def train_step(self,
sample,
model,
criterion,
optimizer,
ignore_grad=False):
return MultilingualTranslationTask.train_step(self, sample, model,
criterion, optimizer)
def add_args(parser):
"""Add task-specific arguments to the parser."""
# fmt: off
MultilingualTranslationTask.add_args(parser)
parser.add_argument('--encoder-latent-layer', action='store_true', help='latent layer selection in encoder')
parser.add_argument('--decoder-latent-layer', action='store_true', help='latent layer selection in decoder')
parser.add_argument('--target-layers', default=-1, type=int,
help='number of effective layers to learn; -1 means no constraint')
parser.add_argument('--sparsity-weight', default=0.0, type=float,
help='weight for sparsity loss')
parser.add_argument('--share-weight', default=0.0, type=float,
help='weight for sharing loss')
parser.add_argument('--soft-update', default=1, type=int,
help='number of updates with soft sampling')
parser.add_argument('--anneal-updates', default=1, type=int,
help='number of updates to anneal the KL loss weight')
parser.add_argument('--prior', default="uniform", type=str,
help='prior used for computing KL loss')
def setup_task(cls, args, **kwargs):
dicts, training = MultilingualTranslationTask.prepare(args, **kwargs)
return cls(args, dicts, training)
def add_args(parser):
"""Add task-specific arguments to the parser."""
MultilingualTranslationTask.add_args(parser)
def inference_step(self, generator, models, sample, prefix_tokens=None):
return MultilingualTranslationTask.inference_step(
self, generator, models, sample, prefix_tokens)
def grad_denom(self, sample_sizes, criterion):
return MultilingualTranslationTask.grad_denom(self, sample_sizes, criterion)
def init_logging_output(self, sample):
return MultilingualTranslationTask.init_logging_output(sample)
def aggregate_logging_outputs(self, logging_outputs, criterion):
return MultilingualTranslationTask.aggregate_logging_outputs(
self, logging_outputs, criterion)
def valid_step(self, sample, model, criterion):
return MultilingualTranslationTask.valid_step(self, sample, model,
criterion)
def add_args(parser):
"""Add task-specific arguments to the parser."""
# fmt: off
MultilingualTranslationTask.add_args(parser)
# Speech related args:
parser.add_argument(
"--dataset-from-json",
default=False,
help=
"whether to read the data from a JSON file or from indexed data containing "
"the precomputed filterbanks")
parser.add_argument(
'--skip-normalization',
action='store_true',
default=False,
help='if set, the input filterbanks are not normalized')
parser.add_argument(
'--legacy-audio-fix-lua-indexing',
action='store_true',
default=False,
help=
'if set, the input filterbanks are subtracted 1 to remove +1 for lua indexing'
)
parser.add_argument('--specaugment',
action='store_true',
default=False)
parser.add_argument(
'--frequency-masking-pars',
type=int,
default=13,
help="Maximum number of frequencies that can be masked")
parser.add_argument(
'--time-masking-pars',
type=int,
default=13,
help="Maximum number of time steps that can be masked")
parser.add_argument(
'--frequency-masking-num',
type=int,
default=2,
help="Number of masks to apply along the frequency dimension")
parser.add_argument(
'--time-masking-num',
type=int,
default=2,
help="Number of masks to apply along the time dimension")
parser.add_argument(
'--specaugment-rate',
type=float,
default=1.0,
help="Probability to apply specaugment to a spectrogram")
parser.add_argument(
'--time-stretch',
action='store_true',
help="If set, activates time stretch on spectrograms")
parser.add_argument(
'--time-stretch-rate',
type=float,
default=1.0,
help='Probability to apply time stretch to a spectrogram')
parser.add_argument('--time-stretch-w',
type=int,
default=1,
help='Window size for time stretch')
parser.add_argument('--time-stretch-low',
type=float,
default=0.8,
help='Low side of the stretch range')
parser.add_argument('--time-stretch-high',
type=float,
default=1.25,
help='High side of the stretch range')
# End of speech args
parser.add_argument(
'--langtok-merge-strategy',
default='concat',
type=str,
choices=['concat', 'sum'],
metavar='MRG',
help=
'strategy to use when merging the language token with the input, '
'it can be \'sum\' or \'concat\'')