def build_decoder(args,
special_symbols,
enc_n_units,
vocab,
ctc_weight,
ctc_fc_list,
global_weight,
external_lm=None):
if args.dec_type in ['transformer', 'transformer_xl']:
from neural_sp.models.seq2seq.decoders.transformer import TransformerDecoder
decoder = TransformerDecoder(
special_symbols=special_symbols,
enc_n_units=enc_n_units,
attn_type=args.transformer_attn_type,
n_heads=args.transformer_n_heads,
n_layers=args.dec_n_layers,
d_model=args.transformer_d_model,
d_ff=args.transformer_d_ff,
ffn_bottleneck_dim=getattr(args, 'transformer_ffn_bottleneck_dim',
0),
layer_norm_eps=args.transformer_layer_norm_eps,
ffn_activation=args.transformer_ffn_activation,
pe_type=args.transformer_dec_pe_type,
vocab=vocab,
tie_embedding=args.tie_embedding,
dropout=args.dropout_dec,
dropout_emb=args.dropout_emb,
dropout_att=args.dropout_att,
dropout_layer=args.dropout_dec_layer,
dropout_head=args.dropout_head,
lsm_prob=args.lsm_prob,
ctc_weight=ctc_weight,
ctc_lsm_prob=args.ctc_lsm_prob,
ctc_fc_list=ctc_fc_list,
backward=(dir == 'bwd'),
global_weight=global_weight,
mtl_per_batch=args.mtl_per_batch,
param_init=args.transformer_param_init,
memory_transformer=args.dec_type == 'transformer_xl',
mem_len=args.mem_len,
mocha_chunk_size=args.mocha_chunk_size,
mocha_n_heads_mono=args.mocha_n_heads_mono,
mocha_n_heads_chunk=args.mocha_n_heads_chunk,
mocha_init_r=args.mocha_init_r,
mocha_eps=args.mocha_eps,
mocha_std=args.mocha_std,
mocha_no_denominator=args.mocha_no_denominator,
mocha_1dconv=args.mocha_1dconv,
mocha_quantity_loss_weight=args.mocha_quantity_loss_weight,
mocha_head_divergence_loss_weight=args.
mocha_head_divergence_loss_weight,
latency_metric=args.mocha_latency_metric,
latency_loss_weight=args.mocha_latency_loss_weight,
mocha_first_layer=args.mocha_first_layer,
share_chunkwise_attention=getattr(args,
'share_chunkwise_attention',
False),
external_lm=external_lm,
lm_fusion=args.lm_fusion)
elif args.dec_type in ['lstm_transducer', 'gru_transducer']:
from neural_sp.models.seq2seq.decoders.rnn_transducer import RNNTransducer
decoder = RNNTransducer(
special_symbols=special_symbols,
enc_n_units=enc_n_units,
rnn_type=args.dec_type,
n_units=args.dec_n_units,
n_projs=args.dec_n_projs,
n_layers=args.dec_n_layers,
bottleneck_dim=args.dec_bottleneck_dim,
emb_dim=args.emb_dim,
vocab=vocab,
dropout=args.dropout_dec,
dropout_emb=args.dropout_emb,
lsm_prob=args.lsm_prob,
ctc_weight=ctc_weight,
ctc_lsm_prob=args.ctc_lsm_prob,
ctc_fc_list=ctc_fc_list,
external_lm=external_lm if args.lm_init else None,
global_weight=global_weight,
mtl_per_batch=args.mtl_per_batch,
param_init=args.param_init)
else:
from neural_sp.models.seq2seq.decoders.las import RNNDecoder
decoder = RNNDecoder(
special_symbols=special_symbols,
enc_n_units=enc_n_units,
rnn_type=args.dec_type,
n_units=args.dec_n_units,
n_projs=args.dec_n_projs,
n_layers=args.dec_n_layers,
bottleneck_dim=args.dec_bottleneck_dim,
emb_dim=args.emb_dim,
vocab=vocab,
tie_embedding=args.tie_embedding,
attn_type=args.attn_type,
attn_dim=args.attn_dim,
attn_sharpening_factor=args.attn_sharpening_factor,
attn_sigmoid_smoothing=args.attn_sigmoid,
attn_conv_out_channels=args.attn_conv_n_channels,
attn_conv_kernel_size=args.attn_conv_width,
attn_n_heads=args.attn_n_heads,
dropout=args.dropout_dec,
dropout_emb=args.dropout_emb,
dropout_att=args.dropout_att,
lsm_prob=args.lsm_prob,
ss_prob=args.ss_prob,
ss_type=args.ss_type,
ctc_weight=ctc_weight,
