def load_dataset(self, split, combine=False, **kwargs):
"""Load a given dataset split (e.g., train, valid, test)."""
def get_path(key, split):
return os.path.join(self.cfg.data, key, split)
def make_dataset(key, dictionary):
split_path = get_path(key, split)
try:
dataset = data_utils.load_indexed_dataset(
split_path,
dictionary,
combine=combine,
)
except Exception as e:
if "StorageException: [404] Path not found" in str(e):
logger.warning(f"dataset {e} not found")
dataset = None
else:
raise e
return dataset
input0 = make_dataset("input0", self.source_dictionary)
assert input0 is not None, "could not find dataset: {}".format(
get_path("input0", split))
input1 = make_dataset("input1", self.source_dictionary)
if self.cfg.init_token is not None:
input0 = PrependTokenDataset(input0, self.cfg.init_token)
if input1 is None:
src_tokens = input0
else:
if self.cfg.separator_token is not None:
input1 = PrependTokenDataset(input1, self.cfg.separator_token)
src_tokens = ConcatSentencesDataset(input0, input1)
with data_utils.numpy_seed(self.cfg.seed):
shuffle = np.random.permutation(len(src_tokens))
src_tokens = maybe_shorten_dataset(
src_tokens,
split,
self.cfg.shorten_data_split_list,
self.cfg.shorten_method,
self.max_positions(),
self.cfg.seed,
)
dataset = {
"id": IdDataset(),
"net_input": {
"src_tokens":
RightPadDataset(
src_tokens,
pad_idx=self.source_dictionary.pad(),
),
"src_lengths":
NumelDataset(src_tokens, reduce=False),
},
"nsentences": NumSamplesDataset(),
"ntokens": NumelDataset(src_tokens, reduce=True),
}
if self.cfg.add_prev_output_tokens:
prev_tokens_dataset = RightPadDataset(
RollDataset(src_tokens, 1),
pad_idx=self.dictionary.pad(),
)
dataset["net_input"].update(
prev_output_tokens=prev_tokens_dataset, )
if not self.cfg.regression_target:
label_dataset = make_dataset("label", self.label_dictionary)
if label_dataset is not None:
dataset.update(target=OffsetTokensDataset(
StripTokenDataset(
label_dataset,
id_to_strip=self.label_dictionary.eos(),
),
offset=-self.label_dictionary.nspecial,
))
else:
label_path = "{0}.label".format(get_path("label", split))
if os.path.exists(label_path):
def parse_regression_target(i, line):
values = line.split()
assert (
len(values) == self.cfg.num_classes
), f'expected num_classes={self.cfg.num_classes} regression target values on line {i}, found: "{line}"'
return [float(x) for x in values]
with open(label_path) as h:
dataset.update(target=RawLabelDataset([
parse_regression_target(i, line.strip())
for i, line in enumerate(h.readlines())
]))
nested_dataset = NestedDictionaryDataset(
dataset,
sizes=[src_tokens.sizes],
)
if self.cfg.no_shuffle:
dataset = nested_dataset
else:
dataset = SortDataset(
nested_dataset,
# shuffle
sort_order=[shuffle],
)
logger.info("Loaded {0} with #samples: {1}".format(
split, len(dataset)))
self.datasets[split] = dataset
return self.datasets[split]
def load_dataset(self, split, epoch=1, combine=False, **kwargs):
"""Load a given dataset split.
Args:
split (str): name of the split (e.g., train, valid, test)
"""
paths = utils.split_paths(self.args.data)
assert len(paths) > 0
data_path = paths[(epoch - 1) % len(paths)]
split_path = os.path.join(data_path, split)
dataset = data_utils.load_indexed_dataset(
split_path,
self.source_dictionary,
self.args.dataset_impl,
combine=combine,
)
if dataset is None:
raise FileNotFoundError("Dataset not found: {} ({})".format(
split, split_path))
dataset = maybe_shorten_dataset(
dataset,
split,
self.args.shorten_data_split_list,
self.args.shorten_method,
self.args.tokens_per_sample,
self.args.seed,
)
# create continuous blocks of tokens
dataset = TokenBlockDataset(
dataset,
dataset.sizes,
self.args.tokens_per_sample - 1, # one less for <s>
pad=self.source_dictionary.pad(),
eos=self.source_dictionary.eos(),
break_mode=self.args.sample_break_mode,
)
logger.info("loaded {} blocks from: {}".format(len(dataset),
split_path))
# prepend beginning-of-sentence token (<s>, equiv. to [CLS] in BERT)
dataset = PrependTokenDataset(dataset, self.source_dictionary.bos())
# create masked input and targets
mask_whole_words = (get_whole_word_mask(self.args,
self.source_dictionary)
if self.args.mask_whole_words else None)
src_dataset, tgt_dataset = MaskTokensDataset.apply_mask(
dataset,
self.source_dictionary,
pad_idx=self.source_dictionary.pad(),
mask_idx=self.mask_idx,
seed=self.args.seed,
mask_prob=self.args.mask_prob,
leave_unmasked_prob=self.args.leave_unmasked_prob,
random_token_prob=self.args.random_token_prob,
freq_weighted_replacement=self.args.freq_weighted_replacement,
mask_whole_words=mask_whole_words,
mask_multiple_length=self.args.mask_multiple_length,
mask_stdev=self.args.mask_stdev,
fix_mask_len=self.args.fix_mask_len)
with data_utils.numpy_seed(self.args.seed):
shuffle = np.random.permutation(len(src_dataset))
self.datasets[split] = SortDataset(
NestedDictionaryDataset(
{
"id":
IdDataset(),
"net_input": {
"src_tokens":
RightPadDataset(
src_dataset,
pad_idx=self.source_dictionary.pad(),
),
"src_lengths":
NumelDataset(src_dataset, reduce=False),
},
"target":
RightPadDataset(
tgt_dataset,
pad_idx=self.source_dictionary.pad(),
),
"nsentences":
NumSamplesDataset(),
"ntokens":
NumelDataset(src_dataset, reduce=True),
},
sizes=[src_dataset.sizes],
),
sort_order=[
shuffle,
src_dataset.sizes,
],
)
def load_dataset(self, split, combine=False, **kwargs):
"""Load a given dataset split (e.g., train, valid, test)."""
