def cli_main():
tmp_parser = options.get_parser('Preprocessing', 'translation')
tmp_args, _ = tmp_parser.parse_known_args()
parser = options.get_parser('Preprocessing', 'translation')
tasks.get_task(tmp_args.task).add_args(parser)
options.add_preprocess_args(parser)
args = parser.parse_args()
main(args)
def get_parser_with_args():
parser = options.get_parser("Collect Top-K Probs",
default_task="pytorch_translate")
pytorch_translate_options.add_verbosity_args(parser)
pytorch_translate_options.add_dataset_args(parser, gen=True)
generation_group = options.add_generation_args(parser)
generation_group.add_argument(
"--source-binary-file",
default="",
help="Path for the binary file containing source eval examples. "
"(Overrides --source-text-file. Must be used in conjunction with "
"--target-binary-file).",
)
generation_group.add_argument(
"--target-binary-file",
default="",
help="Path for the binary file containing target eval examples. "
"(Overrides --target-text-file. Must be used in conjunction with "
"--source-binary-file).",
)
generation_group.add_argument(
"--k-probs-to-collect",
type=int,
default=8,
help="Number of probabilities to collect for each output step.",
)
generation_group.add_argument(
"--top-k-probs-binary-file",
type=str,
default="",
help="File into which to save top-K probabilities for each token.",
)
return parser
def get_parser_with_args():
parser = options.get_parser("Trainer")
parser.add_argument(
"--log-verbose",
action="store_true",
help="Whether to output more verbose logs for debugging/profiling.",
)
pytorch_translate_options.add_dataset_args(parser, train=True, gen=True)
options.add_distributed_training_args(parser)
# Adds args related to training (validation and stopping criterions).
optimization_group = options.add_optimization_args(parser)
pytorch_translate_options.expand_optimization_args(optimization_group)
# Adds args related to checkpointing.
checkointing_group = options.add_checkpoint_args(parser)
pytorch_translate_options.expand_checkpointing_args(checkointing_group)
# Add model related args
options.add_model_args(parser)
# Adds args for generating intermediate BLEU eval while training.
generation_group = options.add_generation_args(parser)
pytorch_translate_options.expand_generation_args(generation_group,
train=True)
# Adds args related to input data files (preprocessing, numberizing, and
# binarizing text files; creating vocab files)
pytorch_translate_options.add_preprocessing_args(parser)
return parser
def get_parser_with_args():
parser = options.get_parser('Generation')
options.add_dataset_args(parser, gen=True)
options.add_generation_args(parser)
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.')
group.add_argument(
'--source-text-file',
default='',
metavar='FILE',
help='Path to raw text file containing examples in source dialect. '
'This overrides what would be loaded from the data dir.',
)
group.add_argument(
'--target-text-file',
default='',
metavar='FILE',
help='Path to raw text file containing examples in target dialect. '
'This overrides what would be loaded from the data dir.',
)
return parser
def get_parser_with_args():
parser = options.get_parser("Generation")
options.add_dataset_args(parser, gen=True)
options.add_generation_args(parser)
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.",
)
group.add_argument(
"--source-text-file",
default="",
metavar="FILE",
help="Path to raw text file containing examples in source dialect. "
"This overrides what would be loaded from the data dir.",
)
group.add_argument(
"--target-text-file",
default="",
metavar="FILE",
help="Path to raw text file containing examples in target dialect. "
"This overrides what would be loaded from the data dir.",
)
return parser
def main():
parser = options.get_parser('Generation')
parser.add_argument('--path', metavar='FILE', required=True, action='append',
help='path(s) to model file(s)')
options.add_dataset_args(parser)
options.add_generation_args(parser)
args = parser.parse_args()
print(args)
use_cuda = torch.cuda.is_available() and not args.cpu
# Load ensemble
print('| loading model(s) from {}'.format(', '.join(args.path)))
models, model_args = utils.load_ensemble_for_inference(args.path, data_dir=args.data)
src_dict, dst_dict = models[0].src_dict, models[0].dst_dict
print('| [{}] dictionary: {} types'.format(model_args.source_lang, len(src_dict)))
print('| [{}] dictionary: {} types'.format(model_args.target_lang, len(dst_dict)))
# Optimize ensemble for generation
for model in models:
model.make_generation_fast_(
beamable_mm_beam_size=None if args.no_beamable_mm else args.beam)
# Initialize generator
translator = SequenceGenerator(
models, beam_size=args.beam, stop_early=(not args.no_early_stop),
normalize_scores=(not args.unnormalized), len_penalty=args.lenpen,
unk_penalty=args.unkpen)
if use_cuda:
translator.cuda()
# Load alignment dictionary for unknown word replacement
# (None if no unknown word replacement, empty if no path to align dictionary)
align_dict = utils.load_align_dict(args.replace_unk)
print('| Type the input sentence and press return:')
for src_str in sys.stdin:
src_str = src_str.strip()
src_tokens = tokenizer.Tokenizer.tokenize(src_str, src_dict, add_if_not_exist=False).long()
if use_cuda:
src_tokens = src_tokens.cuda()
translations = translator.generate(Variable(src_tokens.view(1, -1)))
hypos = translations[0]
print('O\t{}'.format(src_str))
# Process top predictions
for hypo in hypos[:min(len(hypos), args.nbest)]:
hypo_tokens, hypo_str, alignment = utils.post_process_prediction(
hypo_tokens=hypo['tokens'].int().cpu(),
src_str=src_str,
alignment=hypo['alignment'].int().cpu(),
align_dict=align_dict,
dst_dict=dst_dict,
remove_bpe=args.remove_bpe)
print('H\t{}\t{}'.format(hypo['score'], hypo_str))
print('A\t{}'.format(' '.join(map(str, alignment))))
def get_generation_parser(interactive=False, default_task='translation'):
parser = get_parser('Generation', default_task)
add_dataset_args(parser, gen=True)
add_distributed_training_args(parser, default_world_size=1)
add_generation_args(parser)
add_checkpoint_args(parser)
if interactive:
add_interactive_args(parser)
return parser
def get_training_and_generation_parser(default_task='translation'):
parser = options.get_parser('Trainer', default_task)
options.add_dataset_args(parser, train=True, gen=True)
options.add_generation_args(parser)
options.add_distributed_training_args(parser)
options.add_model_args(parser)
options.add_optimization_args(parser)
options.add_checkpoint_args(parser)
return parser
def get_parser_with_args():
"""Create argument parser with arguments specific to this script"""
parser = options.get_parser("Whitebox attack",
default_task="pytorch_translate_adversarial")
# Data related arguments
data_group = pytorch_translate_options.add_dataset_args(parser, gen=True)