ctc_lsm_prob=args.ctc_lsm_prob,
ctc_fc_list=ctc_fc_list,
mbr_training=args.mbr_training,
mbr_ce_weight=args.mbr_ce_weight,
external_lm=external_lm,
lm_fusion=args.lm_fusion,
lm_init=args.lm_init,
backward=(dir == 'bwd'),
global_weight=global_weight,
mtl_per_batch=args.mtl_per_batch,
param_init=args.param_init,
mocha_chunk_size=args.mocha_chunk_size,
mocha_n_heads_mono=args.mocha_n_heads_mono,
mocha_init_r=args.mocha_init_r,
mocha_eps=args.mocha_eps,
mocha_std=args.mocha_std,
mocha_no_denominator=args.mocha_no_denominator,
mocha_1dconv=args.mocha_1dconv,
mocha_quantity_loss_weight=args.mocha_quantity_loss_weight,
latency_metric=args.mocha_latency_metric,
latency_loss_weight=args.mocha_latency_loss_weight,
gmm_attn_n_mixtures=args.gmm_attn_n_mixtures,
replace_sos=args.replace_sos,
distillation_weight=args.distillation_weight,
discourse_aware=args.discourse_aware)
return decoder
def __init__(self, args, save_path=None):
super(ModelBase, self).__init__()
self.save_path = save_path
# for encoder, decoder
self.input_type = args.input_type
self.input_dim = args.input_dim
self.enc_type = args.enc_type
self.enc_n_units = args.enc_n_units
if args.enc_type in ['blstm', 'bgru', 'conv_blstm', 'conv_bgru']:
self.enc_n_units *= 2
self.dec_type = args.dec_type
# for OOV resolution
self.enc_n_layers = args.enc_n_layers
self.enc_n_layers_sub1 = args.enc_n_layers_sub1
self.subsample = [int(s) for s in args.subsample.split('_')]
# for decoder
self.vocab = args.vocab
self.vocab_sub1 = args.vocab_sub1
self.vocab_sub2 = args.vocab_sub2
self.blank = 0
self.unk = 1
self.eos = 2
self.pad = 3
# NOTE: reserved in advance
# for the sub tasks
self.main_weight = 1 - args.sub1_weight - args.sub2_weight
self.sub1_weight = args.sub1_weight
self.sub2_weight = args.sub2_weight
self.mtl_per_batch = args.mtl_per_batch
self.task_specific_layer = args.task_specific_layer
# for CTC
self.ctc_weight = min(args.ctc_weight, self.main_weight)
self.ctc_weight_sub1 = min(args.ctc_weight_sub1, self.sub1_weight)
self.ctc_weight_sub2 = min(args.ctc_weight_sub2, self.sub2_weight)
# for backward decoder
self.bwd_weight = min(args.bwd_weight, self.main_weight)
self.fwd_weight = self.main_weight - self.bwd_weight - self.ctc_weight
self.fwd_weight_sub1 = self.sub1_weight - self.ctc_weight_sub1
self.fwd_weight_sub2 = self.sub2_weight - self.ctc_weight_sub2
# Feature extraction
self.gaussian_noise = args.gaussian_noise
self.n_stacks = args.n_stacks
self.n_skips = args.n_skips
self.n_splices = args.n_splices
self.is_specaug = args.n_freq_masks > 0 or args.n_time_masks > 0
self.specaug = None
if self.is_specaug:
assert args.n_stacks == 1 and args.n_skips == 1
assert args.n_splices == 1
self.specaug = SpecAugment(F=args.freq_width,
T=args.time_width,
n_freq_masks=args.n_freq_masks,
n_time_masks=args.n_time_masks,
p=args.time_width_upper)
# Frontend
self.ssn = None
if args.sequence_summary_network:
assert args.input_type == 'speech'
self.ssn = SequenceSummaryNetwork(args.input_dim,
n_units=512,
n_layers=3,
bottleneck_dim=100,
dropout=0,
param_init=args.param_init)
# Encoder
self.enc = select_encoder(args)
if args.freeze_encoder:
for p in self.enc.parameters():
p.requires_grad = False
# main task
directions = []
if self.fwd_weight > 0 or self.ctc_weight > 0:
directions.append('fwd')
if self.bwd_weight > 0:
directions.append('bwd')
for dir in directions:
# Load the LM for LM fusion
if args.lm_fusion and dir == 'fwd':