def get_path(type, split):
return os.path.join(self.args.data, type, split)
def make_dataset(type, dictionary):
split_path = get_path(type, split)
dataset = data_utils.load_indexed_dataset(
split_path,
dictionary,
self.args.dataset_impl,
combine=combine,
)
return dataset
input0 = make_dataset('input0', self.source_dictionary)
assert input0 is not None, 'could not find dataset: {}'.format(
get_path('input0', split))
input1 = make_dataset('input1', self.source_dictionary)
assert input1 is not None, 'could not find dataset: {}'.format(
get_path('input1', split))
assert len(input0) == len(input1), 'input pair different length'
if self.args.init_token is not None:
input0 = PrependTokenDataset(input0, self.args.init_token)
input1 = PrependTokenDataset(input1, self.args.init_token)
with data_utils.numpy_seed(self.args.seed):
shuffle = np.random.permutation(len(input0))
if self.args.truncate_sequence:
input0 = TruncateDataset(input0, self.args.max_positions)
input1 = TruncateDataset(input1, self.args.max_positions)
dataset = {
'id': IdDataset(),
'net_input0': {
'src_tokens':
RightPadDataset(
input0,
pad_idx=self.source_dictionary.pad(),
),
'src_lengths':
NumelDataset(input0, reduce=False),
},
'net_input1': {
'src_tokens':
RightPadDataset(
input1,
pad_idx=self.source_dictionary.pad(),
),
'src_lengths':
NumelDataset(input1, reduce=False),
},
'nsentences': NumSamplesDataset(),
'ntokens0': NumelDataset(input0, reduce=True),
'ntokens1': NumelDataset(input1, reduce=True),
}
label_path = "{0}.label".format(get_path('label', split))
if os.path.exists(label_path):
dataset.update(target=RawLabelDataset(
[float(x.strip()) for x in open(label_path).readlines()]))
nested_dataset = NestedDictionaryDataset(
dataset,
sizes=[np.maximum(input0.sizes, input1.sizes)],
)
if self.args.no_shuffle:
dataset = nested_dataset
else:
dataset = SortDataset(
nested_dataset,
# shuffle
sort_order=[shuffle],
)
logger.info("Loaded {0} with #samples: {1}".format(
split, len(dataset)))
self.datasets[split] = dataset
return self.datasets[split]
def load_dataset(self, split, epoch=1, combine=False, **kwargs):
"""Load a given dataset split.
Args:
split (str): name of the split (e.g., train, valid, test)
"""
paths = utils.split_paths(self.args.data)
assert len(paths) > 0
src_tokens = {}
tgt_tokens = {}
tgt_values = {}
for field in configs.fields:
split_path = os.path.join(self.args.data, field, split)
dataset = data_utils.load_indexed_dataset(
split_path,
self.source_dictionary[field],
self.args.dataset_impl,
combine=combine,
)
if dataset is None:
raise FileNotFoundError(
"Dataset not found: {} ({})".format(split, split_path)
)
dataset = maybe_shorten_dataset(
dataset,
split,
self.args.shorten_data_split_list,
self.args.shorten_method,
self.args.tokens_per_sample,
self.args.seed,
)
# create continuous blocks of tokens
dataset = TokenBlockDataset(
dataset,
dataset.sizes,
self.args.tokens_per_sample - 1, # one less for <s>
pad=self.source_dictionary[field].pad(),
eos=self.source_dictionary[field].eos(),
break_mode=self.args.sample_break_mode,
)
logger.info("loaded {} blocks from: {}".format(len(dataset), split_path))
# prepend beginning-of-sentence token (<s>, equiv. to [CLS] in BERT)
dataset = PrependTokenDataset(dataset, self.source_dictionary[field].bos())
if field == configs.static_field:
src_dataset_code, tgt_dataset_code = MaskTokensDataset.apply_mask(
dataset,
self.source_dictionary[field],
pad_idx=self.source_dictionary[field].pad(),
mask_idx=self.mask_idx_dict[field],
seed=self.args.seed,
mask_prob=self.args.mask_prob,
leave_unmasked_prob=self.args.leave_unmasked_prob,
random_token_prob=self.args.random_token_prob,
freq_weighted_replacement=self.args.freq_weighted_replacement,
)
src_tokens[field] = RightPadDataset(
src_dataset_code,
pad_idx=self.source_dictionary[field].pad()
)
tgt_tokens[field] = RightPadDataset(
tgt_dataset_code,
pad_idx=self.source_dictionary[field].pad()
)
elif field in configs.byte_fields:
src_dataset_value, tgt_dataset_value = MaskValuesDataset.apply_mask(
