def load_dataset(self,
split,
src_bin_path,
tgt_bin_path,
weights_file=None):
corpus = pytorch_translate_data.ParallelCorpusConfig(
source=pytorch_translate_data.CorpusConfig(
dialect=self.args.source_lang, data_file=src_bin_path),
target=pytorch_translate_data.CorpusConfig(
dialect=self.args.target_lang, data_file=tgt_bin_path),
weights_file=weights_file,
)
if self.args.log_verbose:
print("Starting to load binarized data files.", flush=True)
if not os.path.exists(corpus.source.data_file):
raise ValueError(
f"{corpus.source.data_file} for {split} not found!")
if not os.path.exists(corpus.target.data_file):
raise ValueError(
f"{corpus.target.data_file} for {split} not found!")
dst_dataset = pytorch_translate_data.InMemoryNumpyDataset.create_from_file(
corpus.target.data_file)
weights_dataset = None
if corpus.weights_file and os.path.exists(corpus.weights_file):
weights_dataset = weighted_data.IndexedWeightsDataset(
corpus.weights_file)
assert len(dst_dataset) == len(weights_dataset)
if self.char_source_dict is not None:
src_dataset = char_data.InMemoryNumpyWordCharDataset.create_from_file(
corpus.source.data_file)
self.datasets[split] = char_data.LanguagePairSourceCharDataset(
src=src_dataset,
src_sizes=src_dataset.sizes,
src_dict=self.source_dictionary,
tgt=dst_dataset,
tgt_sizes=dst_dataset.sizes,
tgt_dict=self.target_dictionary,
weights=weights_dataset,
)
else:
src_dataset = pytorch_translate_data.InMemoryNumpyDataset.create_from_file(
corpus.source.data_file)
self.datasets[split] = weighted_data.WeightedLanguagePairDataset(
src=src_dataset,
src_sizes=src_dataset.sizes,
src_dict=self.source_dictionary,
tgt=dst_dataset,
tgt_sizes=dst_dataset.sizes,
tgt_dict=self.target_dictionary,
weights=weights_dataset,
)
if self.args.log_verbose:
print("Finished loading dataset", flush=True)
print(f"| {split} {len(self.datasets[split])} examples")
def load_dataset(self,
split,
src_bin_path,
tgt_bin_path,
weights_file=None,
is_train=False):
"""
Currently this method does not support character models.
"""
corpus = pytorch_translate_data.ParallelCorpusConfig(
source=pytorch_translate_data.CorpusConfig(
dialect=self.args.source_lang, data_file=src_bin_path),
target=pytorch_translate_data.CorpusConfig(
dialect=self.args.target_lang, data_file=tgt_bin_path),
weights_file=weights_file,
)
if self.args.log_verbose:
print("Starting to load binarized data files.", flush=True)
data_utils.validate_corpus_exists(corpus=corpus, split=split)
dst_dataset = pytorch_translate_data.InMemoryNumpyDataset.create_from_file(
corpus.target.data_file)
src_dataset = pytorch_translate_data.InMemoryNumpyDataset.create_from_file(
corpus.source.data_file)
if is_train:
self.datasets[split] = TeacherDataset(
src=src_dataset,
src_sizes=src_dataset.sizes,
src_dict=self.src_dict,
tgt=dst_dataset,
tgt_sizes=dst_dataset.sizes,
tgt_dict=self.tgt_dict,
teacher_models=self.teacher_models,
top_k_teacher_tokens=self.top_k_teacher_tokens,
top_k_teacher_scores=self.top_k_teacher_scores,
top_k_teacher_indices=self.top_k_teacher_indices,
left_pad_source=False,
)
else:
self.datasets[split] = weighted_data.WeightedLanguagePairDataset(
src=src_dataset,
src_sizes=src_dataset.sizes,
src_dict=self.src_dict,
tgt=dst_dataset,
tgt_sizes=dst_dataset.sizes,
tgt_dict=self.tgt_dict,
weights=None,
left_pad_source=False,
)
if self.args.log_verbose:
print("Finished loading dataset", flush=True)
print(f"| {split} {len(self.datasets[split])} examples")
def load_parallel_dataset(
source_lang,
target_lang,
src_bin_path,
tgt_bin_path,
source_dictionary,
target_dictionary,
split,
remove_eos_from_source,
append_eos_to_target=True,
char_source_dict=None,
log_verbose=True,
):
corpus = pytorch_translate_data.ParallelCorpusConfig(
source=pytorch_translate_data.CorpusConfig(
dialect=source_lang, data_file=src_bin_path
),
target=pytorch_translate_data.CorpusConfig(
dialect=target_lang, data_file=tgt_bin_path
),
weights_file=None,
)
if log_verbose:
print("Starting to load binarized data files.", flush=True)
