def __init__(self, idim, odim, args):
"""Initialize multi-speaker modules.
Args:
idim (int): dimension of inputs
odim (int): dimension of outputs
args (Namespace): argument Namespace containing options
"""
torch.nn.Module.__init__(self)
self.mtlalpha = args.mtlalpha
assert 0.0 <= self.mtlalpha <= 1.0, "mtlalpha should be [0.0, 1.0]"
self.rnnt_mode = args.rnnt_mode
self.etype = args.etype
self.verbose = args.verbose
self.char_list = args.char_list
self.outdir = args.outdir
self.reporter = Reporter()
self.num_spkrs = args.num_spkrs
self.spa = args.spa
self.pit = PIT(self.num_spkrs)
# below means the last number becomes eos/sos ID
# note that sos/eos IDs are identical
self.sos = odim - 1
self.eos = odim - 1
# subsample info
# +1 means input (+1) and layers outputs (args.elayer_sd + args.elayers)
subsample = np.ones(args.elayers_sd + args.elayers + 1, dtype=np.int)
if args.etype.endswith("p") and not args.etype.startswith("vgg"):
ss = args.subsample.split("_")
for j in range(min(args.elayers_sd + args.elayers + 1, len(ss))):
subsample[j] = int(ss[j])
else:
logging.warning(
'Subsampling is not performed for vgg*. It is performed in max pooling layers at CNN.'
)
logging.info('subsample: ' + ' '.join([str(x) for x in subsample]))
self.subsample = subsample
# label smoothing info
if args.lsm_type and os.path.isfile(args.train_json):
logging.info("Use label smoothing with " + args.lsm_type)
labeldist = label_smoothing_dist(odim,
args.lsm_type,
transcript=args.train_json)
else:
labeldist = None
if getattr(args, "use_frontend",
False): # use getattr to keep compatibility
# Relative importing because of using python3 syntax
from espnet.nets.pytorch_backend.frontends.feature_transform \
import feature_transform_for
from espnet.nets.pytorch_backend.frontends.frontend \
import frontend_for
self.frontend = frontend_for(args, idim)
self.feature_transform = feature_transform_for(
args, (idim - 1) * 2)
idim = args.n_mels
else:
self.frontend = None
# encoder
self.enc = encoder_for(args, idim, self.subsample)
# ctc
self.ctc = ctc_for(args, odim, reduce=False)
if args.rnnt_mode == 'rnnt-att':
# attention
num_att = self.num_spkrs if args.spa else 1
self.att = att_for(args, num_att)
# decoder
self.dec = decoder_for(args, odim, self.att)
else:
# prediction
self.dec = decoder_for(args, odim)
# weight initialization
self.init_like_chainer()
# options for beam search
if 'report_cer' in vars(args) and (args.report_cer or args.report_wer):
recog_args = {
'beam_size': args.beam_size,
'penalty': args.penalty,
'ctc_weight': args.ctc_weight,
'maxlenratio': args.maxlenratio,
'minlenratio': args.minlenratio,
'lm_weight': args.lm_weight,
'rnnlm': args.rnnlm,
'nbest': args.nbest,
'space': args.sym_space,
'blank': args.sym_blank
}
self.recog_args = argparse.Namespace(**recog_args)
self.report_cer = args.report_cer
self.report_wer = args.report_wer
else:
self.report_cer = False
self.report_wer = False
self.rnnlm = None
self.logzero = -10000000000.0
self.loss = None
self.acc = None
def __init__(self, idim, odim, args):
"""Initialize transducer modules.