dataset,
self.source_dictionary[field],
pad_idx=self.source_dictionary[field].pad(),
mask_idx=self.mask_idx_dict[field],
seed=self.args.seed,
mask_prob=self.args.mask_prob,
leave_unmasked_prob=self.args.leave_unmasked_prob,
random_token_prob=self.args.random_token_prob,
freq_weighted_replacement=self.args.freq_weighted_replacement,
)
src_tokens[field] = RightPadDataset(
src_dataset_value,
pad_idx=self.source_dictionary[field].pad()
)
# dummy tokens are treated as 1
# TODO: assert there should not be any dummy tokens here
tgt_values[field] = BytevalueDataset(tgt_dataset_value, self.source_dictionary[field])
else:
src_tokens[field] = RightPadDataset(
dataset,
pad_idx=self.source_dictionary[field].pad()
)
with data_utils.numpy_seed(self.args.seed):
shuffle = np.random.permutation(len(src_dataset_code))
self.datasets[split] = SortDataset(
NestedDictionaryDataset(
{
"id": IdDataset(),
"net_input": {
"src_tokens": src_tokens,
"src_lengths": NumelDataset(src_dataset_code, reduce=False),
},
"target": {
"tgt_tokens": tgt_tokens,
"tgt_values": tgt_values
},
"nsentences": NumSamplesDataset(),
"ntokens": NumelDataset(src_dataset_code, reduce=True),
},
sizes=[src_dataset_code.sizes],
),
sort_order=[
shuffle,
src_dataset_code.sizes,
],
)
def load_dataset(self, split, epoch=1, combine=False, data_path=None, return_only=False, **kwargs):
"""Load a given dataset split.
Args:
split (str): name of the split (e.g., train, valid, test)
"""
def binarize(s, append_bos=False):
if self.bpe is not None:
s = self.bpe.encode(s)
tokens = self.vocab.encode_line(
s, append_eos=True, add_if_not_exist=False,
).long()
if append_bos and self.args.init_token is not None:
tokens = torch.cat([tokens.new([self.args.init_token]), tokens])
return tokens
if data_path is None:
data_path = os.path.join(self.args.data, split + '.jsonl')
if not os.path.exists(data_path):
raise FileNotFoundError('Cannot find data: {}'.format(data_path))
src_tokens = [[] for i in range(self.args.num_classes)]
src_lengths = [[] for i in range(self.args.num_classes)]
labels = []
with open(data_path) as h:
for line in h:
example = json.loads(line.strip())
if 'answerKey' in example:
label = ord(example['answerKey']) - ord('A')
labels.append(label)
question = example['question']['stem']
assert len(example['question']['choices']) == self.args.num_classes
# format: `<s> Q: Where would I not want a fox? </s> A: hen house </s>`
question = 'Q: ' + question
question_toks = binarize(question, append_bos=True)
for i, choice in enumerate(example['question']['choices']):
src = 'A: ' + choice['text']
src_bin = torch.cat([question_toks, binarize(src)])
src_tokens[i].append(src_bin)
src_lengths[i].append(len(src_bin))
assert all(len(src_tokens[0]) == len(src_tokens[i]) for i in range(self.args.num_classes))
assert len(src_tokens[0]) == len(src_lengths[0])
assert len(labels) == 0 or len(labels) == len(src_tokens[0])
for i in range(self.args.num_classes):
src_lengths[i] = np.array(src_lengths[i])
src_tokens[i] = ListDataset(src_tokens[i], src_lengths[i])
src_lengths[i] = ListDataset(src_lengths[i])
dataset = {
'id': IdDataset(),
'nsentences': NumSamplesDataset(),
'ntokens': NumelDataset(src_tokens[0], reduce=True),
}
for i in range(self.args.num_classes):
dataset.update({
'net_input{}'.format(i + 1): {
'src_tokens': RightPadDataset(
src_tokens[i],
pad_idx=self.source_dictionary.pad(),
),
'src_lengths': src_lengths[i],
}
})
if len(labels) > 0:
dataset.update({'target': RawLabelDataset(labels)})
dataset = NestedDictionaryDataset(
dataset,
sizes=[np.maximum.reduce([src_token.sizes for src_token in src_tokens])],
)
with data_utils.numpy_seed(self.args.seed):
dataset = SortDataset(
dataset,
# shuffle
sort_order=[np.random.permutation(len(dataset))],
)
print('| Loaded {} with {} samples'.format(split, len(dataset)))
self.datasets[split] = dataset
return self.datasets[split]
def load_dataset(self, split, combine=False, **kwargs):
"""Load a given dataset split (e.g., train, valid, test)."""