validate_corpus_exists(corpus=corpus, split=split)
tgt_dataset = pytorch_translate_data.InMemoryNumpyDataset.create_from_file(
corpus.target.data_file
)
if char_source_dict is not None:
src_dataset = char_data.InMemoryNumpyWordCharDataset.create_from_file(
corpus.source.data_file
)
else:
src_dataset = pytorch_translate_data.InMemoryNumpyDataset.create_from_file(
corpus.source.data_file
)
parallel_dataset = weighted_data.WeightedLanguagePairDataset(
src=src_dataset,
src_sizes=src_dataset.sizes,
src_dict=source_dictionary,
tgt=tgt_dataset,
tgt_sizes=tgt_dataset.sizes,
tgt_dict=target_dictionary,
remove_eos_from_source=remove_eos_from_source,
append_eos_to_target=append_eos_to_target,
)
return parallel_dataset, src_dataset, tgt_dataset
def load_dataset(self,
split,
src_bin_path,
tgt_bin_path,
seed=None,
use_noiser=False):
"""
Load a dataset split. Seed and noiser are only used for loading train
data, not eval data.
"""
parallel_dataset, src_dataset, tgt_dataset = data_utils.load_parallel_dataset(
source_lang=self.source_lang,
target_lang=self.target_lang,
src_bin_path=src_bin_path,
tgt_bin_path=tgt_bin_path,
source_dictionary=self.source_dictionary,
target_dictionary=self.target_dictionary,
split=split,
remove_eos_from_source=not self.args.append_eos_to_source,
append_eos_to_target=True,
char_source_dict=self.char_source_dict,
log_verbose=self.args.log_verbose,
)
dataset_map = OrderedDict([(f"{self.source_lang}-{self.target_lang}",
parallel_dataset)])
if use_noiser:
if getattr(self.args, "denoising_source_parallel", False):
dataset_map[(
f"{self.source_lang}-{self.source_lang}"
)] = weighted_data.WeightedLanguagePairDataset(
src=noising.NoisingDataset(
src_dataset=src_dataset,
src_dict=self.source_dictionary,
seed=seed,
noiser=self.source_noiser,
),
tgt=src_dataset,
src_sizes=src_dataset.sizes,
src_dict=self.source_dictionary,
remove_eos_from_source=not self.args.append_eos_to_source,
append_eos_to_target=True,
)
if getattr(self.args, "denoising_target_parallel", False):
dataset_map[(
f"{self.target_lang}-{self.target_lang}"
)] = weighted_data.WeightedLanguagePairDataset(
src=noising.NoisingDataset(
src_dataset=tgt_dataset,
src_dict=self.target_dictionary,
seed=seed,
noiser=self.target_noiser,
),
tgt=tgt_dataset,
src_sizes=tgt_dataset.sizes,
src_dict=self.target_dictionary,
remove_eos_from_source=not self.args.append_eos_to_source,
append_eos_to_target=True,
)
if getattr(self.args, "denoising_source_mono", False):
source_mono_dataset = self.load_monolingual_dataset(
self.args.train_mono_source_binary_path)
dataset_map[(
f"{self.source_lang}-{self.source_lang}_"
f"{constants.MONOLINGUAL_DATA_IDENTIFIER}"
)] = weighted_data.WeightedLanguagePairDataset(
src=noising.NoisingDataset(
src_dataset=source_mono_dataset,
src_dict=self.source_dictionary,
seed=seed,
noiser=self.source_noiser,
),
tgt=source_mono_dataset,
src_sizes=source_mono_dataset.sizes,
src_dict=self.source_dictionary,
remove_eos_from_source=not self.args.append_eos_to_source,
append_eos_to_target=True,
)
if getattr(self.args, "denoising_target_mono", False):
target_mono_dataset = self.load_monolingual_dataset(
self.args.train_mono_target_binary_path)
dataset_map[(
f"{self.target_lang}-{self.target_lang}_"
f"{constants.MONOLINGUAL_DATA_IDENTIFIER}"
)] = weighted_data.WeightedLanguagePairDataset(
src=noising.NoisingDataset(
src_dataset=target_mono_dataset,
src_dict=self.target_dictionary,
seed=seed,
noiser=self.target_noiser,
),
tgt=target_mono_dataset,
src_sizes=target_mono_dataset.sizes,
src_dict=self.target_dictionary,
remove_eos_from_source=not self.args.append_eos_to_source,
append_eos_to_target=True,
)
self.datasets[split] = RoundRobinZipDatasets(dataset_map)
if self.args.log_verbose:
print("Finished loading dataset", flush=True)
print(f"| {split} {len(self.datasets[split])} datasets")
def load_dataset(
self, split, src_bin_path, tgt_bin_path, forward_model=None, backward_model=None
):
"""Load a dataset split."""
