorCalculator(args.char_list,
args.sym_space,
args.sym_blank,
args.report_cer,
args.report_wer)
else:
self.error_calculator = None
self.rnnlm = None
# yzl23 config
self.remove_blank_in_ctc_mode = True
self.reset_parameters(args) # reset params at the last
logging.warning(
"Model total size: {}M, requires_grad size: {}M".format(
self.count_parameters(),
self.count_parameters(requires_grad=True)))
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,
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_weight_learnable=False,
cross_weight=0.0,
cross_self=False,
cross_src=False,
cross_to_asr=True,
cross_to_st=True):
"""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))
self.embed_asr = 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))
self.embed_asr = 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))
self.embed_asr = 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.dual_decoders = repeat(
num_blocks, lambda: DualDecoderLayer(
attention_dim,
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),
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),
cross_self_attn=MultiHeadedAttention(
attention_heads, attention_dim, self_attention_dropout_rate
) if (cross_self and cross_to_st) else None,
cross_self_attn_asr=MultiHeadedAttention(
attention_heads, attention_dim, self_attention_dropout_rate
) if (cross_self and cross_to_asr) else None,
cross_src_attn=MultiHeadedAttention(
attention_heads, attention_dim, self_attention_dropout_rate
) if (cross_src and cross_to_st) else None,
cross_src_attn_asr=MultiHeadedAttention(
attention_heads, attention_dim, self_attention_dropout_rate
) if (cross_src and cross_to_asr) else None,
dropout_rate=dropout_rate,
normalize_before=normalize_before,
concat_after=concat_after,
cross_operator=cross_operator,
cross_weight_learnable=cross_weight_learnable,
cross_weight=cross_weight,
cross_to_asr=cross_to_asr,
cross_to_st=cross_to_st))
if self.normalize_before:
self.after_norm = LayerNorm(attention_dim)
self.after_norm_asr = LayerNorm(attention_dim)
if use_output_layer:
self.output_layer = torch.nn.Linear(attention_dim, odim)
self.output_layer_asr = torch.nn.Linear(attention_dim, odim)
else:
self.output_layer = None
self.output_layer_asr = None