lm_fusion = RNNLM(args.lm_conf)
lm_fusion, _ = load_checkpoint(lm_fusion, args.lm_fusion)
else:
lm_fusion = None
# TODO(hirofumi): for backward RNNLM
# Load the LM for LM initialization
if args.lm_init and dir == 'fwd':
lm_init = RNNLM(args.lm_conf)
lm_init, _ = load_checkpoint(lm_init, args.lm_init)
else:
lm_init = None
# TODO(hirofumi): for backward RNNLM
# Decoder
if args.dec_type == 'transformer':
dec = TransformerDecoder(
eos=self.eos,
unk=self.unk,
pad=self.pad,
blank=self.blank,
enc_n_units=self.enc.output_dim,
attn_type=args.transformer_attn_type,
attn_n_heads=args.transformer_attn_n_heads,
n_layers=args.dec_n_layers,
d_model=args.d_model,
d_ff=args.d_ff,
vocab=self.vocab,
tie_embedding=args.tie_embedding,
pe_type=args.pe_type,
layer_norm_eps=args.layer_norm_eps,
dropout=args.dropout_dec,
dropout_emb=args.dropout_emb,
dropout_att=args.dropout_att,
lsm_prob=args.lsm_prob,
focal_loss_weight=args.focal_loss_weight,
focal_loss_gamma=args.focal_loss_gamma,
ctc_weight=self.ctc_weight if dir == 'fwd' else 0,
ctc_lsm_prob=args.ctc_lsm_prob,
ctc_fc_list=[
int(fc) for fc in args.ctc_fc_list.split('_')
] if args.ctc_fc_list is not None
and len(args.ctc_fc_list) > 0 else [],
backward=(dir == 'bwd'),
global_weight=self.main_weight -
self.bwd_weight if dir == 'fwd' else self.bwd_weight,
mtl_per_batch=args.mtl_per_batch)
elif 'transducer' in args.dec_type:
dec = RNNTransducer(
eos=self.eos,
unk=self.unk,
pad=self.pad,
blank=self.blank,
enc_n_units=self.enc.output_dim,
rnn_type=args.dec_type,
n_units=args.dec_n_units,
n_projs=args.dec_n_projs,
n_layers=args.dec_n_layers,
residual=args.dec_residual,
bottleneck_dim=args.dec_bottleneck_dim,
emb_dim=args.emb_dim,
vocab=self.vocab,
dropout=args.dropout_dec,
dropout_emb=args.dropout_emb,
lsm_prob=args.lsm_prob,
ctc_weight=self.ctc_weight if dir == 'fwd' else 0,
ctc_lsm_prob=args.ctc_lsm_prob,
ctc_fc_list=[
int(fc) for fc in args.ctc_fc_list.split('_')
] if args.ctc_fc_list is not None
and len(args.ctc_fc_list) > 0 else [],
lm_init=lm_init,
lmobj_weight=args.lmobj_weight,
share_lm_softmax=args.share_lm_softmax,
global_weight=self.main_weight -
self.bwd_weight if dir == 'fwd' else self.bwd_weight,
mtl_per_batch=args.mtl_per_batch,
param_init=args.param_init)
else:
dec = RNNDecoder(
eos=self.eos,
unk=self.unk,
pad=self.pad,
blank=self.blank,
enc_n_units=self.enc.output_dim,
attn_type=args.attn_type,
attn_dim=args.attn_dim,
attn_sharpening_factor=args.attn_sharpening,
attn_sigmoid_smoothing=args.attn_sigmoid,
attn_conv_out_channels=args.attn_conv_n_channels,
attn_conv_kernel_size=args.attn_conv_width,
attn_n_heads=args.attn_n_heads,
rnn_type=args.dec_type,
n_units=args.dec_n_units,
n_projs=args.dec_n_projs,
n_layers=args.dec_n_layers,
loop_type=args.dec_loop_type,
residual=args.dec_residual,
bottleneck_dim=args.dec_bottleneck_dim,
emb_dim=args.emb_dim,
vocab=self.vocab,
tie_embedding=args.tie_embedding,
dropout=args.dropout_dec,
dropout_emb=args.dropout_emb,
dropout_att=args.dropout_att,
zoneout=args.zoneout,
ss_prob=args.ss_prob,
ss_type=args.ss_type,
lsm_prob=args.lsm_prob,
focal_loss_weight=args.focal_loss_weight,
focal_loss_gamma=args.focal_loss_gamma,
ctc_weight=self.ctc_weight if dir == 'fwd' else 0,
ctc_lsm_prob=args.ctc_lsm_prob,
ctc_fc_list=[
int(fc) for fc in args.ctc_fc_list.split('_')
] if args.ctc_fc_list is not None
and len(args.ctc_fc_list) > 0 else [],
input_feeding=args.input_feeding,