corpus = ptt_data.ParallelCorpusConfig(
source=ptt_data.CorpusConfig(
dialect=self.source_lang, data_file=src_bin_path
),
target=ptt_data.CorpusConfig(
dialect=self.target_lang, data_file=tgt_bin_path
),
weights_file=None,
)
if self.args.log_verbose:
print("Starting to load binarized data files.", flush=True)
data_utils.validate_corpus_exists(corpus=corpus, split=split)
forward_tgt_dataset = ptt_data.InMemoryNumpyDataset.create_from_file(
corpus.target.data_file
)
backward_tgt_dataset = ptt_data.InMemoryNumpyDataset.create_from_file(
corpus.source.data_file
)
forward_src_dataset = ptt_data.InMemoryNumpyDataset.create_from_file(
corpus.source.data_file
)
backward_src_dataset = ptt_data.InMemoryNumpyDataset.create_from_file(
corpus.target.data_file
)
forward_parallel_dataset = weighted_data.WeightedLanguagePairDataset(
src=forward_src_dataset,
src_sizes=forward_src_dataset.sizes,
src_dict=self.source_dictionary,
tgt=forward_tgt_dataset,
tgt_sizes=forward_tgt_dataset.sizes,
tgt_dict=self.target_dictionary,
remove_eos_from_source=self.remove_eos_from_source,
append_eos_to_target=True,
)
backward_parallel_dataset = weighted_data.WeightedLanguagePairDataset(
src=backward_src_dataset,
src_sizes=backward_src_dataset.sizes,
src_dict=self.target_dictionary,
tgt=backward_tgt_dataset,
tgt_sizes=backward_tgt_dataset.sizes,
tgt_dict=self.source_dictionary,
remove_eos_from_source=self.remove_eos_from_source,
append_eos_to_target=True,
)
dataset_map = OrderedDict(
[
(f"{self.source_lang}-{self.target_lang}", forward_parallel_dataset),
(f"{self.target_lang}-{self.source_lang}", backward_parallel_dataset),
]
)
assert (forward_model and backward_model) or (
forward_model is None and backward_model is None
), (
"Only one of forward or backward models can't be null;"
" both have to be non-null or null"
)
if forward_model and backward_model:
fwd_generator = beam_decode.SequenceGenerator(
models=[forward_model], tgt_dict=self.source_dictionary
)
bwd_generator = beam_decode.SequenceGenerator(
models=[backward_model], tgt_dict=self.target_dictionary
)
def monolingual_dataset(
path,
dictionary,
is_source=False,
num_examples_limit: Optional[int] = None,
):
dataset = self.load_monolingual_dataset(
path, is_source=is_source, num_examples_limit=num_examples_limit
)
return LanguagePairDataset(
src=dataset,
src_sizes=dataset.sizes,
src_dict=dictionary,
tgt=None,
tgt_sizes=None,
tgt_dict=None,
)
monolingual_num_examples_limit = None
if self.args.monolingual_ratio is not None:
monolingual_num_examples_limit = int(
self.args.monolingual_ratio * len(forward_parallel_dataset)
)
src_dataset = monolingual_dataset(
path=self.args.train_mono_source_binary_path,
dictionary=self.source_dictionary,
is_source=True,
num_examples_limit=monolingual_num_examples_limit,
)
tgt_dataset = monolingual_dataset(
path=self.args.train_mono_target_binary_path,
dictionary=self.target_dictionary,
is_source=False,
num_examples_limit=monolingual_num_examples_limit,
)
dataset_map[
f"{self.source_lang}-"
f"{self.target_lang}_{constants.MONOLINGUAL_DATA_IDENTIFIER}"