inputs_path = Path(self.args.data) / "{split}".format(split=split)
src_tokens = data_utils.load_indexed_dataset(
str(inputs_path),
self.source_dictionary,
self.args.dataset_impl,
combine=combine,
)
assert src_tokens is not None, "could not find dataset: {}".format(
inputs_path)
with data_utils.numpy_seed(self.args.seed):
shuffle = np.random.permutation(len(src_tokens))
src_tokens = PrependTokenDataset(src_tokens,
self.source_dictionary.bos())
targets_path = Path(self.args.data) / "{}.nonterm".format(split)
labelled_spans = data_utils.load_indexed_dataset(
str(targets_path),
self.label_dictionary,
self.args.dataset_impl,
combine=combine,
)
assert labelled_spans is not None, "could not find labels: {}".format(
targets_path)
raise NotImplementedError
target_spans = LabelledSpanDataset(labelled_spans, return_spans=True)
labels = LabelledSpanDataset(labelled_spans, return_spans=False)
# all possible word spans in each sequence
word_spans = WordSpanDataset(src_tokens, self.source_dictionary,
self.is_word_initial)
all_spans = ProductSpanDataset(word_spans)
dataset = {
"id":
IdDataset(),
"net_input": {
"src_tokens":
RightPadDataset(src_tokens,
pad_idx=self.source_dictionary.pad()),
"nsrc_tokens":
NumelDataset(src_tokens),
"src_spans":
RightPadDataset(all_spans,
pad_idx=self.label_dictionary.pad()),
"nsrc_spans":
NumSpanDataset(all_spans),
},
"targets":
RightPadDataset(labels, pad_idx=self.label_dictionary.pad()),
"target_spans":
RightPadDataset(target_spans, pad_idx=self.label_dictionary.pad()),
"ntargets":
NumelDataset(labels),
"nsentences":
NumSamplesDataset(),
"ntokens":
NumelDataset(src_tokens, reduce=True),
"nwords":
NumWordsDataset(src_tokens, self.dictionary, self.is_word_initial),
"word_spans":
RightPadDataset(word_spans, pad_idx=self.label_dictionary.pad()),
}
nested_dataset = NestedDictionaryDatasetFix(dataset,
sizes=[src_tokens.sizes])
if self.args.no_shuffle:
dataset = nested_dataset
else:
dataset = SortDataset(nested_dataset, sort_order=[shuffle])
logger.info("Loaded {0} with #samples: {1}".format(
split, len(dataset)))
self.datasets[split] = dataset
return self.datasets[split]
def load_dataset(self, split, combine=False, **kwargs):
"""Load a given dataset split (e.g., train, valid, test)."""
def get_path(type, split):
return os.path.join(self.args.data, type, split)
def make_dataset(type, dictionary):
split_path = get_path(type, split)
dataset = data_utils.load_indexed_dataset(
split_path,
self.source_dictionary,
self.args.dataset_impl,
combine=combine,
)
return dataset
# inputs are loaded similarly to sentence_prediction
input0 = make_dataset("input0", self.source_dictionary) # question
input1 = make_dataset("input1", self.source_dictionary) # context
# src_tokens: <init_token> input0 <separator_token> input1 <eos_token>
if self.args.init_token is not None:
input0 = PrependTokenDataset(input0, self.args.init_token)
if self.args.separator_token is not None:
input1 = PrependTokenDataset(input1, self.args.separator_token)
if self.args.max_context_length is not None:
input1 = TruncateDataset(input1, self.args.max_option_length)
src_tokens = ConcatSentencesDataset(input0, input1)
if self.args.truncate_sequence:
src_tokens = TruncateDataset(src_tokens, self.args.max_positions)
with data_utils.numpy_seed(self.args.seed):
shuffle = np.random.permutation(len(src_tokens))
dataset = {
"id": IdDataset(),
"net_input": {
"src_tokens":
RightPadDataset(
src_tokens,
pad_idx=self.source_dictionary.pad(),
),
"src_lengths":
NumelDataset(src_tokens, reduce=False),
"input0_lengths":
NumelDataset(
input0, reduce=False
), # question length (init_token possibly included)
},
"nsentences": NumSamplesDataset(),
"ntokens": NumelDataset(src_tokens, reduce=True),
}
# labels (spans) are loaded similarly to sentence_ranking
label_path = "{}.label".format(get_path("label", split))
def _process_label(positions, input0_length, truncate_sequence,
max_positions):
"""Process a span [start:end] to the input range.
After processing, tokens can be accessed by tokens[start:end+1].
TODO: change inputs to reflect this change in the first place.
"""
start, end = [
pos + input0_length + (self.args.separator_token is not None)
for pos in positions
]
end -= 1 # [0, 511]
if truncate_sequence:
if start >= max_positions:
start, end = max_positions - 1, max_positions - 1 # not predictable
elif end >= max_positions:
end = max_positions - 1
return start, end
if os.path.exists(label_path):
with open(label_path) as h:
dataset.update(target=RawLabelDataset([
_process_label(
tuple(int(pos) for pos in x.split()),
dataset["net_input"]["input0_lengths"][i],
self.args.truncate_sequence,
self.max_positions(),
) for i, x in enumerate(
h.readlines()) # (start_position, end_position)
]))
nested_dataset = NestedDictionaryDataset(
dataset,
sizes=[src_tokens.sizes],
)
if self.args.no_shuffle:
dataset = nested_dataset
else:
dataset = SortDataset(
nested_dataset,
# shuffle
sort_order=[shuffle],
)
logger.info("Loaded {0} with #samples: {1}".format(
split, len(dataset)))
self.datasets[split] = dataset
return self.datasets[split]
def load_dataset(self, split, combine=False, **kwargs):
"""Load a given dataset split (e.g., train, valid, test)."""