# Adds args used by the standalone generate binary.
data_group.add_argument(
"--source-vocab-file",
default="",
metavar="FILE",
help="Path to text file representing the Dictionary to use.",
)
data_group.add_argument(
"--char-source-vocab-file",
default="",
metavar="FILE",
help=("Same as --source-vocab-file except using characters. "
"(For use with char_source models only.)"),
)
data_group.add_argument(
"--target-vocab-file",
default="",
metavar="FILE",
help="Path to text file representing the Dictionary to use.",
)
data_group.add_argument(
"--source-text-file",
default="",
metavar="FILE",
help="Path to raw text file containing examples in source dialect. "
"This overrides what would be loaded from the data dir. ",
)
data_group.add_argument(
"--target-text-file",
default="",
metavar="FILE",
help="Path to raw text file containing examples in target dialect. "
"This overrides what would be loaded from the data dir.",
)
data_group.add_argument(
"--adversarial-output-file",
default="",
type=str,
metavar="FILE",
help="Path to text file to store the generated adversarial examples.",
)
# Adversarial attack specific group
adversarial_options.add_adversarial_args(parser, attack_only=True)
return parser
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_preprocessing_parser():
parser = get_parser("Preprocessing", "translation")
group = parser.add_argument_group("Preprocessing")
# fmt: off
parser.add_argument('--format', metavar='INP',
help='Input format for audio files')
group.add_argument("--trainpref", metavar="FP", default=None,
help="train file prefix")
group.add_argument("--validpref", metavar="FP", default=None,
help="comma separated, valid file prefixes")
group.add_argument("--testpref", metavar="FP", default=None,
help="comma separated, test file prefixes")
group.add_argument("--destdir", metavar="DIR", default="data-bin",
help="destination dir")
parser.add_argument('--legacy-audio-fix-lua-indexing', action='store_true', default=False,
help='if set, the input filterbanks are added 1 for compatibility with lua indexing fix')
# TODO: add parallel implementation
# fmt: on
return parser
def get_parser_with_args(default_task="pytorch_translate"):
parser = options.get_parser("Trainer", default_task=default_task)
pytorch_translate_options.add_verbosity_args(parser, train=True)
pytorch_translate_options.add_dataset_args(parser, train=True, gen=True)
options.add_distributed_training_args(parser)
# Adds args related to training (validation and stopping criterions).
optimization_group = options.add_optimization_args(parser)
pytorch_translate_options.expand_optimization_args(optimization_group)
# Adds args related to checkpointing.
checkpointing_group = options.add_checkpoint_args(parser)
pytorch_translate_options.expand_checkpointing_args(checkpointing_group)
# Add model related args
options.add_model_args(parser)
# Adds args for generating intermediate BLEU eval while training.
generation_group = options.add_generation_args(parser)
pytorch_translate_options.expand_generation_args(generation_group, train=True)
# Adds args related to input data files (preprocessing, numberizing, and
# binarizing text files; creating vocab files)
pytorch_translate_options.add_preprocessing_args(parser)
return parser
def get_rendering_parser(default_task="single_object_rendering"):
parser = options.get_parser("Rendering", default_task)
options.add_dataset_args(parser, gen=True)
add_rendering_args(parser)
return parser
for arg in vars(parsed_args).keys():
setattr(args, arg, getattr(parsed_args, arg))
task = tasks.setup_task(args)
# Load dataset splits
task.load_dataset(args.gen_subset)
print('| {} {} {} examples'.format(args.data, args.gen_subset,
len(task.dataset(args.gen_subset))))
eval_dataset(task, model, task.dataset(args.gen_subset), args.out_file,
args.thresholds, args.compute_metrics, use_cuda)
if __name__ == '__main__':
parser = options.get_parser('Evaluate Single Sentence Classifier',
'sentence_classification')
options.add_common_eval_args(parser)
options.add_dataset_args(parser, gen=True)
parser.add_argument('--out-file', type=str, help='output filename')
parser.add_argument('--thresholds',
nargs='+',
type=float,
help='thresholds to try or use')
parser.add_argument(
'--compute-metrics',
action='store_true',
help='if set, uses the labels to compute metrics for each threshold')
args = options.parse_args_and_arch(parser)
main(args)
def main():
parser = options.get_parser('Trainer')
dataset_args = options.add_dataset_args(parser)
dataset_args.add_argument('--max-tokens',
default=6000,
type=int,
metavar='N',
help='maximum number of tokens in a batch')
dataset_args.add_argument(
'--train-subset',
default='train',
metavar='SPLIT',
choices=['train', 'valid', 'test'],
help='data subset to use for training (train, valid, test)')
dataset_args.add_argument(
'--valid-subset',
default='valid',
metavar='SPLIT',
help='comma separated list ofdata subsets '
' to use for validation (train, valid, valid1,test, test1)')
options.add_optimization_args(parser)
options.add_checkpoint_args(parser)
options.add_model_args(parser)
args = utils.parse_args_and_arch(parser)
print(args)
if args.no_progress_bar:
progress_bar.enabled = False
progress_bar.print_interval = args.log_interval
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
torch.manual_seed(args.seed)
# Load dataset
dataset = data.load_with_check(args.data, ['train', 'valid'],
args.source_lang, args.target_lang)
if args.source_lang is None or args.target_lang is None:
# record inferred languages in args, so that it's saved in checkpoints
args.source_lang, args.target_lang = dataset.src, dataset.dst
print('| [{}] dictionary: {} types'.format(dataset.src,
len(dataset.src_dict)))
print('| [{}] dictionary: {} types'.format(dataset.dst,
len(dataset.dst_dict)))
for split in ['train', 'valid']:
print('| {} {} {} examples'.format(args.data, split,
len(dataset.splits[split])))
if not torch.cuda.is_available():
raise NotImplementedError('Training on CPU is not supported')
num_gpus = torch.cuda.device_count()
print('| using {} GPUs (with max tokens per GPU = {})'.format(
num_gpus, args.max_tokens))
# Build model and criterion
model = utils.build_model(args, dataset.src_dict, dataset.dst_dict)
criterion = utils.build_criterion(args, dataset.src_dict, dataset.dst_dict)
print('| model {}, criterion {}'.format(args.arch,
criterion.__class__.__name__))
# Start multiprocessing
trainer = MultiprocessingTrainer(args, model, criterion)
# Load the latest checkpoint if one is available
checkpoint_path = os.path.join(args.save_dir, args.restore_file)
extra_state = trainer.load_checkpoint(checkpoint_path)
if extra_state is not None:
epoch = extra_state['epoch']
batch_offset = extra_state['batch_offset']
print('| loaded checkpoint {} (epoch {})'.format(
checkpoint_path, epoch))
if batch_offset == 0:
epoch += 1
else:
epoch, batch_offset = 1, 0
# Train until the learning rate gets too small
val_loss = None
max_epoch = args.max_epoch or math.inf
lr = trainer.get_lr()
train_meter = StopwatchMeter()
train_meter.start()
while lr > args.min_lr and epoch <= max_epoch:
# train for one epoch
train(args, epoch, batch_offset, trainer, dataset, num_gpus)
# evaluate on validate set
for k, subset in enumerate(args.valid_subset.split(',')):
val_loss = validate(args, epoch, trainer, dataset, subset,
num_gpus)
if k == 0:
if not args.no_save:
# save checkpoint
save_checkpoint(trainer, args, epoch, 0, val_loss)
# only use first validation loss to update the learning schedule
lr = trainer.lr_step(val_loss, epoch)
epoch += 1
batch_offset = 0
train_meter.stop()
print('| done training in {:.1f} seconds'.format(train_meter.sum))
# Stop multiprocessing
trainer.stop()
def get_parser_with_args():
parser = options.get_parser("Generation", default_task="pytorch_translate")
pytorch_translate_options.add_verbosity_args(parser)
pytorch_translate_options.add_dataset_args(parser, gen=True)
generation_group = options.add_generation_args(parser)
pytorch_translate_options.expand_generation_args(generation_group)
generation_group.add_argument(
"--source-vocab-file",
default="",
metavar="FILE",
help="Path to text file representing the Dictionary to use.",
)
generation_group.add_argument(
"--char-source-vocab-file",
default="",
metavar="FILE",
help=(
"Same as --source-vocab-file except using characters. "
"(For use with char_source models only.)"