] = BacktranslationDataset(
tgt_dataset=TransformEosDataset(
dataset=tgt_dataset,
eos=self.target_dictionary.eos(),
# Remove EOS from the input before backtranslation.
remove_eos_from_src=True,
),
backtranslation_fn=bwd_generator.generate,
max_len_a=self.args.max_len_a,
max_len_b=self.args.max_len_b,
output_collater=TransformEosDataset(
dataset=tgt_dataset,
eos=self.target_dictionary.eos(),
# The original input (now the target) doesn't have
# an EOS, so we need to add one. The generated
# backtranslation (now the source) will have an EOS,
# so we want to remove it.
append_eos_to_tgt=True,
remove_eos_from_src=True,
).collater,
)
dataset_map[
f"{self.target_lang}-"
f"{self.source_lang}_{constants.MONOLINGUAL_DATA_IDENTIFIER}"
] = BacktranslationDataset(
tgt_dataset=src_dataset,
backtranslation_fn=fwd_generator.generate,
max_len_a=self.args.max_len_a,
max_len_b=self.args.max_len_b,
output_collater=TransformEosDataset(
dataset=src_dataset,
eos=self.source_dictionary.eos(),
# The original input (now the target) doesn't have
# an EOS, so we need to add one. The generated
# backtranslation (now the source) will have an EOS,
# so we want to remove it.
append_eos_to_tgt=True,
remove_eos_from_src=True,
).collater,
)
# print before loading RoundRobinZipDatasets to help catch any bugs
for dataset_key, dataset in dataset_map.items():
print(f"| {split}: {dataset_key} {len(dataset)} examples in dataset")
self.datasets[split] = RoundRobinZipDatasets(dataset_map)
print(
f"| {split} {len(self.datasets[split])} examples in RoundRobinZipDatasets"
)
if self.args.log_verbose:
print("Finished loading dataset", flush=True)
def load_binarized_dataset(
train_corpus: ParallelCorpusConfig,
eval_corpus: ParallelCorpusConfig,
train_split: str,
eval_split: str,
args: argparse.Namespace,
use_char_source: bool = False,
) -> data.LanguageDatasets:
if is_multilingual(args): # Dummy dictionaries
source_dict = pytorch_translate_dictionary.Dictionary()
target_dict = pytorch_translate_dictionary.Dictionary()
else:
source_dict = pytorch_translate_dictionary.Dictionary.load(
args.source_vocab_file)
target_dict = pytorch_translate_dictionary.Dictionary.load(
args.target_vocab_file)
if use_char_source:
char_source_dict = pytorch_translate_dictionary.Dictionary.load(
args.char_source_vocab_file)
# this attribute is used for CharSourceModel construction
args.char_source_dict_size = len(char_source_dict)
dataset = data.LanguageDatasets(
src=train_corpus.source.dialect,
dst=train_corpus.target.dialect,
src_dict=source_dict,
dst_dict=target_dict,
)
for split, corpus in [(train_split, train_corpus),
(eval_split, eval_corpus)]:
if not os.path.exists(corpus.source.data_file):
raise ValueError(
f"{corpus.source.data_file} for {split} not found!")
if not os.path.exists(corpus.target.data_file):
raise ValueError(
f"{corpus.target.data_file} for {split} not found!")
