def get_parser_with_args():
parser = options.get_parser("Generation")
options.add_dataset_args(parser, gen=True)
options.add_generation_args(parser)
pytorch_translate_generate.add_args(parser)
group = parser.add_argument_group("Generation")
group.add_argument(
"--source-vocab-file",
default="",
metavar="FILE",
help="Path to text file representing the Dictionary to use.",
)
group.add_argument(
"--target-vocab-file",
default="",
metavar="FILE",
help="Path to text file representing the Dictionary to use.",
)
# Add args related to benchmarking.
group = parser.add_argument_group("Benchmarking")
group.add_argument(
"--increment",
default=5,
type=int,
help="Difference in lengths between synthesized sentences. "
"Must be integer >=1.",
)
group.add_argument(
"--max-length",
default=100,
type=int,
help="Maximum allowed length for synthesized sentences. "
"Should be greater than --increment.",
)
group.add_argument(
"--samples-per-length",
default=1,
type=int,
help="Number of sentences to be synthesized at each length. ",
)
return parser
def get_parser_with_args():
parser = options.get_parser("Trainer")
options.add_dataset_args(parser, train=True, gen=True)
options.add_distributed_training_args(parser)
options.add_optimization_args(parser)
options.add_checkpoint_args(parser)
options.add_model_args(parser)
options.add_generation_args(parser)
parser.add_argument(
"--log-verbose",
action="store_true",
help="Whether to output more verbose logs for debugging/profiling.",
)
# Adds args related to training (validation and stopping criterions).
group = parser.add_argument_group("Optimization")
group.add_argument(
"--subepoch-validate-interval",
default=0,
type=int,
metavar="N",
help="Calculates loss over the validation set every N batch updates. "
"Note that validation is done at the end of every epoch regardless. "
"A value of <= 0 disables this.",
)
group.add_argument(
"--stop-time-hr",
default=-1,
type=int,
metavar="N",
help="Stops training after N hours have elapsed. "
"A value of < 0 disables this.",
)
group.add_argument(
"--stop-no-best-validate-loss",
default=-1,
type=int,
metavar="N",
help="Stops training after N validations have been run without "
"achieving a better loss than before. Note that this is affected by "
"--validation-interval in how frequently we run validation in the "
"first place. A value of < 0 disables this.",
)
group.add_argument(
"--stop-no-best-bleu-eval",
default=-1,
type=int,
metavar="N",
help="Stops training after N evals have been run without "
"achieving a better BLEU score than before. Note that this is affected "
"by --generate-bleu-eval-interval in how frequently we run BLEU eval "
"in the first place. A value of < 0 disables this.",
)
# Adds args related to input data files (preprocessing, numberizing, and
# binarizing text files; creating vocab files)
preprocess.add_args(parser)
# Adds args related to checkpointing.
group = parser.add_argument_group("Checkpointing")
group.add_argument(
"--no-end-of-epoch-checkpoints",
action="store_true",
help="Disables saving checkpoints at the end of the epoch. "
"This differs from --no-save and --no-epoch-checkpoints in that it "
"still allows for intra-epoch checkpoints if --save-interval is set.",
)
group.add_argument(
"--max-checkpoints-kept",
default=-1,
type=int,
metavar="N",
help="Keep at most the last N checkpoints file around. "
"A value < -1 keeps all. "
"When --generate-bleu-eval-avg-checkpoints is used and is > N, the "
"number of checkpoints kept around is automatically adjusted "
"to allow BLEU to work properly.",
)
# Adds args for generating intermediate BLEU eval while training.
# generate.add_args() adds args used by both train.py and the standalone
# generate binary, while the flags defined here are used only by train.py.
generate.add_args(parser)
group = parser.add_argument_group("Generation")
group.add_argument(
"--generate-bleu-eval-per-epoch",
action="store_true",
help="Whether to generate BLEU score eval after each epoch.",
)
group.add_argument(
"--generate-bleu-eval-interval",
default=0,
type=int,
metavar="N",
help="Does BLEU eval every N batch updates. Note that "
"--save-interval also affects this - we can only eval as "
"frequently as a checkpoint is written. A value of <= 0 "
"disables this.",
)
group.add_argument(
"--generate-bleu-eval-avg-checkpoints",
default=1,
type=int,
metavar="N",
help="Maximum number of last N checkpoints to average over when "
"doing BLEU eval. Must be >= 1.",
)
group.add_argument(
"--continuous-averaging-after-epochs",
type=int,
default=-1,
help=("Average parameter values after each step since previous "
"checkpoint, beginning after the specified number of epochs. "),
)
return parser
def get_parser_with_args():
parser = options.get_parser('Trainer')
options.add_dataset_args(parser, train=True, gen=True)
options.add_distributed_training_args(parser)
options.add_optimization_args(parser)
options.add_checkpoint_args(parser)
options.add_model_args(parser)
options.add_generation_args(parser)
parser.add_argument(
'--log-verbose',
action='store_true',
help='Whether to output more verbose logs for debugging/profiling.',
)