),
)
generation_group.add_argument(
"--target-vocab-file",
default="",
metavar="FILE",
help="Path to text file representing the Dictionary to use.",
)
generation_group.add_argument(
"--multiling-source-lang",
action="append",
metavar="SRC",
help=(
"Must be set for decoding with multilingual models. "
"Must match an entry from --multiling-encoder-lang from training."
),
)
generation_group.add_argument(
"--multiling-target-lang",
action="append",
metavar="TARGET",
help=(
"Must be set for decoding with multilingual models. "
"Must match an entry from --multiling-decoder-lang from training."
),
)
# Add args related to benchmarking.
group = parser.add_argument_group("Benchmarking")
group.add_argument(
"--runs-per-length",
default=10,
type=int,
help="Number of times to run generation on each length.",
)
group.add_argument(
"--examples-per-length",
default=1,
type=int,
help="Sentences of each length to include in each eval (batched if >1).",
)
return parser
def get_tuning_parser(default_task='translation'):
parser = options.get_parser('Reranking tuning', default_task)
add_reranking_args(parser)
add_tuning_args(parser)
return parser
def get_reranking_parser(default_task='translation'):
parser = options.get_parser('Generation and reranking', default_task)
add_reranking_args(parser)
return parser
def main():
parser = options.get_parser('Generation')
parser.add_argument('--path',
metavar='FILE',
required=True,
action='append',
help='path(s) to model file(s)')
dataset_args = options.add_dataset_args(parser)
dataset_args.add_argument('--batch-size',
default=32,
type=int,
metavar='N',
help='batch size')
dataset_args.add_argument(
'--gen-subset',
default='test',
metavar='SPLIT',
help='data subset to generate (train, valid, test)')
dataset_args.add_argument('--num-shards',
default=1,
type=int,
metavar='N',
help='shard generation over N shards')
dataset_args.add_argument(
'--shard-id',
default=0,
type=int,
metavar='ID',
help='id of the shard to generate (id < num_shards)')
options.add_generation_args(parser)
args = parser.parse_args()
if args.no_progress_bar and args.log_format is None:
args.log_format = 'none'
# print(args)
use_cuda = torch.cuda.is_available() and not args.cpu
if hasattr(torch, 'set_grad_enabled'):
torch.set_grad_enabled(False)
# Load dataset
if args.replace_unk is None:
dataset = data.load_dataset(args.data, [args.gen_subset],
args.source_lang, args.target_lang)
else:
dataset = data.load_raw_text_dataset(args.data, [args.gen_subset],
args.source_lang,
args.target_lang)
if args.source_lang is None or args.target_lang is None:
# record inferred languages in args
args.source_lang, args.target_lang = dataset.src, dataset.dst
# Load ensemble
# print('| loading model(s) from {}'.format(', '.join(args.path)))
models, _ = utils.load_ensemble_for_inference(args.path, dataset.src_dict,
dataset.dst_dict)
# print('| [{}] dictionary: {} types'.format(dataset.src, len(dataset.src_dict)))
# print('| [{}] dictionary: {} types'.format(dataset.dst, len(dataset.dst_dict)))
# print('| {} {} {} examples'.format(args.data, args.gen_subset, len(dataset.splits[args.gen_subset])))
# Optimize ensemble for generation
for model in models:
model.make_generation_fast_(
beamable_mm_beam_size=None if args.no_beamable_mm else args.beam)
# Initialize generator
translator = SequenceGenerator(models,
beam_size=args.beam,
stop_early=(not args.no_early_stop),
normalize_scores=(not args.unnormalized),
len_penalty=args.lenpen,
unk_penalty=args.unkpen)
if use_cuda:
translator.cuda()
# Load alignment dictionary for unknown word replacement
# (None if no unknown word replacement, empty if no path to align dictionary)
align_dict = utils.load_align_dict(args.replace_unk)
# Generate and compute BLEU score
#scorer = bleu.Scorer(dataset.dst_dict.pad(), dataset.dst_dict.eos(), dataset.dst_dict.unk())
max_positions = min(model.max_encoder_positions() for model in models)
itr = dataset.eval_dataloader(args.gen_subset,
max_sentences=args.batch_size,
max_positions=max_positions,
skip_invalid_size_inputs_valid_test=args.
skip_invalid_size_inputs_valid_test)
if args.num_shards > 1:
if args.shard_id < 0 or args.shard_id >= args.num_shards:
raise ValueError('--shard-id must be between 0 and num_shards')
itr = data.sharded_iterator(itr, args.num_shards, args.shard_id)
num_sentences = 0
with utils.build_progress_bar(args, itr) as t:
wps_meter = TimeMeter()
gen_timer = StopwatchMeter()
translations = translator.generate_batched_itr(
t,
maxlen_a=args.max_len_a,
maxlen_b=args.max_len_b,
cuda_device=0 if use_cuda else None,
timer=gen_timer)
correct = 0
total = 0
for sample_id, src_tokens, target_tokens, hypos in translations:
# Process input and ground truth
target_tokens = target_tokens.int().cpu()