Args:
idim (int): dimension of inputs
odim (int): dimension of outputs
args (Namespace): argument Namespace containing options
"""
super(E2E, self).__init__()
torch.nn.Module.__init__(self)
self.rnnt_mode = args.rnnt_mode
self.etype = args.etype
self.verbose = args.verbose
self.char_list = args.char_list
self.outdir = args.outdir
self.space = args.sym_space
self.blank = args.sym_blank
self.reporter = Reporter()
# note that eos is the same as sos (equivalent ID)
self.sos = odim - 1
self.eos = odim - 1
# subsample info
# +1 means input (+1) and layers outputs (args.elayer)
subsample = np.ones(args.elayers + 1, dtype=np.int)
if args.etype.endswith("p") and not args.etype.startswith("vgg"):
ss = args.subsample.split("_")
for j in range(min(args.elayers + 1, len(ss))):
subsample[j] = int(ss[j])
else:
logging.warning(
'Subsampling is not performed for vgg*. It is performed in max pooling layers at CNN.')
logging.info('subsample: ' + ' '.join([str(x) for x in subsample]))
self.subsample = subsample
if args.use_frontend:
# Relative importing because of using python3 syntax
from espnet.nets.pytorch_backend.frontends.feature_transform \
import feature_transform_for
from espnet.nets.pytorch_backend.frontends.frontend \
import frontend_for
self.frontend = frontend_for(args, idim)
self.feature_transform = feature_transform_for(args, (idim - 1) * 2)
idim = args.n_mels
else:
self.frontend = None
# encoder
self.enc = encoder_for(args, idim, self.subsample)
if args.rnnt_mode == 'rnnt-att':
# attention
self.att = att_for(args)
# decoder
self.dec = decoder_for(args, odim, self.att)
else:
# prediction
self.dec = decoder_for(args, odim)
# weight initialization
self.init_like_chainer()
# options for beam search
if 'report_cer' in vars(args) and (args.report_cer or args.report_wer):
recog_args = {'beam_size': args.beam_size, 'nbest': args.nbest,
'space': args.sym_space,
'score_norm_transducer': args.score_norm_transducer}
self.recog_args = argparse.Namespace(**recog_args)
self.report_cer = args.report_cer
self.report_wer = args.report_wer
else:
self.report_cer = False
self.report_wer = False
self.logzero = -10000000000.0
self.rnnlm = None
self.loss = None
def __init__(self, idim, odim, args):
"""Initialize transducer modules.
Args:
idim (int): dimension of inputs
odim (int): dimension of outputs
args (Namespace): argument Namespace containing options
"""
super(E2E, self).__init__()
torch.nn.Module.__init__(self)
self.rnnt_mode = args.rnnt_mode
self.etype = args.etype
self.verbose = args.verbose
self.char_list = args.char_list
self.outdir = args.outdir
self.space = args.sym_space
self.blank = args.sym_blank
self.reporter = Reporter()
self.beam_size = args.beam_size
# note that eos is the same as sos (equivalent ID)
self.sos = odim - 1
self.eos = odim - 1
# subsample info
self.subsample = get_subsample(args, mode='asr', arch='rnn-t')
if args.use_frontend:
# Relative importing because of using python3 syntax
from espnet.nets.pytorch_backend.frontends.feature_transform \
import feature_transform_for
from espnet.nets.pytorch_backend.frontends.frontend \
import frontend_for
self.frontend = frontend_for(args, idim)
self.feature_transform = feature_transform_for(
args, (idim - 1) * 2)
idim = args.n_mels
else:
self.frontend = None
# encoder
self.enc = encoder_for(args, idim, self.subsample)
if args.rnnt_mode == 'rnnt-att':
# attention
self.att = att_for(args)
# decoder
self.dec = decoder_for(args, odim, self.att)
else:
# prediction
self.dec = decoder_for(args, odim)
# weight initialization
self.init_like_chainer()
# options for beam search
if 'report_cer' in vars(args) and (args.report_cer or args.report_wer):
recog_args = {
'beam_size': args.beam_size,
'nbest': args.nbest,
'space': args.sym_space,
'score_norm_transducer': args.score_norm_transducer
}
self.recog_args = argparse.Namespace(**recog_args)
self.report_cer = args.report_cer
self.report_wer = args.report_wer
else:
self.report_cer = False
self.report_wer = False
self.logzero = -10000000000.0
self.rnnlm = None
self.loss = None
