def train_translation_model(data_dir, extra_flags):
parser = train.get_parser_with_args()
args = options.parse_args_and_arch(parser, [
"--save-dir",
data_dir,
"--train-source-text-file",
os.path.join(data_dir, "train.in"),
"--train-target-text-file",
os.path.join(data_dir, "train.out"),
"--eval-source-text-file",
os.path.join(data_dir, "valid.in"),
"--eval-target-text-file",
os.path.join(data_dir, "valid.out"),
"--source-max-vocab-size",
"26",
"--target-max-vocab-size",
"26",
"--max-tokens",
"500",
"--optimizer",
"sgd",
"--lr",
"0.05",
"--lr-scheduler",
"fixed",
"--lr-shrink",
"0.95",
"--momentum",
"0.0",
"--clip-norm",
"5.0",
"--sentence-avg",
"--criterion",
"label_smoothed_cross_entropy",
"--label-smoothing",
"0.1",
"--beam",
"3",
"--stop-no-best-bleu-eval",
"5",
"--unk-reward",
"0.5",
"--generate-bleu-eval-avg-checkpoints",
"10",
"--generate-bleu-eval-per-epoch",
"--max-epoch",
"1",
"--stop-time-hr",
"1",
"--no-progress-bar",
"--distributed-world-size",
"1",
"--source-lang",
"in",
"--target-lang",
"out",
] + (extra_flags or []))
train.validate_and_set_default_args(args)
train.main(args, train.single_process_main)
trainer=trainer,
task=task,
epoch_itr=epoch_itr,
)
for num_updates, stats in train_iterator:
if distributed_utils.is_master(args):
for k, v in LOGGED_STATS_MAP.items():
if stats[k] is not None:
if v == "mid_epoch":
tensorboard_logger.log_value(k, float(stats[k]),
num_updates)
else:
tensorboard_logger.log_value(k, float(stats[k]),
extra_state["epoch"])
if __name__ == "__main__":
parser = train.get_parser_with_args()
parser.add_argument(
"--tensorboard-dir",
type=str,
default="run-1234",
help=
"The log directory for tensorboard, used as `tensorboard --logdir DIR`",
)
args = options.parse_args_and_arch(parser)
train.validate_and_set_default_args(args)
print(args)
train.main(args, single_process_tensorboard)
def train_translation_model(
data_dir,
extra_flags,
criterion=None,
set_empty_data_positional_arg=False,
set_lang_args=True,
):
parser = train.get_parser_with_args()
args = options.parse_args_and_arch(
parser,
([""] if set_empty_data_positional_arg else [])
+ [
"--save-dir",
data_dir,
"--train-source-text-file",
os.path.join(data_dir, "train.in"),
"--train-target-text-file",
os.path.join(data_dir, "train.out"),
"--eval-source-text-file",
os.path.join(data_dir, "valid.in"),
"--eval-target-text-file",
os.path.join(data_dir, "valid.out"),
"--source-max-vocab-size",
"26",
"--target-max-vocab-size",
"26",
"--max-tokens",
"500",
"--optimizer",
"sgd",
"--lr",
"0.05",
"--lr-scheduler",
"fixed",
"--lr-shrink",
"0.95",
"--momentum",
"0.0",
"--clip-norm",
"5.0",
"--sentence-avg",
"--beam",
"3",
"--stop-no-best-bleu-eval",
"5",
"--unk-reward",
"0.5",
"--num-avg-checkpoints",
"10",
"--max-epoch",
"1",
"--stop-time-hr",
"1",
"--no-progress-bar",
"--distributed-world-size",
"1",
"--local-num-gpus",
"1" if torch.cuda.device_count() >= 1 else "0",
]
+ (["--source-lang", "in", "--target-lang", "out"] if set_lang_args else [])
+ (extra_flags or [])
+ (
criterion
or [
"--criterion",
"label_smoothed_cross_entropy",
"--label-smoothing",
"0.1",
]
),
)
train.validate_and_set_default_args(args)
train.main(args)