dst_dataset = InMemoryNumpyDataset.create_from_file(
corpus.target.data_file)
weights_dataset = None
if corpus.weights_file and os.path.exists(corpus.weights_file):
weights_dataset = weighted_data.IndexedWeightsDataset(
corpus.weights_file)
assert len(dst_dataset) == len(weights_dataset)
if use_char_source:
src_dataset = char_data.InMemoryNumpyWordCharDataset.create_from_file(
corpus.source.data_file)
dataset.splits[split] = char_data.LanguagePairSourceCharDataset(
src=src_dataset,
dst=dst_dataset,
pad_idx=source_dict.pad(),
eos_idx=source_dict.eos(),
weights=weights_dataset,
)
else:
src_dataset = InMemoryNumpyDataset.create_from_file(
corpus.source.data_file)
dataset.splits[split] = weighted_data.WeightedLanguagePairDataset(
src=src_dataset,
dst=dst_dataset,
pad_idx=source_dict.pad(),
eos_idx=source_dict.eos(),
weights=weights_dataset,
)
return dataset
def load_dataset(self,
split,
src_bin_path,
tgt_bin_path,
seed=None,
noiser=None):
"""
Load a dataset split. Seed and noiser are only used for loading train
data, not eval data.
"""
corpus = pytorch_translate_data.ParallelCorpusConfig(
source=pytorch_translate_data.CorpusConfig(
dialect=self.source_lang, data_file=src_bin_path),
target=pytorch_translate_data.CorpusConfig(
dialect=self.target_lang, data_file=tgt_bin_path),
weights_file=None,
)
if self.args.log_verbose:
print("Starting to load binarized data files.", flush=True)
data_utils.validate_corpus_exists(corpus=corpus, split=split)
tgt_dataset = pytorch_translate_data.InMemoryNumpyDataset.create_from_file(
corpus.target.data_file)
if self.char_source_dict is not None:
src_dataset = char_data.InMemoryNumpyWordCharDataset.create_from_file(
corpus.source.data_file)
else:
src_dataset = pytorch_translate_data.InMemoryNumpyDataset.create_from_file(
corpus.source.data_file)
parallel_dataset = weighted_data.WeightedLanguagePairDataset(
src=src_dataset,
src_sizes=src_dataset.sizes,
src_dict=self.source_dictionary,
tgt=tgt_dataset,
tgt_sizes=tgt_dataset.sizes,
tgt_dict=self.target_dictionary,
remove_eos_from_source=not self.args.append_eos_to_source,
append_eos_to_target=True,
)
dataset_map = OrderedDict([(f"{self.source_lang}-{self.target_lang}",
parallel_dataset)])
if noiser is not None:
if getattr(self.args, "denoising_source_parallel", False):
dataset_map[(
f"{self.source_lang}-{self.source_lang}"
)] = weighted_data.WeightedLanguagePairDataset(
src=noising.NoisingDataset(
src_dataset=src_dataset,
src_dict=self.source_dictionary,
seed=seed,
noiser=self.noiser,
),
tgt=src_dataset,
src_sizes=src_dataset.sizes,
src_dict=self.source_dictionary,
remove_eos_from_source=not self.args.append_eos_to_source,
append_eos_to_target=True,
)
if getattr(self.args, "denoising_target_parallel", False):
dataset_map[(
f"{self.target_lang}-{self.target_lang}"
)] = weighted_data.WeightedLanguagePairDataset(
src=noising.NoisingDataset(
src_dataset=tgt_dataset,
src_dict=self.target_dictionary,
seed=seed,
noiser=self.noiser,
),
tgt=tgt_dataset,
src_sizes=tgt_dataset.sizes,
src_dict=self.target_dictionary,
remove_eos_from_source=not self.args.append_eos_to_source,
append_eos_to_target=True,
)
if getattr(self.args, "denoising_source_mono", False):
source_mono_dataset = self.load_monolingual_dataset(
self.args.train_mono_source_binary_path)
dataset_map[(
f"{self.source_lang}-{self.source_lang}_"
f"{constants.MONOLINGUAL_DATA_IDENTIFIER}"
)] = weighted_data.WeightedLanguagePairDataset(
src=noising.NoisingDataset(
src_dataset=source_mono_dataset,
src_dict=self.source_dictionary,
seed=seed,
noiser=self.noiser,
),
tgt=source_mono_dataset,
src_sizes=source_mono_dataset.sizes,
src_dict=self.source_dictionary,
remove_eos_from_source=not self.args.append_eos_to_source,
append_eos_to_target=True,
)
if getattr(self.args, "denoising_target_mono", False):
target_mono_dataset = self.load_monolingual_dataset(
self.args.train_mono_target_binary_path)
dataset_map[(
f"{self.target_lang}-{self.target_lang}_"
f"{constants.MONOLINGUAL_DATA_IDENTIFIER}"
)] = weighted_data.WeightedLanguagePairDataset(
src=noising.NoisingDataset(
src_dataset=target_mono_dataset,
src_dict=self.target_dictionary,
seed=seed,
noiser=self.noiser,
),
tgt=target_mono_dataset,
src_sizes=target_mono_dataset.sizes,
src_dict=self.target_dictionary,
remove_eos_from_source=not self.args.append_eos_to_source,
append_eos_to_target=True,
)
self.datasets[split] = RoundRobinZipDatasets(dataset_map)
if self.args.log_verbose:
print("Finished loading dataset", flush=True)
print(f"| {split} {len(self.datasets[split])} datasets")
def load_dataset(self,
split,
src_bin_path,
tgt_bin_path,
forward_model=None,
backward_model=None):
"""Load a dataset split."""