# Either retrieve the original sentences or regenerate them from tokens.
if align_dict is not None:
src_str = dataset.splits[
args.gen_subset].src.get_original_text(sample_id)
target_str = dataset.splits[
args.gen_subset].dst.get_original_text(sample_id)
else:
src_str = dataset.src_dict.string(src_tokens, args.remove_bpe)
target_str = dataset.dst_dict.string(target_tokens,
args.remove_bpe,
escape_unk=True)
# if not args.quiet:
# print('S-{}\t{}'.format(sample_id, src_str))
# print('T-{}\t{}'.format(sample_id, target_str))
total += 1
# Process top predictions
for i, hypo in enumerate(hypos[:min(len(hypos), args.nbest)]):
hypo_tokens, hypo_str, alignment = utils.post_process_prediction(
hypo_tokens=hypo['tokens'].int().cpu(),
src_str=src_str,
alignment=hypo['alignment'].int().cpu(),
align_dict=align_dict,
dst_dict=dataset.dst_dict,
remove_bpe=args.remove_bpe)
#if src_str == 'walk around right thrice after jump opposite left twice':
# import pdb; pdb.set_trace()
# if not args.quiet:
# print('H-{}\t{}\t{}'.format(sample_id, hypo['score'], hypo_str))
# print('A-{}\t{}'.format(sample_id, ' '.join(map(str, alignment))))
# Score only the top hypothesis
if i == 0:
if align_dict is not None or args.remove_bpe is not None:
# Convert back to tokens for evaluation with unk replacement and/or without BPE
target_tokens = tokenizer.Tokenizer.tokenize(
target_str,
dataset.dst_dict,
add_if_not_exist=True)
#scorer.add(target_tokens, hypo_tokens)
mat = ''
for row in hypo['attention']:
for column in row:
mat += str(column) + '\t'
mat += '\n'
tar = '/' + target_str
tra = '=' + str(target_str == hypo_str)
to_write.write(mat)
to_write.write(src_str)
to_write.write('\n')
to_write.write(hypo_str)
to_write.write('\n')
to_write.write(tar)
to_write.write('\n')
to_write.write(tra)
to_write.write('\n')
to_write.write('-----------')
to_write.write('\n')
if hypo_str == target_str:
correct += 1
wps_meter.update(src_tokens.size(0))
t.log({'wps': round(wps_meter.avg)})
num_sentences += 1
print('| Correct : {} - Total: {}. Accuracy: {:.5f}'.format(
correct, total, correct / total))
def get_tuning_parser(default_task="translation"):
parser = options.get_parser("Reranking tuning", default_task)
add_reranking_args(parser)
add_tuning_args(parser)
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.",
)
# 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 main():
parser = options.get_parser('Generation')
parser.add_argument('--path',
metavar='FILE',
required=True,
action='append',
help='path(s) to model file(s)')
options.add_dataset_args(parser)
options.add_generation_args(parser)
args = parser.parse_args()
print(args)
use_cuda = torch.cuda.is_available() and not args.cpu
# Load ensemble
print('| loading model(s) from {}'.format(', '.join(args.path)))
models, model_args = utils.load_ensemble_for_inference(args.path,
data_dir=args.data)
src_dict, dst_dict = models[0].src_dict, models[0].dst_dict
print('| [{}] dictionary: {} types'.format(model_args.source_lang,
len(src_dict)))
print('| [{}] dictionary: {} types'.format(model_args.target_lang,
len(dst_dict)))
# Optimize ensemble for generation
for model in models:
model.make_generation_fast_(
beamable_mm_beam_size=None if args.no_beamable_mm else args.beam)
# Initialize generator
translator = SequenceGenerator(models,
beam_size=args.beam,
stop_early=(not args.no_early_stop),
normalize_scores=(not args.unnormalized),
len_penalty=args.lenpen,
unk_penalty=args.unkpen)
if use_cuda:
translator.cuda()
# Load alignment dictionary for unknown word replacement
# (None if no unknown word replacement, empty if no path to align dictionary)
align_dict = utils.load_align_dict(args.replace_unk)
print('| Type the input sentence and press return:')
for src_str in sys.stdin:
src_str = src_str.strip()
src_tokens = tokenizer.Tokenizer.tokenize(
src_str, src_dict, add_if_not_exist=False).long()
if use_cuda:
src_tokens = src_tokens.cuda()
translations = translator.generate(Variable(src_tokens.view(1, -1)))
hypos = translations[0]
print('O\t{}'.format(src_str))
# Process top predictions
for hypo in hypos[:min(len(hypos), args.nbest)]:
hypo_tokens, hypo_str, alignment = utils.post_process_prediction(
hypo_tokens=hypo['tokens'].int().cpu(),
src_str=src_str,
alignment=hypo['alignment'].int().cpu(),
align_dict=align_dict,
dst_dict=dst_dict,
remove_bpe=args.remove_bpe)
print('H\t{}\t{}'.format(hypo['score'], hypo_str))
print('A\t{}'.format(' '.join(map(str, alignment))))
def main():
parser = options.get_parser('Trainer')
dataset_args = options.add_dataset_args(parser)
dataset_args.add_argument('--max-tokens',
default=6000,
type=int,
metavar='N',
help='maximum number of tokens in a batch')
dataset_args.add_argument('--max-sentences',
type=int,
metavar='N',
help='maximum number of sentences in a batch')
dataset_args.add_argument(
'--train-subset',
default='train',
metavar='SPLIT',
choices=['train', 'valid', 'test'],
help='data subset to use for training (train, valid, test)')
dataset_args.add_argument(
'--valid-subset',
default='valid',
metavar='SPLIT',
help='comma separated list of data subsets '
' to use for validation (train, valid, valid1,test, test1)')
dataset_args.add_argument(
'--max-sentences-valid',
type=int,
metavar='N',
help='maximum number of sentences in a validation batch')
options.add_optimization_args(parser)
options.add_checkpoint_args(parser)
options.add_model_args(parser)
args = utils.parse_args_and_arch(parser)
if args.no_progress_bar and args.log_format is None:
args.log_format = 'simple'
if args.max_sentences_valid is None:
args.max_sentences_valid = args.max_sentences
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
torch.manual_seed(args.seed)
# Load dataset
splits = ['train', 'valid']
if data.has_binary_files(args.data, splits):
dataset = data.load_dataset(args.data, splits, args.source_lang,
args.target_lang)
else:
dataset = data.load_raw_text_dataset(args.data, splits,
args.source_lang,
args.target_lang)
if args.source_lang is None or args.target_lang is None:
# record inferred languages in args, so that it's saved in checkpoints
args.source_lang, args.target_lang = dataset.src, dataset.dst
if not torch.cuda.is_available():
raise NotImplementedError('Training on CPU is not supported')
args.num_gpus = torch.cuda.device_count()
print(args)
print('| [{}] dictionary: {} types'.format(dataset.src,
len(dataset.src_dict)))
print('| [{}] dictionary: {} types'.format(dataset.dst,
len(dataset.dst_dict)))
for split in splits:
print('| {} {} {} examples'.format(args.data, split,
len(dataset.splits[split])))
print(
'| using {} GPUs (with max tokens per GPU = {} and max sentences per GPU = {})'
.format(args.num_gpus, args.max_tokens, args.max_sentences))
# Build model and criterion
model = utils.build_model(args, dataset.src_dict, dataset.dst_dict)
criterion = utils.build_criterion(args, dataset.src_dict, dataset.dst_dict)
print('| model {}, criterion {}'.format(args.arch,
criterion.__class__.__name__))
print('| num. model params: {}'.format(
sum(p.data.numel() for p in model.parameters())))
# The max number of positions can be different for train and valid
# e.g., RNNs may support more positions at test time than seen in training
max_positions_train = (min(args.max_source_positions,
model.max_encoder_positions()),
min(args.max_target_positions,
model.max_decoder_positions()))
max_positions_valid = (model.max_encoder_positions(),
model.max_decoder_positions())
# Start multiprocessing
trainer = MultiprocessingTrainer(args, model, criterion)
# Create files to save losses
traincsv_path = os.path.join(args.save_dir, 'train_losses.csv')
validcsv_path = os.path.join(args.save_dir, 'valid_losses.csv')
output_path = [traincsv_path, validcsv_path]
for path in output_path:
with open(path, 'w+') as csvfile:
csvwriter = csv.writer(csvfile, delimiter=',')
csvwriter.writerow(['Epoch', 'Perplexity', 'Loss'])
csvfile.close()
# Load the latest checkpoint if one is available
checkpoint_path = os.path.join(args.save_dir, args.restore_file)