backward=(dir == 'bwd'),
lm_fusion=lm_fusion,
lm_fusion_type=args.lm_fusion_type,
discourse_aware=args.discourse_aware,
lm_init=lm_init,
lmobj_weight=args.lmobj_weight,
share_lm_softmax=args.share_lm_softmax,
global_weight=self.main_weight -
self.bwd_weight if dir == 'fwd' else self.bwd_weight,
mtl_per_batch=args.mtl_per_batch,
adaptive_softmax=args.adaptive_softmax,
param_init=args.param_init,
replace_sos=args.replace_sos)
setattr(self, 'dec_' + dir, dec)
# sub task
for sub in ['sub1', 'sub2']:
if getattr(self, sub + '_weight') > 0:
if args.dec_type == 'transformer':
raise NotImplementedError
else:
dec_sub = RNNDecoder(
eos=self.eos,
unk=self.unk,
pad=self.pad,
blank=self.blank,
enc_n_units=self.enc_n_units,
attn_type=args.attn_type,
attn_dim=args.attn_dim,
attn_sharpening_factor=args.attn_sharpening,
attn_sigmoid_smoothing=args.attn_sigmoid,
attn_conv_out_channels=args.attn_conv_n_channels,
attn_conv_kernel_size=args.attn_conv_width,
attn_n_heads=1,
rnn_type=args.dec_type,
n_units=args.dec_n_units,
n_projs=args.dec_n_projs,
n_layers=args.dec_n_layers,
loop_type=args.dec_loop_type,
residual=args.dec_residual,
bottleneck_dim=args.dec_bottleneck_dim,
emb_dim=args.emb_dim,
tie_embedding=args.tie_embedding,
vocab=getattr(self, 'vocab_' + sub),
dropout=args.dropout_dec,
dropout_emb=args.dropout_emb,
dropout_att=args.dropout_att,
ss_prob=args.ss_prob,
ss_type=args.ss_type,
lsm_prob=args.lsm_prob,
focal_loss_weight=args.focal_loss_weight,
focal_loss_gamma=args.focal_loss_gamma,
ctc_weight=getattr(self, 'ctc_weight_' + sub),
ctc_lsm_prob=args.ctc_lsm_prob,
ctc_fc_list=[
int(fc) for fc in getattr(args, 'ctc_fc_list_' +
sub).split('_')
] if getattr(args, 'ctc_fc_list_' + sub) is not None
and len(getattr(args, 'ctc_fc_list_' + sub)) > 0 else
[],
input_feeding=args.input_feeding,
global_weight=getattr(self, sub + '_weight'),
mtl_per_batch=args.mtl_per_batch,
param_init=args.param_init)
setattr(self, 'dec_fwd_' + sub, dec_sub)
if args.input_type == 'text':
if args.vocab == args.vocab_sub1:
# Share the embedding layer between input and output
self.embed = dec.embed
else:
self.embed = Embedding(vocab=args.vocab_sub1,
emb_dim=args.emb_dim,
dropout=args.dropout_emb,
ignore_index=self.pad)
# Recurrent weights are orthogonalized
if args.rec_weight_orthogonal:
self.reset_parameters(args.param_init,
dist='orthogonal',
keys=['rnn', 'weight'])
# Initialize bias in forget gate with 1
# self.init_forget_gate_bias_with_one()
# Fix all parameters except for the gating parts in deep fusion
if args.lm_fusion_type == 'deep' and args.lm_fusion:
for n, p in self.named_parameters():
if 'output' in n or 'output_bn' in n or 'linear' in n:
p.requires_grad = True
else:
p.requires_grad = False
def build_decoder(args,
special_symbols,
enc_n_units,
vocab,
ctc_weight,
ctc_fc_list,
global_weight,
lm_fusion=None,
lm_init=None):
if args.dec_type == 'transformer':
if args.attn_type == 'cif':
raise NotImplementedError
else:
from neural_sp.models.seq2seq.decoders.transformer import TransformerDecoder
decoder = TransformerDecoder(
special_symbols=special_symbols,
enc_n_units=enc_n_units,
attn_type=args.transformer_attn_type,
n_heads=args.transformer_n_heads,
n_layers=args.dec_n_layers,
d_model=args.transformer_d_model,
d_ff=args.transformer_d_ff,
layer_norm_eps=args.transformer_layer_norm_eps,
ffn_activation=args.transformer_ffn_activation,
pe_type=args.transformer_dec_pe_type,
vocab=vocab,