corpus = ptt_data.ParallelCorpusConfig(
source=ptt_data.CorpusConfig(dialect=self.source_lang,
data_file=src_bin_path),
target=ptt_data.CorpusConfig(dialect=self.target_lang,
data_file=tgt_bin_path),
weights_file=None,
)
if self.args.log_verbose:
print("Starting to load binarized data files.", flush=True)
data_utils.validate_corpus_exists(corpus=corpus, split=split)
forward_tgt_dataset = ptt_data.InMemoryNumpyDataset.create_from_file(
corpus.target.data_file)
backward_tgt_dataset = ptt_data.InMemoryNumpyDataset.create_from_file(
corpus.source.data_file)
forward_src_dataset = ptt_data.InMemoryNumpyDataset.create_from_file(
corpus.source.data_file)
backward_src_dataset = ptt_data.InMemoryNumpyDataset.create_from_file(
corpus.target.data_file)
forward_parallel_dataset = weighted_data.WeightedLanguagePairDataset(
src=forward_src_dataset,
src_sizes=forward_src_dataset.sizes,
src_dict=self.source_dictionary,
tgt=forward_tgt_dataset,
tgt_sizes=forward_tgt_dataset.sizes,
tgt_dict=self.target_dictionary,
remove_eos_from_source=self.remove_eos_from_source,
append_eos_to_target=True,
)
backward_parallel_dataset = weighted_data.WeightedLanguagePairDataset(
src=backward_src_dataset,
src_sizes=backward_src_dataset.sizes,
src_dict=self.target_dictionary,
tgt=backward_tgt_dataset,
tgt_sizes=backward_tgt_dataset.sizes,
tgt_dict=self.source_dictionary,
remove_eos_from_source=self.remove_eos_from_source,
append_eos_to_target=True,
)
dataset_map = OrderedDict([
(f"{self.source_lang}-{self.target_lang}",
forward_parallel_dataset),
(f"{self.target_lang}-{self.source_lang}",
backward_parallel_dataset),
])
assert (forward_model and backward_model) or (
forward_model is None and backward_model is None), (
"Only one of forward or backward models can't be null;"
" both have to be non-null or null")
if forward_model and backward_model:
dataset_map[
f"{self.source_lang}-"
f"{self.target_lang}_{constants.MONOLINGUAL_DATA_IDENTIFIER}"] = BacktranslationDataset(
tgt_dataset=self.load_monolingual_dataset(
self.args.train_mono_target_binary_path),
tgt_dict=self.target_dictionary,
backtranslation_model=backward_model,
max_len_a=self.args.max_len_a,
max_len_b=self.args.max_len_b,
remove_eos_at_src=True,
generator_class=beam_decode.SequenceGenerator,
)
dataset_map[
f"{self.target_lang}-"
f"{self.source_lang}_{constants.MONOLINGUAL_DATA_IDENTIFIER}"] = BacktranslationDataset(
tgt_dataset=self.load_monolingual_dataset(
self.args.train_mono_source_binary_path),
tgt_dict=self.source_dictionary,
backtranslation_model=forward_model,
max_len_a=self.args.max_len_a,
max_len_b=self.args.max_len_b,
remove_eos_at_src=True,
generator_class=beam_decode.SequenceGenerator,
)
self.datasets[split] = RoundRobinZipDatasets(dataset_map)
if self.args.log_verbose:
print("Finished loading dataset", flush=True)
print(f"| {split} {len(self.datasets[split])} datasets")
