def __init__(
self,
idim: int,
odim: int,
args: Namespace,
ignore_id: int = -1,
blank_id: int = 0,
training: bool = True,
):
"""Construct an E2E object for transducer model."""
torch.nn.Module.__init__(self)
args = fill_missing_args(args, self.add_arguments)
self.is_transducer = True
self.use_auxiliary_enc_outputs = (True if (
training and args.use_aux_transducer_loss) else False)
self.subsample = get_subsample(
args,
mode="asr",
arch="transformer" if args.etype == "custom" else "rnn-t")
if self.use_auxiliary_enc_outputs:
n_layers = (((len(args.enc_block_arch) * args.enc_block_repeat) -
1) if args.enc_block_arch is not None else
(args.elayers - 1))
aux_enc_output_layers = valid_aux_encoder_output_layers(
args.aux_transducer_loss_enc_output_layers,
n_layers,
args.use_symm_kl_div_loss,
self.subsample,
)
else:
aux_enc_output_layers = []
if args.etype == "custom":
if args.enc_block_arch is None:
raise ValueError(
"When specifying custom encoder type, --enc-block-arch"
"should also be specified in training config. See"
"egs/vivos/asr1/conf/transducer/train_*.yaml for more info."
)
self.encoder = CustomEncoder(
idim,
args.enc_block_arch,
input_layer=args.custom_enc_input_layer,
repeat_block=args.enc_block_repeat,
self_attn_type=args.custom_enc_self_attn_type,
positional_encoding_type=args.
custom_enc_positional_encoding_type,
positionwise_activation_type=args.
custom_enc_pw_activation_type,
conv_mod_activation_type=args.
custom_enc_conv_mod_activation_type,
aux_enc_output_layers=aux_enc_output_layers,
)
encoder_out = self.encoder.enc_out
else:
self.enc = encoder_for(
args,
idim,
self.subsample,
aux_enc_output_layers=aux_enc_output_layers,
)
encoder_out = args.eprojs
if args.dtype == "custom":
if args.dec_block_arch is None:
raise ValueError(
"When specifying custom decoder type, --dec-block-arch"
"should also be specified in training config. See"
"egs/vivos/asr1/conf/transducer/train_*.yaml for more info."
)
self.decoder = CustomDecoder(
odim,
args.dec_block_arch,
input_layer=args.custom_dec_input_layer,
repeat_block=args.dec_block_repeat,
positionwise_activation_type=args.
custom_dec_pw_activation_type,
dropout_rate_embed=args.dropout_rate_embed_decoder,
blank_id=blank_id,
)
decoder_out = self.decoder.dunits
else:
self.dec = RNNDecoder(
odim,
args.dtype,
args.dlayers,
args.dunits,
args.dec_embed_dim,
dropout_rate=args.dropout_rate_decoder,
dropout_rate_embed=args.dropout_rate_embed_decoder,
blank_id=blank_id,
)
decoder_out = args.dunits
self.transducer_tasks = TransducerTasks(
encoder_out,
decoder_out,
args.joint_dim,
odim,
joint_activation_type=args.joint_activation_type,
transducer_loss_weight=args.transducer_weight,
ctc_loss=args.use_ctc_loss,
ctc_loss_weight=args.ctc_loss_weight,
ctc_loss_dropout_rate=args.ctc_loss_dropout_rate,
lm_loss=args.use_lm_loss,
lm_loss_weight=args.lm_loss_weight,
lm_loss_smoothing_rate=args.lm_loss_smoothing_rate,
aux_transducer_loss=args.use_aux_transducer_loss,
aux_transducer_loss_weight=args.aux_transducer_loss_weight,
aux_transducer_loss_mlp_dim=args.aux_transducer_loss_mlp_dim,
aux_trans_loss_mlp_dropout_rate=args.
aux_transducer_loss_mlp_dropout_rate,
symm_kl_div_loss=args.use_symm_kl_div_loss,
symm_kl_div_loss_weight=args.symm_kl_div_loss_weight,
fastemit_lambda=args.fastemit_lambda,
blank_id=blank_id,
ignore_id=ignore_id,
training=training,
)
if training and (args.report_cer or args.report_wer):
self.error_calculator = ErrorCalculator(
self.decoder if args.dtype == "custom" else self.dec,
self.transducer_tasks.joint_network,
args.char_list,
args.sym_space,
args.sym_blank,
args.report_cer,
args.report_wer,
)
else:
self.error_calculator = None
self.etype = args.etype
self.dtype = args.dtype
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.default_parameters(args)
self.loss = None
self.rnnlm = None
def __init__(self,
idim,
odim,
args,
ignore_id=-1,
blank_id=0,
training=True):
"""Construct an E2E object for transducer model."""