def get_path(type, split):
return os.path.join(self.args.data, type, split)
def make_dataset(type, dictionary):
split_path = get_path(type, split)
dataset = data_utils.load_indexed_dataset(
split_path,
dictionary,
self.args.dataset_impl,
combine=combine,
)
return dataset
src_tokens = make_dataset('data', self.source_dictionary)
if self.args.init_token is not None:
src_tokens = PrependTokenDataset(src_tokens, self.args.init_token)
with data_utils.numpy_seed(self.args.seed):
shuffle = np.random.permutation(len(src_tokens))
if self.args.truncate_sequence:
src_tokens = TruncateDataset(src_tokens, self.args.max_positions)
dataset = {
'id': IdDataset(),
'net_input': {
'src_tokens':
RightPadDataset(
src_tokens,
pad_idx=self.source_dictionary.pad(),
),
'src_lengths':
NumelDataset(src_tokens, reduce=False),
},
'nsentences': NumSamplesDataset(),
'ntokens': NumelDataset(src_tokens, reduce=True),
}
label_dataset = make_dataset('label', self.label_dictionary)
if label_dataset is not None:
dataset.update(target=OffsetTokensDataset(
StripTokenDataset(
label_dataset,
id_to_strip=self.label_dictionary.eos(),
),
offset=-self.label_dictionary.nspecial,
))
nested_dataset = NestedDictionaryDataset(
dataset,
sizes=[src_tokens.sizes],
)
if self.args.no_shuffle:
dataset = nested_dataset
else:
dataset = SortDataset(
nested_dataset,
# shuffle
sort_order=[shuffle],
)
logger.info("Loaded {0} with #samples: {1}".format(
split, len(dataset)))
self.datasets[split] = dataset
return self.datasets[split]
def load_dataset(self, split, epoch=1, combine=False, **kwargs):
"""Load a given dataset split.
Args:
split (str): name of the split (e.g., train, valid, test)
"""
dataset = self._load_dataset_split(split, epoch, combine)
# create masked input and targets
mask_whole_words = (get_whole_word_mask(self.args,
self.source_dictionary)
if self.cfg.mask_whole_words else None)
src_dataset, tgt_dataset = MaskTokensDataset.apply_mask(
dataset,
self.source_dictionary,
pad_idx=self.source_dictionary.pad(),
mask_idx=self.mask_idx,
seed=self.cfg.seed,
mask_prob=self.cfg.mask_prob,
leave_unmasked_prob=self.cfg.leave_unmasked_prob,
random_token_prob=self.cfg.random_token_prob,
freq_weighted_replacement=self.cfg.freq_weighted_replacement,
mask_whole_words=mask_whole_words,
mask_multiple_length=self.cfg.mask_multiple_length,
mask_stdev=self.cfg.mask_stdev,
)
with data_utils.numpy_seed(self.cfg.seed):
shuffle = np.random.permutation(len(src_dataset))
target_dataset = RightPadDataset(
tgt_dataset,
pad_idx=self.source_dictionary.pad(),
)
input_dict = {
"src_tokens":
RightPadDataset(
src_dataset,
pad_idx=self.source_dictionary.pad(),
),
"src_lengths":
NumelDataset(src_dataset, reduce=False),
}
if self.cfg.include_target_tokens:
input_dict["target_tokens"] = target_dataset
self.datasets[split] = SortDataset(
NestedDictionaryDataset(
{
"id": IdDataset(),
"net_input": input_dict,
"target": target_dataset,
"nsentences": NumSamplesDataset(),
"ntokens": NumelDataset(src_dataset, reduce=True),
},
sizes=[src_dataset.sizes],
),
sort_order=[
shuffle,
src_dataset.sizes,
],
)
def load_dataset(self, split, epoch=1, combine=False, **kwargs):
"""Load a given dataset split.
Args:
split (str): name of the split (e.g., train, valid, test)
"""
paths = utils.split_paths(self.cfg.data)
assert len(paths) > 0
if split != getattr(self.cfg, "train_subset", None):
# if not training data set, use the first shard for valid and test
paths = paths[:1]
data_path = paths[(epoch - 1) % len(paths)]
# infer langcode
src, tgt = self.cfg.source_lang, self.cfg.target_lang
prefix = os.path.join(data_path, '{}.{}-{}.'.format(split, src, tgt))
src_dataset = data_utils.load_indexed_dataset(prefix + src,
self.src_dict,
self.cfg.dataset_impl)
tag_dataset = data_utils.load_indexed_dataset(prefix + tgt,
self.tag_dict,
self.cfg.dataset_impl)
src_dataset = StripTokenDataset(
src_dataset, id_to_strip=self.source_dictionary.eos())
tag_dataset = StripTokenDataset(tag_dataset,
id_to_strip=self.tag_dictionary.eos())
tag_pad = self.source_dictionary.pad()
tag_offset = tag_pad + 1
dataset = {
'id':
IdDataset(),
'net_input': {
'src_tokens':
RightPadDataset(src_dataset,
pad_idx=self.source_dictionary.pad()),
'src_lengths':
NumelDataset(src_dataset, reduce=False),
},
'nsentences':
NumSamplesDataset(),
'ntokens':
NumelDataset(src_dataset, reduce=True),
'target':
RightPadDataset(
OffsetTokensDataset(tag_dataset,
offset=-self.tag_dictionary.nspecial +
tag_offset),
pad_idx=tag_pad,
),
}
dataset = NestedDictionaryDataset(
dataset,
sizes=[src_dataset.sizes],
)
logger.info(
str([self.src_dict[k]
for k in dataset[0]['net_input.src_tokens']]))
logger.info(
str([
self.tag_dict[k + self.tag_dictionary.nspecial - tag_offset]
for k in dataset[0]['target']
]))
self.datasets[split] = dataset
def load_dataset(self,
split,
epoch=0,
combine=False,
data_path=None,
return_only=False,
**kwargs):
"""Load a given dataset split.
