def __init__(self, idims, odim, args):
"""Initialize this class with python-level args.
Args:
idims (list): list of the number of an input feature dim.
odim (int): The number of output vocab.
args (Namespace): arguments
"""
super(E2E, self).__init__()
torch.nn.Module.__init__(self)
self.mtlalpha = args.mtlalpha
assert 0.0 <= self.mtlalpha <= 1.0, "mtlalpha should be [0.0, 1.0]"
self.verbose = args.verbose
# NOTE: for self.build method
args.char_list = getattr(args, "char_list", None)
self.char_list = args.char_list
self.outdir = args.outdir
self.space = args.sym_space
self.blank = args.sym_blank
self.reporter = Reporter()
self.num_encs = args.num_encs
self.share_ctc = args.share_ctc
# 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
self.subsample_list = get_subsample(args,
mode="asr",
arch="rnn_mulenc")
# 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
# speech translation related
self.replace_sos = getattr(args, "replace_sos",
False) # use getattr to keep compatibility
self.frontend = None
# encoder
self.enc = encoder_for(args, idims, self.subsample_list)
# ctc
self.ctc = ctc_for(args, odim)
# attention
self.att = att_for(args)
# hierarchical attention network
han = att_for(args, han_mode=True)
self.att.append(han)
# decoder
self.dec = decoder_for(args, odim, self.sos, self.eos, self.att,
labeldist)
if args.mtlalpha > 0 and self.num_encs > 1:
# weights-ctc,
# e.g. ctc_loss = w_1*ctc_1_loss + w_2 * ctc_2_loss + w_N * ctc_N_loss
self.weights_ctc_train = args.weights_ctc_train / np.sum(
args.weights_ctc_train) # normalize
self.weights_ctc_dec = args.weights_ctc_dec / np.sum(
args.weights_ctc_dec) # normalize
logging.info("ctc weights (training during training): " +
" ".join([str(x) for x in self.weights_ctc_train]))
logging.info("ctc weights (decoding during training): " +
" ".join([str(x) for x in self.weights_ctc_dec]))
else:
self.weights_ctc_dec = [1.0]
self.weights_ctc_train = [1.0]
# weight initialization
self.init_like_chainer()
# options for beam search
if 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,
"tgt_lang": False,
"ctc_weights_dec": self.weights_ctc_dec,
}
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, mono_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
"""
super(E2E, self).__init__()
torch.nn.Module.__init__(self)
self.mtlalpha = args.mtlalpha
assert 0.0 <= self.mtlalpha <= 1.0, "mtlalpha should be [0.0, 1.0]"
self.etype = args.etype
self.verbose = args.verbose
# NOTE: for self.build method
self.outdir = args.outdir
# target matching system organization
self.oversampling = args.oversampling
self.residual = args.residual
self.outer = args.outer
self.poster = torch.nn.Linear(args.eprojs, odim * self.oversampling)
self.poster_mono = torch.nn.Linear(args.eprojs,
mono_odim * self.oversampling)
# below means the last number becomes eos/sos ID
# note that sos/eos IDs are identical
self.sos = odim - 1
self.eos = odim - 1
self.sos_mono = mono_odim - 1
self.eos_mono = mono_odim - 1
self.odim = odim
self.mono_odim = mono_odim
self.ignore_id = ignore_id
self.subsample = get_subsample(args, mode="asr", arch="rnn")
self.reporter = Reporter()
# 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
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)
# weight initialization
if args.initializer == "lecun":
self.init_like_chainer()
elif args.initializer == "orthogonal":
self.init_orthogonal()
elif args.initializer == "xavier":
self.init_xavier()
else:
raise NotImplementedError("unknown initializer: " +
args.initializer)
if args.report_cer or args.report_wer:
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
self.logzero = -10000000000.0
self.loss = None
self.acc = None
def __init__(self, idim, odim, args):
super(E2E, self).__init__()
self.mtlalpha = args.mtlalpha
assert 0.0 <= self.mtlalpha <= 1.0, "mtlalpha should be [0.0, 1.0]"
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:
logging.info("Use label smoothing with " + args.lsm_type)
labeldist = label_smoothing_dist(odim,
args.lsm_type,
transcript=args.train_json)
else:
labeldist = None
# encoder
self.enc = encoder_for(args, idim, self.subsample)
# ctc
self.ctc = ctc_for(args, odim, reduce=False)
# 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.sos, self.eos, self.att,
labeldist)
# 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, idims, odim, args):
"""Initialize this class with python-level args.