extra_state = trainer.load_checkpoint(checkpoint_path)
if extra_state is not None:
epoch = extra_state['epoch']
batch_offset = extra_state['batch_offset']
print('| loaded checkpoint {} (epoch {})'.format(
checkpoint_path, epoch))
if batch_offset == 0:
epoch += 1
else:
epoch, batch_offset = 1, 0
# Train until the learning rate gets too small
val_loss = None
max_epoch = args.max_epoch or math.inf
lr = trainer.get_lr()
train_meter = StopwatchMeter()
train_meter.start()
while lr > args.min_lr and epoch <= max_epoch:
# train for one epoch
train(args, epoch, batch_offset, trainer, dataset, max_positions_train,
traincsv_path)
# evaluate on validate set
for k, subset in enumerate(args.valid_subset.split(',')):
val_loss = validate(args, epoch, trainer, dataset,
max_positions_valid, subset, validcsv_path)
if k == 0:
if not args.no_save:
# save checkpoint
save_checkpoint(trainer, args, epoch, 0, val_loss)
# only use first validation loss to update the learning schedule
lr = trainer.lr_step(val_loss, epoch)
epoch += 1
batch_offset = 0
train_meter.stop()
print('| done training in {:.1f} seconds'.format(train_meter.sum))
# Stop multiprocessing
trainer.stop()
def get_reranking_parser(default_task="translation"):
parser = options.get_parser("Generation and reranking", default_task)
add_reranking_args(parser)
return parser
def get_parser_with_args():
parser = options.get_parser("Generation")
pytorch_translate_options.add_verbosity_args(parser)
pytorch_translate_options.add_dataset_args(parser, gen=True)
generation_group = options.add_generation_args(parser)
pytorch_translate_options.expand_generation_args(generation_group)
# Adds args used by the standalone generate binary.
generation_group.add_argument(
"--source-vocab-file",
default="",
metavar="FILE",
help="Path to text file representing the Dictionary to use.",
)
generation_group.add_argument(
"--char-source-vocab-file",
default="",
metavar="FILE",
help=(
"Same as --source-vocab-file except using characters. "
"(For use with char_source models only.)"
),
)
generation_group.add_argument(
"--target-vocab-file",
default="",
metavar="FILE",
help="Path to text file representing the Dictionary to use.",
)
generation_group.add_argument(
"--source-text-file",
default="",
nargs="+",
metavar="FILE",
help="Path to raw text file containing examples in source dialect. "
"This overrides what would be loaded from the data dir. "
"You can specify multiple source files (eg. for use in combination "
"with --source-ensembling). By default this will only translate the "
"first source file",
)
generation_group.add_argument(
"--target-text-file",
default="",
metavar="FILE",
help="Path to raw text file containing examples in target dialect. "
"This overrides what would be loaded from the data dir.",
)
generation_group.add_argument(
"--source-binary-file",
default="",
help="Path for the binary file containing source eval examples. "
"(Overrides --source-text-file. Must be used in conjunction with "
"--target-binary-file).",
)
generation_group.add_argument(
"--target-binary-file",
default="",
help="Path for the binary file containing target eval examples. "
"(Overrides --target-text-file. Must be used in conjunction with "
"--source-binary-file).",
)
generation_group.add_argument(
"--translation-output-file",
default="",
type=str,
metavar="FILE",
help="Path to text file to store the output of the model. ",
)
generation_group.add_argument(
"--translation-probs-file",
default="",
type=str,
metavar="FILE",
help="Path to text file to store the probs of translation output. ",
)
generation_group.add_argument(
"--multiling-source-lang-id",
type=int,
default=None,
help=(
"Must be set for decoding with multilingual models. Set to i if "
"the source language is the i-th language in the training parameter "
"--multiling-encoder-lang (0-indexed)"
),
)
generation_group.add_argument(
"--multiling-target-lang-id",
type=int,
default=None,
help=(
"Must be set for decoding with multilingual models. Set to i if "
"the target language is the i-th language in the training parameter "
"--multiling-decoder-lang (0-indexed)"
),
)
generation_group.add_argument(
"--source-ensembling",
action="store_true",
help="If this flag is present, the model will ensemble the predictions "
"conditioned on multiple source sentences (one per source-text-file)",
)
return parser
def main():
parser = options.get_parser('Generation')
parser.add_argument('--path', metavar='FILE', required=True, action='append',
help='path(s) to model file(s)')
dataset_args = options.add_dataset_args(parser)
dataset_args.add_argument('-i', '--interactive', action='store_true',
help='generate translations in interactive mode')
dataset_args.add_argument('--batch-size', default=32, type=int, metavar='N',
help='batch size')
dataset_args.add_argument('--gen-subset', default='test', metavar='SPLIT',
help='data subset to generate (train, valid, test)')
options.add_generation_args(parser)
args = parser.parse_args()
print(args)
if args.no_progress_bar:
progress_bar.enabled = False
use_cuda = torch.cuda.is_available() and not args.cpu
# Load model and dataset
print('| loading model(s) from {}'.format(', '.join(args.path)))
models, dataset = utils.load_ensemble_for_inference(args.path, args.data)
print('| [{}] dictionary: {} types'.format(dataset.src, len(dataset.src_dict)))
print('| [{}] dictionary: {} types'.format(dataset.dst, len(dataset.dst_dict)))
if not args.interactive:
print('| {} {} {} examples'.format(args.data, args.gen_subset, len(dataset.splits[args.gen_subset])))
# Optimize model for generation
for model in models:
model.make_generation_fast_(not args.no_beamable_mm)
# Initialize generator
translator = SequenceGenerator(models, dataset.dst_dict, beam_size=args.beam,
stop_early=(not args.no_early_stop),
normalize_scores=(not args.unnormalized),
len_penalty=args.lenpen)
align_dict = {}
if args.unk_replace_dict != '':
assert args.interactive, "Unkown words replacing requires access to original source and is only" \
"supported in interactive mode"
with open(args.unk_replace_dict, 'r') as f:
for line in f:
l = line.split()
align_dict[l[0]] = l[1]
def replace_unk(hypo_str, align_str, src, unk):
hypo_tokens = hypo_str.split()
src_tokens = tokenizer.tokenize_line(src)
align_idx = [int(i) for i in align_str.split()]
for i, ht in enumerate(hypo_tokens):
if ht == unk:
src_token = src_tokens[align_idx[i]]
if src_token in align_dict:
hypo_tokens[i] = align_dict[src_token]
else:
hypo_tokens[i] = src_token
return ' '.join(hypo_tokens)
if use_cuda:
translator.cuda()
bpe_symbol = '@@ ' if args.remove_bpe else None
def display_hypotheses(id, src, orig, ref, hypos):
id_str = '' if id is None else '-{}'.format(id)
src_str = to_sentence(dataset.src_dict, src, bpe_symbol)
print('S{}\t{}'.format(id_str, src_str))
if orig is not None:
print('O{}\t{}'.format(id_str, orig.strip()))
if ref is not None:
print('T{}\t{}'.format(id_str, to_sentence(dataset.dst_dict, ref, bpe_symbol, ref_unk=True)))
for hypo in hypos:
hypo_str = to_sentence(dataset.dst_dict, hypo['tokens'], bpe_symbol)
align_str = ' '.join(map(str, hypo['alignment']))
if args.unk_replace_dict != '':
hypo_str = replace_unk(hypo_str, align_str, orig, unk_symbol(dataset.dst_dict))
print('H{}\t{}\t{}'.format(
id_str, hypo['score'], hypo_str))
print('A{}\t{}'.format(id_str, align_str))
if args.interactive:
for line in sys.stdin:
tokens = tokenizer.Tokenizer.tokenize(line, dataset.src_dict, add_if_not_exist=False).long()
start = dataset.src_dict.pad() + 1
positions = torch.arange(start, start + len(tokens)).type_as(tokens)
if use_cuda:
positions = positions.cuda()
tokens = tokens.cuda()
translations = translator.generate(Variable(tokens.view(1, -1)), Variable(positions.view(1, -1)))
hypos = translations[0]
display_hypotheses(None, tokens, line, None, hypos[:min(len(hypos), args.nbest)])
else:
def maybe_remove_bpe(tokens):
"""Helper for removing BPE symbols from a hypothesis."""