Args:
split (str): name of the split (e.g., train, valid, test)
"""
def binarize(s, append_bos=False):
if self.bpe is not None:
s = self.bpe.encode(s)
tokens = self.vocab.encode_line(
s,
append_eos=True,
add_if_not_exist=False,
).long()
if append_bos and self.args.init_token is not None:
tokens = torch.cat(
[tokens.new([self.args.init_token]), tokens])
return tokens
# self.data_path_table={'train_input':os.path.join(self.args.data,'Training Data','subtaskA_data_all.csv'),\
# 'train_answer':os.path.join(self.args.data,'Training Data','subtaskA_answers_all.csv'),\
# 'valid_input':os.path.join(self.args.data,'Trial Data','taskA_trial_data.csv'),\
# 'valid_answer':os.path.join(self.args.data,'Trial Data','taskA_trial_answer.csv')\
# }
# self.data_path_table={'train_input':os.path.join(self.args.data,'trainval','subtaskA_data_all.csv'),\
# 'train_answer':os.path.join(self.args.data,'trainval','subtaskA_answers_all.csv'),\
# 'valid_input':os.path.join(self.args.data,'Dev Data','subtaskA_dev_data.csv'),\
# 'valid_answer':os.path.join(self.args.data,'Dev Data','subtaskA_gold_answers.csv')\
# }
self.data_path_table={'train_input':os.path.join(self.args.data,'trainvaldev','subtaskA_data_all_plusplus.csv'),\
'train_answer':os.path.join(self.args.data,'trainvaldev','subtaskA_answers_all.csv'),\
'valid_input':os.path.join(self.args.data,'Dev Data','subtaskA_dev_data_plusplus.csv'),\
'valid_answer':os.path.join(self.args.data,'Dev Data','subtaskA_gold_answers.csv')\
}
# self.data_path_table={'train_input':os.path.join(self.args.data,'subtaskA_data_all.csv'),\
# 'train_answer':os.path.join(self.args.data,'subtaskA_answers_all.csv'),\
# 'valid_input':os.path.join(self.args.data,'taskA_trial_data.csv'),\
# 'valid_answer':os.path.join(self.args.data,'taskA_trial_answer.csv')\
# }
data_path_input = self.data_path_table[split + '_input']
data_path_answer = self.data_path_table[split + '_answer']
if not os.path.exists(data_path_input):
raise FileNotFoundError(
'Cannot find data: {}'.format(data_path_input))
if not os.path.exists(data_path_answer):
raise FileNotFoundError(
'Cannot find data: {}'.format(data_path_answer))
src_tokens = [[] for i in range(self.args.num_classes)]
src_lengths = [[] for i in range(self.args.num_classes)]
src_ids = []
labels = []
label_ids = []
with open(data_path_input) as f:
reader = csv.reader(f)
for row in islice(reader, 1, None):
src_ids.append(row[0])
for i in range(self.args.num_classes):
src = row[i + 1]
evidence = row[i + 3]
if src.isupper():
src = src.capitalize()
src = src + ' Context: ' + evidence
src_bin = binarize(src, append_bos=True)
src_tokens[i].append(src_bin)
src_lengths[i].append(len(src_bin))
assert all(
len(src_tokens[0]) == len(src_tokens[i])
for i in range(self.args.num_classes))
assert len(src_tokens[0]) == len(src_lengths[0])
with open(data_path_answer) as f:
reader = csv.reader(f)
for row in reader:
label_ids.append(row[0])
label = 1 - int(row[1])
labels.append(label)
assert len(labels) == 0 or len(labels) == len(src_tokens[0])
assert all(src_ids[i] == label_ids[i] for i in range(len(src_ids)))
for i in range(self.args.num_classes):
src_lengths[i] = np.array(src_lengths[i])
src_tokens[i] = ListDataset(src_tokens[i], src_lengths[i])
src_lengths[i] = ListDataset(src_lengths[i])
dataset = {
'id': IdDataset(),
'nsentences': NumSamplesDataset(),
'ntokens': NumelDataset(src_tokens[0], reduce=True),
}
for i in range(self.args.num_classes):
dataset.update({
'net_input{}'.format(i + 1): {
'src_tokens':
RightPadDataset(
src_tokens[i],
pad_idx=self.source_dictionary.pad(),
),
'src_lengths':
src_lengths[i],
}
})
if len(labels) > 0:
dataset.update({'target': RawLabelDataset(labels)})
dataset = NestedDictionaryDataset(
dataset,
sizes=[
np.maximum.reduce(
[src_token.sizes for src_token in src_tokens])
],
)
with data_utils.numpy_seed(self.args.seed):
dataset = SortDataset(
dataset,
# shuffle
sort_order=[np.random.permutation(len(dataset))],
)
print('| Loaded {} with {} samples'.format(split, len(dataset)))
self.datasets[split] = dataset
return self.datasets[split]
def build_dataset_for_inference(self, src_tokens, src_lengths):
return RightPadDataset(src_tokens,
pad_idx=self.source_dictionary.pad())
def load_dataset(self, split, combine=False, **kwargs):
"""Load a given dataset split (e.g., train, valid, test)."""