# Adds args related to training (validation and stopping criterions).
group = parser.add_argument_group('Optimization')
group.add_argument(
'--subepoch-validate-interval',
default=0,
type=int,
metavar='N',
help='Calculates loss over the validation set every N batch updates. '
'Note that validation is done at the end of every epoch regardless. '
'A value of <= 0 disables this.',
)
group.add_argument(
'--stop-time-hr',
default=-1,
type=int,
metavar='N',
help='Stops training after N hours have elapsed. '
'A value of < 0 disables this.',
)
group.add_argument(
'--stop-no-best-validate-loss',
default=-1,
type=int,
metavar='N',
help='Stops training after N validations have been run without '
'achieving a better loss than before. Note that this is affected by '
'--validation-interval in how frequently we run validation in the '
'first place. A value of < 0 disables this.',
)
group.add_argument(
'--stop-no-best-bleu-eval',
default=-1,
type=int,
metavar='N',
help='Stops training after N evals have been run without '
'achieving a better BLEU score than before. Note that this is affected '
'by --generate-bleu-eval-interval in how frequently we run BLEU eval '
'in the first place. A value of < 0 disables this.',
)
# Args related to dataset.
group = parser.add_argument_group('Dataset and data loading')
group.add_argument(
'--source-vocab-file',
default='',
metavar='FILE',
help='Path to text file representing the fairseq Dictionary to use. '
'If left empty, the dict is auto-generated from source training data.',
)
group.add_argument(
'--source-max-vocab-size',
default=-1,
type=int,
metavar='N',
help='If a new vocab file needs to be generated, restrict it to the '
'top N most common words. If we re-use an existing vocab file, this '
'flag will have no effect. A value of < 0 means no max size.',
)
group.add_argument(
'--target-vocab-file',
default='',
metavar='FILE',
help='Path to text file representing the fairseq Dictionary to use. '
'If left empty, the dict is auto-generated from target training data.',
)
group.add_argument(
'--target-max-vocab-size',
default=-1,
type=int,
metavar='N',
help='If a new vocab file needs to be generated, restrict it to the '
'top N most common words. If we re-use an existing vocab file, this '
'flag will have no effect. A value of < 0 means no max size.',
)
group.add_argument(
'--train-source-text-file',
default='',
metavar='FILE',
help='Path to raw text file containing source training examples. '
'This overrides what would be loaded from the data dir.',
)
group.add_argument(
'--train-target-text-file',
default='',
metavar='FILE',
help='Path to raw text file containing target training examples. '
'This overrides what would be loaded from the data dir.',
)
group.add_argument(
'--eval-source-text-file',
default='',
metavar='FILE',
help='Path to raw text file containing source eval examples for '
'calculating validation loss and BLEU eval scores. '
'This overrides what would be loaded from the data dir.',
)
group.add_argument(
'--eval-target-text-file',
default='',
metavar='FILE',
help='Path to raw text file containing target eval examples for '
'calculating validation loss and BLEU eval scores. '
'This overrides what would be loaded from the data dir.',
)
# Adds args related to checkpointing.
group = parser.add_argument_group('Checkpointing')
group.add_argument(
'--no-end-of-epoch-checkpoints',
action='store_true',
help='Disables saving checkpoints at the end of the epoch. '
'This differs from --no-save and --no-epoch-checkpoints in that it '
'still allows for intra-epoch checkpoints if --save-interval is set.')
# Adds args for generating intermediate BLEU eval while training.
# generate.add_args() adds args used by both train.py and the standalone
# generate binary, while the flags defined here are used only by train.py.
generate.add_args(parser)
group = parser.add_argument_group('Generation')
group.add_argument(
'--generate-bleu-eval-per-epoch',
action='store_true',
help='Whether to generate BLEU score eval after each epoch.',
)
group.add_argument(
'--generate-bleu-eval-interval',
default=0,
type=int,
metavar='N',
help='Does BLEU eval every N batch updates. Note that '
'--save-interval also affects this - we can only eval as '
'frequently as a checkpoint is written. A value of <= 0 '
'disables this.',
)
group.add_argument(
'--generate-bleu-eval-avg-checkpoints',
default=1,
type=int,
metavar='N',
help='Maximum number of last N checkpoints to average over when '
'doing BLEU eval. Must be >= 1.',
)
group.add_argument(
'--continuous-averaging-after-epochs',
type=int,
default=-1,
help=('Average parameter values after each step since previous '
'checkpoint, beginning after the specified number of epochs. '),
)
return parser
def get_parser_with_args():
parser = options.get_parser("Trainer")
options.add_dataset_args(parser, train=True, gen=True)
options.add_distributed_training_args(parser)
options.add_optimization_args(parser)
options.add_checkpoint_args(parser)
options.add_model_args(parser)
options.add_generation_args(parser)
parser.add_argument(
"--log-verbose",
action="store_true",
help="Whether to output more verbose logs for debugging/profiling.",
)
