def set_default_args(args):
# Prevents generate from printing individual translated sentences when
# calculating BLEU score.
args.quiet = True
# Set default init method for multi-GPU training if the user didn't specify
# them.
if args.distributed_world_size > 1:
args.distributed_init_method = (
f"tcp://localhost:{random.randint(10000, 20000)}" if
not args.distributed_init_method else args.distributed_init_method)
if args.local_num_gpus > args.distributed_world_size:
raise ValueError(
f"--local-num-gpus={args.local_num_gpus} must be "
f"<= --distributed-world-size={args.distributed_world_size}.")
if args.local_num_gpus > torch.cuda.device_count():
raise ValueError(
f"--local-num-gpus={args.local_num_gpus} must be "
f"<= the number of GPUs: {torch.cuda.device_count()}.")
if args.fp16 and getattr(args, "adversary", False):
print(
"Warning: disabling fp16 training since it's not supported by AdversarialTrainer."
)
args.fp16 = False
if not args.source_vocab_file:
args.source_vocab_file = pytorch_translate_dictionary.default_dictionary_path(
save_dir=args.save_dir, dialect=args.source_lang)
if not args.target_vocab_file:
args.target_vocab_file = pytorch_translate_dictionary.default_dictionary_path(
save_dir=args.save_dir, dialect=args.target_lang)
if args.arch in constants.ARCHS_FOR_CHAR_SOURCE and not args.char_source_vocab_file:
args.char_source_vocab_file = (
pytorch_translate_dictionary.default_char_dictionary_path(
save_dir=args.save_dir, dialect=args.source_lang))
if args.arch in constants.ARCHS_FOR_CHAR_TARGET and not args.char_target_vocab_file:
args.char_target_vocab_file = (
pytorch_translate_dictionary.default_char_dictionary_path(
save_dir=args.save_dir, dialect=args.target_lang))
if args.multiling_encoder_lang and not args.multiling_source_vocab_file:
args.multiling_source_vocab_file = [
pytorch_translate_dictionary.default_dictionary_path(
save_dir=args.save_dir, dialect=f"src-{l}")
for l in args.multiling_encoder_lang
]
if args.multiling_decoder_lang and not args.multiling_target_vocab_file:
args.multiling_target_vocab_file = [
pytorch_translate_dictionary.default_dictionary_path(
save_dir=args.save_dir, dialect=f"trg-{l}")
for l in args.multiling_decoder_lang
]
def get_dict_paths(vocabulary_args: Optional[List[str]], langs: List[str],
save_dir: str) -> Dict[str, str]:
"""
Extract dictionary files based on --vocabulary argument, for the given
languages `langs`.
vocabulary_arg: Optional[List[str]] where each element is a str with the format
"lang:vocab_file"
"""
dicts = {}
if vocabulary_args is not None:
for vocab_config in vocabulary_args:
# vocab_config is in the format "lang:vocab_file"
lang, vocab = vocab_config.split(":")
if lang in langs:
dicts[lang] = vocab
for lang in langs:
if lang not in dicts:
dicts[lang] = pytorch_translate_dictionary.default_dictionary_path(
save_dir=save_dir, dialect=lang)
return dicts