def get_path(type, split):
return os.path.join(self.args.data, type, split)
def make_dataset(type, dictionary):
split_path = get_path(type, split)
dataset = data_utils.load_indexed_dataset(
split_path,
dictionary,
self.args.dataset_impl,
combine=combine,
)
return dataset
# input0 is source, input1 is synthetic target, input2 is reference
input0 = make_dataset(self.args.input0, self.source_dictionary)
assert input0 is not None, 'could not find dataset: {}'.format(
get_path(type, split))
input1 = make_dataset(self.args.input1, self.source_dictionary)
if self.args.init_token is not None:
input0 = PrependTokenDataset(input0, self.args.init_token)
if self.args.input2 is not None:
input2 = make_dataset(self.args.input2, self.source_dictionary)
if self.args.input2 is not None and self.add_ref_prob > 0 and split != 'valid':
input3 = PrependTokenDataset(input2, self.args.separator_token)
else:
input3 = None
if input1 is None:
src_tokens = input0
else:
if self.args.separator_token is not None:
input1 = PrependTokenDataset(input1, self.args.separator_token)
if self.args.input2 is not None and self.add_ref_prob > 0. and split != 'valid':
src_tokens = ConcatSentencesDataset(
input0,
input3,
input1,
add_ref_prob=self.add_ref_prob,
drop_ref_rate=self.args.dropout_ref,
pad_idx=self.source_dictionary.pad(),
eos_idx=self.source_dictionary.eos(),
bos_idx=self.source_dictionary.bos())
else:
src_tokens = ConcatSentencesDataset(input0, input1)
with data_utils.numpy_seed(self.args.seed):
shuffle = np.random.permutation(len(src_tokens))
if self.args.truncate_sequence:
src_tokens = TruncateDataset(src_tokens, self.args.max_positions)
if self.args.input2 is not None and self.args.add_tran_loss:
# create masked input and targets
mask_whole_words = get_whole_word_mask(self.args, self.source_dictionary) \
if self.args.mask_whole_words else None
ref_dataset, ref_target_dataset = MaskTokensDataset.apply_mask(
input2,
self.source_dictionary,
pad_idx=self.source_dictionary.pad(),
mask_idx=self.mask_idx,
seed=self.args.seed,
mask_prob=self.args.mask_prob,
leave_unmasked_prob=self.args.leave_unmasked_prob,
random_token_prob=self.args.random_token_prob,
freq_weighted_replacement=self.args.freq_weighted_replacement,
mask_whole_words=mask_whole_words,
)
if self.args.separator_token is not None:
input2 = PrependTokenDataset(ref_dataset,
self.args.separator_token)
parallel_src_tokens = ConcatSentencesDataset(input0, input2)
if self.args.truncate_sequence:
parallel_src_tokens = TruncateDataset(parallel_src_tokens,
self.args.max_positions)
dataset = {
'id': IdDataset(),
'net_input': {
'src_tokens':
RightPadDataset(
src_tokens,
pad_idx=self.source_dictionary.pad(),
),
'src_lengths':
NumelDataset(src_tokens, reduce=False),
},
'nsentences': NumSamplesDataset(),
'ntokens': NumelDataset(src_tokens, reduce=True),
}
if self.args.input2 is not None and self.args.add_tran_loss:
dataset['net_input']['parallel_src_tokens'] = RightPadDataset(
parallel_src_tokens,
pad_idx=self.source_dictionary.pad(),
)
if self.args.add_prev_output_tokens:
prev_tokens_dataset = RightPadDataset(
RollDataset(src_tokens, 1),
pad_idx=self.dictionary.pad(),
)
dataset['net_input'].update(
prev_output_tokens=prev_tokens_dataset, )
if not self.args.regression_target:
label_dataset = make_dataset('label', self.label_dictionary)
if label_dataset is not None:
dataset.update(target=OffsetTokensDataset(
StripTokenDataset(
label_dataset,
id_to_strip=self.label_dictionary.eos(),
),
offset=-self.label_dictionary.nspecial,
))
if self.args.input2 is not None and self.args.add_tran_loss:
# used as translation target when calculating loss
dataset.update(parallel_target=RightPadDataset(
ref_target_dataset,
pad_idx=self.source_dictionary.pad(),
))
else:
label_path = "{0}.label".format(get_path('label', split))
if os.path.exists(label_path):
def parse_regression_target(i, line):
values = line.split()
assert len(values) == self.args.num_classes, \
f'expected num_classes={self.args.num_classes} regression target values on line {i}, found: "{line}"'
return [float(x) for x in values]
dataset.update(target=RawLabelDataset([
parse_regression_target(i, line.strip())
for i, line in enumerate(open(label_path).readlines())
]))
nested_dataset = NestedDictionaryDataset(
dataset,
sizes=[src_tokens.sizes],
all_sizes=src_tokens.all_sizes
if self.args.add_target_num_tokens else None,
padding_idx=self.source_dictionary.pad(),
add_ref_prob=self.add_ref_prob if split != 'valid' else 0.,
)
if self.args.no_shuffle:
dataset = nested_dataset
else:
dataset = SortDataset(
nested_dataset,
# shuffle
sort_order=[shuffle],
)
logger.info("Loaded {0} with #samples: {1}".format(
split, len(dataset)))
self.datasets[split] = dataset
return self.datasets[split]
def load_dataset(self, split, combine=False, **kwargs):
"""Load a given dataset split (e.g., train, valid, test)."""