torch.nn.Module.__init__(self)
args = fill_missing_args(args, self.add_arguments)
self.is_rnnt = True
self.transducer_weight = args.transducer_weight
self.use_aux_task = (True if (args.aux_task_type is not None
and training) else False)
self.use_aux_ctc = args.aux_ctc and training
self.aux_ctc_weight = args.aux_ctc_weight
self.use_aux_cross_entropy = args.aux_cross_entropy and training
self.aux_cross_entropy_weight = args.aux_cross_entropy_weight
if self.use_aux_task:
n_layers = ((len(args.enc_block_arch) * args.enc_block_repeat -
1) if args.enc_block_arch is not None else
(args.elayers - 1))
aux_task_layer_list = valid_aux_task_layer_list(
args.aux_task_layer_list,
n_layers,
)
else:
aux_task_layer_list = []
if "custom" in args.etype:
if args.enc_block_arch is None:
raise ValueError(
"When specifying custom encoder type, --enc-block-arch"
"should also be specified in training config. See"
"egs/vivos/asr1/conf/transducer/train_*.yaml for more info."
)
self.subsample = get_subsample(args,
mode="asr",
arch="transformer")
self.encoder = CustomEncoder(
idim,
args.enc_block_arch,
input_layer=args.custom_enc_input_layer,
repeat_block=args.enc_block_repeat,
self_attn_type=args.custom_enc_self_attn_type,
positional_encoding_type=args.
custom_enc_positional_encoding_type,
positionwise_activation_type=args.
custom_enc_pw_activation_type,
conv_mod_activation_type=args.
custom_enc_conv_mod_activation_type,
aux_task_layer_list=aux_task_layer_list,
)
encoder_out = 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,
aux_task_layer_list=aux_task_layer_list,
)
encoder_out = args.eprojs
if "custom" in args.dtype:
if args.dec_block_arch is None:
raise ValueError(
"When specifying custom decoder type, --dec-block-arch"
"should also be specified in training config. See"
"egs/vivos/asr1/conf/transducer/train_*.yaml for more info."
)
self.decoder = CustomDecoder(
odim,
args.dec_block_arch,
input_layer=args.custom_dec_input_layer,
repeat_block=args.dec_block_repeat,
positionwise_activation_type=args.
custom_dec_pw_activation_type,
dropout_rate_embed=args.dropout_rate_embed_decoder,
)
decoder_out = self.decoder.dunits
if "custom" in args.etype:
self.most_dom_list += args.dec_block_arch[:]
else:
self.most_dom_list = args.dec_block_arch[:]
else:
self.dec = DecoderRNNT(
odim,
args.dtype,
args.dlayers,
args.dunits,
blank_id,
args.dec_embed_dim,
args.dropout_rate_decoder,
args.dropout_rate_embed_decoder,
)
decoder_out = args.dunits
self.joint_network = JointNetwork(odim, encoder_out, decoder_out,
args.joint_dim,
args.joint_activation_type)
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.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.error_calculator = None
self.default_parameters(args)
if training:
self.criterion = TransLoss(args.trans_type, self.blank_id)
decoder = self.decoder if self.dtype == "custom" else self.dec
if args.report_cer or args.report_wer:
self.error_calculator = ErrorCalculator(
decoder,
self.joint_network,
args.char_list,
args.sym_space,
args.sym_blank,
args.report_cer,
args.report_wer,
)
if self.use_aux_task:
self.auxiliary_task = AuxiliaryTask(
decoder,
self.joint_network,
self.criterion,
args.aux_task_type,
args.aux_task_weight,
encoder_out,
args.joint_dim,
)
if self.use_aux_ctc:
self.aux_ctc = ctc_for(
Namespace(
num_encs=1,
eprojs=encoder_out,
dropout_rate=args.aux_ctc_dropout_rate,
ctc_type="warpctc",
),
odim,
)
if self.use_aux_cross_entropy:
self.aux_decoder_output = torch.nn.Linear(decoder_out, odim)
self.aux_cross_entropy = LabelSmoothingLoss(
odim, ignore_id, args.aux_cross_entropy_smoothing)
self.loss = None
self.rnnlm = None