tie_embedding=args.tie_embedding,
dropout=args.dropout_dec,
dropout_emb=args.dropout_emb,
dropout_att=args.dropout_att,
lsm_prob=args.lsm_prob,
ctc_weight=ctc_weight,
ctc_lsm_prob=args.ctc_lsm_prob,
ctc_fc_list=ctc_fc_list,
backward=(dir == 'bwd'),
global_weight=global_weight,
mtl_per_batch=args.mtl_per_batch,
param_init=args.transformer_param_init)
elif args.dec_type == 'transformer_transducer':
raise NotImplementedError
from neural_sp.models.seq2seq.decoders.transformer_transducer import TrasformerTransducer
decoder = TrasformerTransducer(
special_symbols=special_symbols,
enc_n_units=enc_n_units,
attn_type=args.transformer_attn_type,
n_heads=args.transformer_n_heads,
n_layers=args.dec_n_layers,
d_model=args.transformer_d_model,
d_ff=args.transformer_d_ff,
layer_norm_eps=args.transformer_layer_norm_eps,
ffn_activation=args.transformer_ffn_activation,
pe_type=args.transformer_dec_pe_type,
vocab=vocab,
dropout=args.dropout_dec,
dropout_emb=args.dropout_emb,
dropout_att=args.dropout_att,
lsm_prob=args.lsm_prob,
ctc_weight=ctc_weight,
ctc_lsm_prob=args.ctc_lsm_prob,
ctc_fc_list=ctc_fc_list,
lm_init=lm_init,
global_weight=global_weight,
mtl_per_batch=args.mtl_per_batch,
param_init=args.transformer_param_init)
elif args.dec_type in ['lstm_transducer', 'gru_transducer']:
from neural_sp.models.seq2seq.decoders.rnn_transducer import RNNTransducer
decoder = RNNTransducer(special_symbols=special_symbols,
enc_n_units=enc_n_units,
rnn_type=args.dec_type,
n_units=args.dec_n_units,
n_projs=args.dec_n_projs,
n_layers=args.dec_n_layers,
bottleneck_dim=args.dec_bottleneck_dim,
emb_dim=args.emb_dim,
vocab=vocab,
dropout=args.dropout_dec,
dropout_emb=args.dropout_emb,
lsm_prob=args.lsm_prob,
ctc_weight=ctc_weight,
ctc_lsm_prob=args.ctc_lsm_prob,
ctc_fc_list=ctc_fc_list,
lm_init=lm_init,
global_weight=global_weight,
mtl_per_batch=args.mtl_per_batch,
param_init=args.param_init)
else:
from neural_sp.models.seq2seq.decoders.las import RNNDecoder
decoder = RNNDecoder(
special_symbols=special_symbols,
enc_n_units=enc_n_units,
attn_type=args.attn_type,
attn_dim=args.attn_dim,
attn_sharpening_factor=args.attn_sharpening_factor,
attn_sigmoid_smoothing=args.attn_sigmoid,
attn_conv_out_channels=args.attn_conv_n_channels,
attn_conv_kernel_size=args.attn_conv_width,
attn_n_heads=args.attn_n_heads,
rnn_type=args.dec_type,
n_units=args.dec_n_units,
n_projs=args.dec_n_projs,
n_layers=args.dec_n_layers,
bottleneck_dim=args.dec_bottleneck_dim,
emb_dim=args.emb_dim,
vocab=vocab,
tie_embedding=args.tie_embedding,
dropout=args.dropout_dec,
dropout_emb=args.dropout_emb,
dropout_att=args.dropout_att,
ss_prob=args.ss_prob,
ss_type=args.ss_type,
lsm_prob=args.lsm_prob,
ctc_weight=ctc_weight,
ctc_lsm_prob=args.ctc_lsm_prob,
ctc_fc_list=ctc_fc_list,
mbr_weight=args.mbr_weight,
mbr_nbest=args.mbr_nbest,
mbr_softmax_smoothing=args.mbr_softmax_smoothing,
backward=(dir == 'bwd'),
lm_fusion=lm_fusion,
lm_fusion_type=args.lm_fusion_type,
discourse_aware=args.discourse_aware,
lm_init=lm_init,
global_weight=global_weight,
mtl_per_batch=args.mtl_per_batch,
param_init=args.param_init,
mocha_chunk_size=args.mocha_chunk_size,
mocha_adaptive=args.mocha_adaptive,
mocha_1dconv=args.mocha_1dconv,
mocha_quantity_loss_weight=args.mocha_quantity_loss_weight,
mocha_ctc_sync=args.mocha_ctc_sync,
gmm_attn_n_mixtures=args.gmm_attn_n_mixtures,
replace_sos=args.replace_sos,
soft_label_weight=args.soft_label_weight)
return decoder