if not args.remove_bpe:
return tokens
assert (tokens == dataset.dst_dict.pad()).sum() == 0
hypo_minus_bpe = to_sentence(dataset.dst_dict, tokens, bpe_symbol)
return tokenizer.Tokenizer.tokenize(hypo_minus_bpe, dataset.dst_dict, add_if_not_exist=True)
# Generate and compute BLEU score
scorer = bleu.Scorer(dataset.dst_dict.pad(), dataset.dst_dict.eos(), dataset.dst_dict.unk())
itr = dataset.dataloader(args.gen_subset, batch_size=args.batch_size, max_positions=args.max_positions)
num_sentences = 0
with progress_bar(itr, smoothing=0, leave=False) as t:
wps_meter = TimeMeter()
gen_timer = StopwatchMeter()
translations = translator.generate_batched_itr(
t, maxlen_a=args.max_len_a, maxlen_b=args.max_len_b,
cuda_device=0 if use_cuda else None, timer=gen_timer)
for id, src, ref, hypos in translations:
ref = ref.int().cpu()
top_hypo = hypos[0]['tokens'].int().cpu()
scorer.add(maybe_remove_bpe(ref), maybe_remove_bpe(top_hypo))
display_hypotheses(id, src, None, ref, hypos[:min(len(hypos), args.nbest)])
wps_meter.update(src.size(0))
t.set_postfix(wps='{:5d}'.format(round(wps_meter.avg)))
num_sentences += 1
print('| Translated {} sentences ({} tokens) in {:.1f}s ({:.2f} tokens/s)'.format(
num_sentences, gen_timer.n, gen_timer.sum, 1. / gen_timer.avg))
print('| Generate {} with beam={}: {}'.format(args.gen_subset, args.beam, scorer.result_string()))
valid_align_path = args.valid_pre + ".{}-{}.".format(
src, tgt) + args.align_suffix
make_binary_alignment_dataset(valid_align_path,
"valid.align",
src,
tgt,
num_workers=args.workers)
if args.test_pre:
test_align_path = args.test_pre + ".{}-{}.".format(
src, tgt) + args.align_suffix
make_binary_alignment_dataset(test_align_path,
"test.align",
src,
tgt,
num_workers=args.workers)
for src in args.source_langs:
for tgt in args.target_langs:
make_all(src, tgt)
if args.align_suffix:
make_all_alignments(src, tgt)
print("| Wrote preprocessed data to {}".format(args.dest_dir))
if __name__ == "__main__":
parser = options.get_parser('Preprocessing', default_task='translation')
add_preprocess_args(parser)
args = parser.parse_args()
main(args)
def main():
parser = options.get_parser('Trainer')
dataset_args = options.add_dataset_args(parser)
dataset_args.add_argument('--max-tokens', default=6000, type=int, metavar='N',
help='maximum number of tokens in a batch')
dataset_args.add_argument('--max-sentences', type=int, metavar='N',
help='maximum number of sentences in a batch')
dataset_args.add_argument('--train-subset', default='train', metavar='SPLIT',
choices=['train', 'valid', 'test'],
help='data subset to use for training (train, valid, test)')
dataset_args.add_argument('--valid-subset', default='valid', metavar='SPLIT',
help='comma separated list of data subsets '
' to use for validation (train, valid, valid1,test, test1)')
options.add_optimization_args(parser)
options.add_checkpoint_args(parser)
options.add_model_args(parser)
args = utils.parse_args_and_arch(parser)
if args.no_progress_bar and args.log_format is None:
args.log_format = 'simple'
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
torch.manual_seed(args.seed)
# Load dataset
splits = ['train', 'valid']
if data.has_binary_files(args.data, splits):
dataset = data.load_dataset(args.data, splits, args.source_lang, args.target_lang)
else:
dataset = data.load_raw_text_dataset(args.data, splits, args.source_lang, args.target_lang)
if args.source_lang is None or args.target_lang is None:
# record inferred languages in args, so that it's saved in checkpoints
args.source_lang, args.target_lang = dataset.src, dataset.dst
print(args)
print('| [{}] dictionary: {} types'.format(dataset.src, len(dataset.src_dict)))
print('| [{}] dictionary: {} types'.format(dataset.dst, len(dataset.dst_dict)))
for split in splits:
print('| {} {} {} examples'.format(args.data, split, len(dataset.splits[split])))
if not torch.cuda.is_available():
raise NotImplementedError('Training on CPU is not supported')
num_gpus = torch.cuda.device_count()
print('| using {} GPUs (with max tokens per GPU = {} and max sentences per GPU = {})'.format(
num_gpus, args.max_tokens, args.max_sentences))
# Build model and criterion
model = utils.build_model(args, dataset.src_dict, dataset.dst_dict)
criterion = utils.build_criterion(args, dataset.src_dict, dataset.dst_dict)
print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__))
# The max number of positions can be different for train and valid
# e.g., RNNs may support more positions at test time than seen in training
max_positions_train = (args.max_source_positions, args.max_target_positions)
max_positions_valid = (
min(args.max_source_positions, model.max_encoder_positions()),
min(args.max_target_positions, model.max_decoder_positions())
)
# Start multiprocessing
trainer = MultiprocessingTrainer(args, model, criterion)
# Load the latest checkpoint if one is available
checkpoint_path = os.path.join(args.save_dir, args.restore_file)
extra_state = trainer.load_checkpoint(checkpoint_path)
if extra_state is not None:
epoch = extra_state['epoch']
batch_offset = extra_state['batch_offset']
print('| loaded checkpoint {} (epoch {})'.format(checkpoint_path, epoch))
if batch_offset == 0:
epoch += 1
else:
epoch, batch_offset = 1, 0
# Train until the learning rate gets too small
val_loss = None
max_epoch = args.max_epoch or math.inf
lr = trainer.get_lr()
train_meter = StopwatchMeter()
train_meter.start()
while lr > args.min_lr and epoch <= max_epoch:
# train for one epoch
train(args, epoch, batch_offset, trainer, dataset, max_positions_train, num_gpus)