# Adds args related to training (validation and stopping criterions).
group = parser.add_argument_group("Optimization")
group.add_argument(
"--subepoch-validate-interval",
default=0,
type=int,
metavar="N",
help="Calculates loss over the validation set every N batch updates. "
"Note that validation is done at the end of every epoch regardless. "
"A value of <= 0 disables this.",
)
group.add_argument(
"--stop-time-hr",
default=-1,
type=int,
metavar="N",
help="Stops training after N hours have elapsed. "
"A value of < 0 disables this.",
)
group.add_argument(
"--stop-no-best-validate-loss",
default=-1,
type=int,
metavar="N",
help="Stops training after N validations have been run without "
"achieving a better loss than before. Note that this is affected by "
"--validation-interval in how frequently we run validation in the "
"first place. A value of < 0 disables this.",
)
group.add_argument(
"--stop-no-best-bleu-eval",
default=-1,
type=int,
metavar="N",
help="Stops training after N evals have been run without "
"achieving a better BLEU score than before. Note that this is affected "
"by --generate-bleu-eval-interval in how frequently we run BLEU eval "
"in the first place. A value of < 0 disables this.",
)
# Args related to dataset.
group = parser.add_argument_group("Dataset and data loading")
group.add_argument(
"--source-vocab-file",
default="",
metavar="FILE",
help="Path to text file representing the fairseq Dictionary to use. "
"If left empty, the dict is auto-generated from source training data.",
)
group.add_argument(
"--source-max-vocab-size",
default=-1,
type=int,
metavar="N",
help="If a new vocab file needs to be generated, restrict it to the "
"top N most common words. If we re-use an existing vocab file, this "
"flag will have no effect. A value of < 0 means no max size.",
)
group.add_argument(
"--target-vocab-file",
default="",
metavar="FILE",
help="Path to text file representing the fairseq Dictionary to use. "
"If left empty, the dict is auto-generated from target training data.",
)
group.add_argument(
"--target-max-vocab-size",
default=-1,
type=int,
metavar="N",
help="If a new vocab file needs to be generated, restrict it to the "
"top N most common words. If we re-use an existing vocab file, this "
"flag will have no effect. A value of < 0 means no max size.",
)
group.add_argument(
"--train-source-text-file",
default="",
metavar="FILE",
help="Path to raw text file containing source training examples. "
"This overrides what would be loaded from the data dir.",
)
group.add_argument(
"--train-target-text-file",
default="",
metavar="FILE",
help="Path to raw text file containing target training examples. "
"This overrides what would be loaded from the data dir.",
)
group.add_argument(
"--eval-source-text-file",
default="",
metavar="FILE",
help="Path to raw text file containing source eval examples for "
"calculating validation loss and BLEU eval scores. "
"This overrides what would be loaded from the data dir.",
)
group.add_argument(
"--eval-target-text-file",
default="",
metavar="FILE",
help="Path to raw text file containing target eval examples for "
"calculating validation loss and BLEU eval scores. "
"This overrides what would be loaded from the data dir.",
)
group.add_argument(
"--penalized-target-tokens-file",
default="",
metavar="FILE",
help="Path to text file of tokens to receive a penalty in decoding."
"If left empty, no penalty will be applied",
)
# Adds args related to checkpointing.
group = parser.add_argument_group("Checkpointing")
group.add_argument(
"--no-end-of-epoch-checkpoints",
action="store_true",
help="Disables saving checkpoints at the end of the epoch. "
"This differs from --no-save and --no-epoch-checkpoints in that it "
"still allows for intra-epoch checkpoints if --save-interval is set.",
)
group.add_argument(
"--max-checkpoints-kept",
default=-1,
type=int,
metavar="N",
help="Keep at most the last N checkpoints file around. "
"A value < -1 keeps all. "
"When --generate-bleu-eval-avg-checkpoints is used and is > N, the "
"number of checkpoints kept around is automatically adjusted "
"to allow BLEU to work properly.",
)
# Adds args for generating intermediate BLEU eval while training.
# generate.add_args() adds args used by both train.py and the standalone
# generate binary, while the flags defined here are used only by train.py.
generate.add_args(parser)
group = parser.add_argument_group("Generation")
group.add_argument(
"--generate-bleu-eval-per-epoch",
action="store_true",
help="Whether to generate BLEU score eval after each epoch.",
)
group.add_argument(
"--generate-bleu-eval-interval",
default=0,
type=int,
metavar="N",
help="Does BLEU eval every N batch updates. Note that "
"--save-interval also affects this - we can only eval as "
"frequently as a checkpoint is written. A value of <= 0 "
"disables this.",
)
group.add_argument(
"--generate-bleu-eval-avg-checkpoints",
default=1,
type=int,
metavar="N",
help="Maximum number of last N checkpoints to average over when "
"doing BLEU eval. Must be >= 1.",
)
group.add_argument(
"--continuous-averaging-after-epochs",
type=int,
default=-1,
help=("Average parameter values after each step since previous "
"checkpoint, beginning after the specified number of epochs. "),
)
return parser