Args:
idims (list): list of the number of an input feature dim.
odim (int): The number of output vocab.
args (Namespace): arguments
"""
super(E2E, self).__init__()
torch.nn.Module.__init__(self)
self.mtlalpha = args.mtlalpha
assert 0.0 <= self.mtlalpha <= 1.0, "mtlalpha should be [0.0, 1.0]"
self.verbose = args.verbose
# NOTE: for self.build method
args.char_list = getattr(args, "char_list", None)
self.char_list = args.char_list
self.outdir = args.outdir
self.space = args.sym_space
self.blank = args.sym_blank
self.reporter = Reporter()
self.num_encs = args.num_encs
self.share_ctc = args.share_ctc
# 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
self.subsample_list = []
for idx in range(self.num_encs):
# +1 means input (+1) and layers outputs (args.elayer)
subsample = np.ones(args.elayers[idx] + 1, dtype=np.int)
if args.etype[idx].endswith(
"p") and not args.etype[idx].startswith("vgg"):
ss = args.subsample[idx].split("_")
for j in range(min(args.elayers[idx] + 1, len(ss))):
subsample[j] = int(ss[j])
else:
logging.warning(
'Encoder {}: Subsampling is not performed for vgg*. '
'It is performed in max pooling layers at CNN.'.format(
idx + 1))
logging.info('subsample: ' + ' '.join([str(x) for x in subsample]))
self.subsample_list.append(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
# speech translation related
self.replace_sos = getattr(args, "replace_sos",
False) # use getattr to keep compatibility
self.frontend = None
# encoder
self.enc = encoder_for(args, idims, self.subsample_list)
# ctc
self.ctc = ctc_for(args, odim)
# attention
self.att = att_for(args)
# hierarchical attention network
han = att_for(args, han_mode=True)
self.att.append(han)
# decoder
self.dec = decoder_for(args, odim, self.sos, self.eos, self.att,
labeldist)
if args.mtlalpha > 0 and self.num_encs > 1:
# weights-ctc, e.g. ctc_loss = w_1*ctc_1_loss + w_2 * ctc_2_loss + w_N * ctc_N_loss
self.weights_ctc_train = args.weights_ctc_train / np.sum(
args.weights_ctc_train) # normalize
self.weights_ctc_dec = args.weights_ctc_dec / np.sum(
args.weights_ctc_dec) # normalize
logging.info('ctc weights (training during training): ' +
' '.join([str(x) for x in self.weights_ctc_train]))
logging.info('ctc weights (decoding during training): ' +
' '.join([str(x) for x in self.weights_ctc_dec]))
else:
self.weights_ctc_dec = [1.0]
self.weights_ctc_train = [1.0]
# weight initialization
self.init_like_chainer()
# options for beam search
if 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,
'tgt_lang': False,
'ctc_weights_dec': self.weights_ctc_dec
}
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):
super(E2E, self).__init__()
torch.nn.Module.__init__(self)
self.mtlalpha = args.mtlalpha
assert 0.0 <= self.mtlalpha <= 1.0, "mtlalpha should be [0.0, 1.0]"
self.etype = args.etype
self.verbose = args.verbose
# NOTE: for self.build method
args.char_list = getattr(args, "char_list", None)
self.char_list = args.char_list
self.outdir = args.outdir
self.space = args.sym_space
self.blank = args.sym_blank
# self.oracle_length = args.oracle_length
self.reporter = Reporter()
# 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)
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
# 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
# speech translation related
self.replace_sos = getattr(args, "replace_sos",
False) # use getattr to keep compatibility
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)
# attention
self.att = att_for(args)
# decoder
self.dec = decoder_for(args, odim, self.sos, self.eos, self.att,
labeldist)
# weight initialization
self.init_like_chainer()
# options for beam search
if 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,
'tgt_lang': False,
'sampling': args.sampling
}
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
self.loss_nll = torch.nn.NLLLoss()
def __init__(self, idim, odim, args):
"""Construct an E2E object.
:param int idim: dimension of inputs
:param int odim: dimension of outputs
:param Namespace args: argument Namespace containing options
"""
super(E2E, self).__init__()
torch.nn.Module.__init__(self)
# fill missing arguments for compatibility
args = fill_missing_args(args, self.add_arguments)
self.mtlalpha = args.mtlalpha
assert 0.0 <= self.mtlalpha <= 1.0, "mtlalpha should be [0.0, 1.0]"
self.etype = args.etype
self.verbose = args.verbose
# NOTE: for self.build method
args.char_list = getattr(args, "char_list", None)
self.char_list = args.char_list
self.outdir = args.outdir
self.space = args.sym_space
self.blank = args.sym_blank
self.reporter = Reporter()
# 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
self.subsample = get_subsample(args, mode="asr", arch="rnn")
# 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
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)
# attention
self.att = att_for(args)
# decoder
self.dec = decoder_for(args, odim, self.sos, self.eos, self.att,
labeldist)
# weight initialization
self.init_like_chainer()
# options for beam search
if 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 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 multi-speaker E2E module."""