# evaluate on validate set
for k, subset in enumerate(args.valid_subset.split(',')):
val_loss = validate(args, epoch, trainer, dataset, max_positions_valid, subset, num_gpus)
if k == 0:
if not args.no_save:
# save checkpoint
save_checkpoint(trainer, args, epoch, 0, val_loss)
# only use first validation loss to update the learning schedule
lr = trainer.lr_step(val_loss, epoch)
epoch += 1
batch_offset = 0
train_meter.stop()
print('| done training in {:.1f} seconds'.format(train_meter.sum))
# Stop multiprocessing
trainer.stop()
def main():
parser = options.get_parser('Trainer')
dataset_args = options.add_dataset_args(parser)
dataset_args.add_argument('--max-tokens',
default=0,
type=int,
metavar='N',
help='maximum number of tokens in a batch')
dataset_args.add_argument('--batch-size',
default=32,
type=int,
metavar='N',
help='batch size')
dataset_args.add_argument('--test-batch-size',
default=32,
type=int,
metavar='N',
help='batch size for test set')
dataset_args.add_argument('--valid-batch-size',
default=32,
type=int,
metavar='N',
help='batch size for validation set')
dataset_args.add_argument(
'--train-subset',
default='train',
metavar='SPLIT',
choices=['train', 'valid', 'test'],
help='data subset to use for training (train, valid, test)')
dataset_args.add_argument(
'--valid-subset',
default='valid',
metavar='SPLIT',
help='comma separated list ofdata subsets '
' to use for validation (train, valid, valid1,test, test1)')
dataset_args.add_argument('--test-subset',
default='test',
metavar='SPLIT',
help='comma separated list ofdata subset '
'to use for testing (train, valid, test)')
dataset_args.add_argument(
'--valid-script',
nargs='+',
metavar='PATH',
help='path to external validation script (optional).')
options.add_optimization_args(parser)
options.add_checkpoint_args(parser)
options.add_model_args(parser)
args = utils.parse_args_and_arch(parser)
print(args)
if args.no_progress_bar:
progress_bar.enabled = False
progress_bar.print_interval = args.log_interval
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
torch.manual_seed(args.seed)
# Setting args.max_tokens to infinity(same as setting to None)
if args.max_tokens == 0:
args.max_tokens = None
# Load dataset
dataset = data.load_with_check(args.data, args.source_lang,
args.target_lang)
if args.source_lang is None or args.target_lang is None:
# record inferred languages in args, so that it's saved in checkpoints
args.source_lang, args.target_lang = dataset.src, dataset.dst
print('| [{}] dictionary: {} types'.format(dataset.src,
len(dataset.src_dict)))
print('| [{}] dictionary: {} types'.format(dataset.dst,
len(dataset.dst_dict)))
for split in dataset.splits:
print('| {} {} {} examples'.format(args.data, split,
len(dataset.splits[split])))
if not torch.cuda.is_available():
raise NotImplementedError('Training on CPU is not supported')
num_gpus = torch.cuda.device_count()
print('| using {} GPUs (with max tokens per GPU = {})'.format(
num_gpus, args.max_tokens))
# Build model
print('| model {}'.format(args.arch))
model = utils.build_model(args, dataset)
criterion = utils.build_criterion(args, dataset)
# Start multiprocessing
trainer = MultiprocessingTrainer(args, model)
# Load the latest checkpoint if one is available
epoch, batch_offset = trainer.load_checkpoint(
os.path.join(args.save_dir, args.restore_file))
# Train until the learning rate gets too small
val_loss = None
max_epoch = args.max_epoch or math.inf
lr = trainer.get_lr()
train_meter = StopwatchMeter()
train_meter.start()
while lr > args.min_lr and epoch <= max_epoch:
# train for one epoch
train(args, epoch, batch_offset, trainer, criterion, dataset, num_gpus)
# evaluate on validate set
for k, subset in enumerate(args.valid_subset.split(',')):
val_loss = validate(args, epoch, trainer, criterion, dataset,
subset, num_gpus)
if k == 0:
if not args.no_save:
# save checkpoint
trainer.save_checkpoint(
args,
epoch,
0,
val_loss,
validation_script=args.valid_script)
# only use first validation loss to update the learning schedule
lr = trainer.lr_step(val_loss, epoch)
epoch += 1
batch_offset = 0
train_meter.stop()
print('| done training in {:.1f} seconds'.format(train_meter.sum))
# Generate on test set and compute BLEU score
for beam in [1, 5, 10, 20]:
for subset in args.test_subset.split(','):
scorer = score_test(args,
trainer.get_model(),
dataset,
subset,
beam,
cuda_device=(0 if num_gpus > 0 else None))
print('| Test on {} with beam={}: {}'.format(
subset, beam, scorer.result_string()))
# Stop multiprocessing
trainer.stop()
def get_lm_scorer_parser(default_task='language_modeling'):
parser = options.get_parser('Evaluate Language Model', default_task)
options.add_dataset_args(parser, gen=True)
options.add_common_eval_args(parser)
add_lm_scorer_args(parser)
return parser
def get_parser_with_args():
parser = options.get_parser("Generation", default_task="pytorch_translate")
pytorch_translate_options.add_verbosity_args(parser)
pytorch_translate_options.add_dataset_args(parser, gen=True)
generation_group = options.add_generation_args(parser)
pytorch_translate_options.expand_generation_args(generation_group)
# Adds args used by the standalone generate binary.
generation_group.add_argument(
"--source-vocab-file",
default="",
metavar="FILE",
help="Path to text file representing the Dictionary to use.",
)
generation_group.add_argument(
"--char-source-vocab-file",
default="",
metavar="FILE",
help=(
"Same as --source-vocab-file except using characters. "
"(For use with char_source models only.)"
),
)
generation_group.add_argument(
"--target-vocab-file",
default="",
metavar="FILE",
help="Path to text file representing the Dictionary to use.",
)
generation_group.add_argument(
"--source-text-file",
default="",
nargs="+",
metavar="FILE",
help="Path to raw text file containing examples in source dialect. "
"This overrides what would be loaded from the data dir. "
"You can specify multiple source files (eg. for use in combination "
"with --source-ensembling). By default this will only translate the "
"first source file",
)
generation_group.add_argument(
"--target-text-file",
default="",
metavar="FILE",
help="Path to raw text file containing examples in target dialect. "
"This overrides what would be loaded from the data dir.",
)
generation_group.add_argument(
"--source-binary-file",
default="",
help="Path for the binary file containing source eval examples. "
"(Overrides --source-text-file. Must be used in conjunction with "
"--target-binary-file).",
)
generation_group.add_argument(
"--target-binary-file",
default="",
help="Path for the binary file containing target eval examples. "
"(Overrides --target-text-file. Must be used in conjunction with "
"--source-binary-file).",
)
generation_group.add_argument(
"--translation-output-file",
default="",
type=str,
metavar="FILE",
help="Path to text file to store the output of the model. ",
)
generation_group.add_argument(
"--translation-probs-file",
default="",
type=str,
metavar="FILE",
help="Path to text file to store the probs of translation output. ",
)
generation_group.add_argument(
"--multiling-source-lang",
action="append",
metavar="SRC",
help=(
"Must be set for decoding with multilingual models. "
"Must match an entry from --multiling-encoder-lang from training."
),
)
generation_group.add_argument(
"--multiling-target-lang",
action="append",
metavar="TARGET",
help=(
"Must be set for decoding with multilingual models. "
"Must match an entry from --multiling-decoder-lang from training."
),
)
generation_group.add_argument(
"--source-ensembling",
action="store_true",
help="If this flag is present, the model will ensemble the predictions "
"conditioned on multiple source sentences (one per source-text-file)",
)
generation_group.add_argument(
"--competing-completed-beam-search",
action="store_true",
help="If this flag is present, use the alternative beam search "
"implementation in research/beam_search. This beam search keeps completed "
"hypos in the beam and let them compete against hypo expansions in the "
"next time step.",
)
return parser
def main():
parser = options.get_parser('Generation')
parser.add_argument('--path', metavar='FILE', required=True, action='append',
help='path(s) to model file(s)')
dataset_args = options.add_dataset_args(parser)
dataset_args.add_argument('--batch-size', default=32, type=int, metavar='N',
help='batch size')
dataset_args.add_argument('--gen-subset', default='test', metavar='SPLIT',
help='data subset to generate (train, valid, test)')
options.add_generation_args(parser)
args = parser.parse_args()
if args.no_progress_bar and args.log_format is None:
args.log_format = 'none'
print(args)
use_cuda = torch.cuda.is_available() and not args.cpu
# Load dataset
if args.replace_unk is None:
dataset = data.load_dataset(args.data, [args.gen_subset], args.source_lang, args.target_lang)
else:
dataset = data.load_raw_text_dataset(args.data, [args.gen_subset], args.source_lang, args.target_lang)
if args.source_lang is None or args.target_lang is None:
# record inferred languages in args
args.source_lang, args.target_lang = dataset.src, dataset.dst
# Load ensemble
print('| loading model(s) from {}'.format(', '.join(args.path)))
models, _ = utils.load_ensemble_for_inference(args.path, dataset.src_dict, dataset.dst_dict)
print('| [{}] dictionary: {} types'.format(dataset.src, len(dataset.src_dict)))
print('| [{}] dictionary: {} types'.format(dataset.dst, len(dataset.dst_dict)))
print('| {} {} {} examples'.format(args.data, args.gen_subset, len(dataset.splits[args.gen_subset])))
# Optimize ensemble for generation
for model in models:
model.make_generation_fast_(
beamable_mm_beam_size=None if args.no_beamable_mm else args.beam)
# Initialize generator
translator = SequenceGenerator(
models, beam_size=args.beam, stop_early=(not args.no_early_stop),
normalize_scores=(not args.unnormalized), len_penalty=args.lenpen,
unk_penalty=args.unkpen)
if use_cuda:
translator.cuda()
# Load alignment dictionary for unknown word replacement
# (None if no unknown word replacement, empty if no path to align dictionary)
align_dict = utils.load_align_dict(args.replace_unk)
# Generate and compute BLEU score
scorer = bleu.Scorer(dataset.dst_dict.pad(), dataset.dst_dict.eos(), dataset.dst_dict.unk())
max_positions = min(model.max_encoder_positions() for model in models)
itr = dataset.eval_dataloader(
args.gen_subset, max_sentences=args.batch_size, max_positions=max_positions,
skip_invalid_size_inputs_valid_test=args.skip_invalid_size_inputs_valid_test)
num_sentences = 0
with utils.build_progress_bar(args, itr) as t:
wps_meter = TimeMeter()
gen_timer = StopwatchMeter()
translations = translator.generate_batched_itr(
t, maxlen_a=args.max_len_a, maxlen_b=args.max_len_b,
cuda_device=0 if use_cuda else None, timer=gen_timer)
for sample_id, src_tokens, target_tokens, hypos in translations:
# Process input and ground truth
target_tokens = target_tokens.int().cpu()
# Either retrieve the original sentences or regenerate them from tokens.
if align_dict is not None:
src_str = dataset.splits[args.gen_subset].src.get_original_text(sample_id)
target_str = dataset.splits[args.gen_subset].dst.get_original_text(sample_id)
else:
src_str = dataset.src_dict.string(src_tokens, args.remove_bpe)
target_str = dataset.dst_dict.string(target_tokens, args.remove_bpe, escape_unk=True)
if not args.quiet:
print('S-{}\t{}'.format(sample_id, src_str))
print('T-{}\t{}'.format(sample_id, target_str))
# Process top predictions
for i, hypo in enumerate(hypos[:min(len(hypos), args.nbest)]):
hypo_tokens, hypo_str, alignment = utils.post_process_prediction(
hypo_tokens=hypo['tokens'].int().cpu(),
src_str=src_str,
alignment=hypo['alignment'].int().cpu(),
align_dict=align_dict,
dst_dict=dataset.dst_dict,
remove_bpe=args.remove_bpe)
if not args.quiet:
print('H-{}\t{}\t{}'.format(sample_id, hypo['score'], hypo_str))
print('A-{}\t{}'.format(sample_id, ' '.join(map(str, alignment))))
# Score only the top hypothesis
if i == 0:
if align_dict is not None or args.remove_bpe is not None:
# Convert back to tokens for evaluation with unk replacement and/or without BPE
target_tokens = tokenizer.Tokenizer.tokenize(target_str,
dataset.dst_dict,
add_if_not_exist=True)
scorer.add(target_tokens, hypo_tokens)
wps_meter.update(src_tokens.size(0))
t.log({'wps': round(wps_meter.avg)})
num_sentences += 1
print('| Translated {} sentences ({} tokens) in {:.1f}s ({:.2f} tokens/s)'.format(
num_sentences, gen_timer.n, gen_timer.sum, 1. / gen_timer.avg))
print('| Generate {} with beam={}: {}'.format(args.gen_subset, args.beam, scorer.result_string()))