torch.nn.Module.__init__(self)
self.mtlalpha = args.mtlalpha
assert 0.0 <= self.mtlalpha <= 1.0, "mtlalpha should be [0.0, 1.0]"
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
self.subsample = get_subsample(args, mode='asr', arch='rnn_mix')
# 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)
# 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.sos, self.eos, self.att,
labeldist)
# 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, asr_model=None, mt_model=None):
super(E2E, self).__init__()
torch.nn.Module.__init__(self)
self.mtlalpha = args.mtlalpha
assert 0.0 <= self.mtlalpha <= 1.0, "mtlalpha should be [0.0, 1.0]"
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()
# 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)
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
# 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
# speech translation related
self.replace_sos = args.replace_sos
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)
# ctc
self.ctc = ctc_for(args, odim)
# attention
self.att = att_for(args)
# decoder
self.dec = decoder_for(args, odim, self.sos, self.eos, self.att, labeldist)
# weight initialization
self.init_like_chainer()
# pre-training w/ ASR encoder and NMT decoder
if asr_model is not None:
param_dict = dict(asr_model.named_parameters())
for n, p in self.named_parameters():
# overwrite the encoder
if n in param_dict.keys() and p.size() == param_dict[n].size():
if 'enc.enc' in n:
p.data = param_dict[n].data
logging.warning('Overwrite %s' % n)
if mt_model is not None:
param_dict = dict(mt_model.named_parameters())
for n, p in self.named_parameters():
# overwrite the decoder
if n in param_dict.keys() and p.size() == param_dict[n].size():
if 'dec.' in n or 'att' in n:
p.data = param_dict[n].data
logging.warning('Overwrite %s' % n)
# options for beam search
if 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,
'tgt_lang': False}
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):
"""Construct an E2E object.
:param int idim: dimension of inputs
:param int odim: dimension of outputs
:param Namespace args: argument Namespace containing options
"""
super(E2E, self).__init__()
torch.nn.Module.__init__(self)
# fill missing arguments for compatibility
args = fill_missing_args(args, self.add_arguments)
self.mtlalpha = args.mtlalpha
assert 0.0 <= self.mtlalpha <= 1.0, "mtlalpha should be [0.0, 1.0]"
self.etype = args.etype
self.verbose = args.verbose
# NOTE: for self.build method
args.char_list = getattr(args, "char_list", None)
self.char_list = args.char_list
self.outdir = args.outdir
self.space = args.sym_space
self.blank = args.sym_blank
self.reporter = Reporter()
# below means the last number becomes eos/sos ID
# note that sos/eos IDs are identical
self.sos = odim - 1
self.eos = odim - 1
# gs534 - word vocab
bpe = len(self.char_list) > 100 # hack here for bpe flag
self.vocabulary = Vocabulary(args.dictfile, bpe) if args.dictfile != '' else None
# gs534 - create lexicon tree
lextree = None
self.meeting_KB = None
self.n_KBs = getattr(args, 'dynamicKBs', 0)
pretrain_emb = []
if args.meetingKB and args.meetingpath != '':
if self.n_KBs == 0 or not os.path.isdir(os.path.join(args.meetingpath, 'split_0')):
self.meeting_KB = KBmeeting(self.vocabulary, args.meetingpath, args.char_list, bpe)
else:
# arrange multiple KBs
self.meeting_KB = []
for i in range(self.n_KBs):
self.meeting_KB.append(KBmeeting(self.vocabulary,
os.path.join(args.meetingpath, 'split_{}'.format(i)), args.char_list, bpe))
# subsample info
self.subsample = get_subsample(args, mode="asr", arch="rnn")
# 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
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)
# attention
self.att = att_for(args)
# decoder
self.dec = decoder_for(args, odim, self.sos, self.eos, self.att, labeldist,
meetingKB=self.meeting_KB[0] if isinstance(self.meeting_KB, list) else self.meeting_KB)
# weight initialization
self.init_from = getattr(args, 'init_full_model', None)
self.init_like_chainer()
# options for beam search
if 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,
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
def __init__(self, idim, odim, args):
torch.nn.Module.__init__(self)
self.mtlalpha = args.mtlalpha
assert 0.0 <= self.mtlalpha <= 1.0, "mtlalpha should be [0.0, 1.0]"
self.etype = args.etype
self.verbose = args.verbose
self.char_list = args.char_list
self.outdir = args.outdir
self.space = args.sym_space
# self.space = -1
self.blank = args.sym_blank
self.reporter = Reporter()
# 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)
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
# label smoothing info
if args.lsm_type:
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 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)
self.encoder = Encoder(
idim=idim,
center_len=args.transformer_encoder_center_chunk_len,
left_len=args.transformer_encoder_left_chunk_len,
hop_len=args.transformer_encoder_hop_len,
right_len=args.transformer_encoder_right_chunk_len,
abs_pos=args.transformer_encoder_abs_embed,
rel_pos=args.transformer_encoder_rel_embed,
use_mem=args.transformer_encoder_use_memory,
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)
# ctc
self.ctc = ctc_for(args, odim)
# attention
self.att = att_for(args)
# decoder
self.dec = decoder_for(args, odim, self.sos, self.eos, self.att,
labeldist)
# weight initialization
self.init_like_chainer()
# options for beam search
if 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