def get_path(type, split):
return os.path.join(self.args.data, type, split)
def make_dataset(type, dictionary):
split_path = get_path(type, split)
dataset = data_utils.load_indexed_dataset(
split_path,
self.source_dictionary,
self.args.dataset_impl,
combine=combine,
)
return dataset
input0 = make_dataset('input0', self.source_dictionary)
input_options = [
make_dataset('input{idx}'.format(idx=idx + 1),
self.source_dictionary)
for idx in range(self.args.num_classes)
]
if self.args.separator_token is not None:
input0 = PrependTokenDataset(input0, self.args.separator_token)
src_tokens = []
for input_option in input_options:
if self.args.init_token is not None:
input_option = PrependTokenDataset(input_option,
self.args.init_token)
if self.args.max_option_length is not None:
input_option = TruncateDataset(input_option,
self.args.max_option_length)
src_token = ConcatSentencesDataset(input_option, input0)
src_token = maybe_shorten_dataset(
src_token,
split,
self.args.shorten_data_split_whitelist,
self.args.shorten_method,
self.args.max_positions,
self.args.seed,
)
src_tokens.append(src_token)
with data_utils.numpy_seed(self.args.seed):
shuffle = np.random.permutation(len(src_tokens[0]))
dataset = {
'id': IdDataset(),
'nsentences': NumSamplesDataset(),
'ntokens': NumelDataset(src_tokens[0], reduce=True),
}
for src_token_idx in range(len(src_tokens)):
dataset.update({
'net_input{idx}'.format(idx=src_token_idx + 1): {
'src_tokens':
RightPadDataset(
src_tokens[src_token_idx],
pad_idx=self.source_dictionary.pad(),
),
'src_lengths':
NumelDataset(src_tokens[src_token_idx], reduce=False),
}
})
label_path = '{}.label'.format(get_path('label', split))
if os.path.exists(label_path):
with open(label_path) as h:
dataset.update(target=RawLabelDataset(
[int(x.strip()) for x in h.readlines()]))
nested_dataset = NestedDictionaryDataset(
dataset,
sizes=[
np.maximum.reduce(
[src_token.sizes for src_token in src_tokens])
],
)
if self.args.no_shuffle:
dataset = nested_dataset
else:
dataset = SortDataset(
nested_dataset,
# shuffle
sort_order=[shuffle],
)
logger.info("Loaded {0} with #samples: {1}".format(
split, len(dataset)))
self.datasets[split] = dataset
return self.datasets[split]
def load_dataset(self, split, combine=False, **kwargs):
"""Load a given dataset split (e.g., train, valid, test)."""
###encoder 객체 생성
bpe_encoder = MultiprocessingEncoder(self.args.encoder_json,
self.args.vocab_bpe)
bpe_encoder.initializer()
###preprocess_coqa부르기
examples, features = get_CoQA_features(self.args,
bpe_encoder,
self.args.init_token,
self.args.separator_token,
self.dictionary.pad(),
split=split)
self.examples[split] = examples
self.features[split] = features
qas_idx = []
src_tokens = []
src_lengths = []
padding_mask = []
start_pos = []
end_pos = []
is_unk = []
is_yes = []
is_no = []
number = []
option = []
for feature in features:
src = torch.IntTensor(feature.input_tokens).long()
p_mask = torch.IntTensor(feature.p_mask).long()
src_tokens.append(src)
src_lengths.append(len(src))
padding_mask.append(p_mask)
qas_idx.append(feature.qas_id)
start_pos.append(feature.start_position)
end_pos.append(feature.end_position)
is_unk.append(feature.is_unk)
is_yes.append(feature.is_yes)
is_no.append(feature.is_no)
number.append(feature.number)
option.append(feature.option)
src_tokens = ListDataset(src_tokens, src_lengths)
src_lengths = ListDataset(src_lengths)
dataset = {
"id": IdDataset(),
"nsentences": NumSamplesDataset(),
"ntokens": NumelDataset(src_tokens, reduce=True),
"qas_id": RawLabelDataset(qas_idx),
"net_input": {
"src_tokens":
RightPadDataset(src_tokens, pad_idx=self.dictionary.pad()),
"src_lengths":
src_lengths,
"start_position":
RawLabelDataset(start_pos),
"p_mask":
RightPadDataset(padding_mask, pad_idx=self.dictionary.pad()),
},
"start_position": RawLabelDataset(start_pos),
"end_position": RawLabelDataset(end_pos),
"is_unk": RawLabelDataset(is_unk),
"is_yes": RawLabelDataset(is_yes),
"is_no": RawLabelDataset(is_no),
"number": RawLabelDataset(number),
"option": RawLabelDataset(option),
}
dataset = NestedDictionaryDataset(
dataset,
sizes=[np.maximum.reduce([src_tokens.sizes])],
)
with data_utils.numpy_seed(self.args.seed):
dataset = SortDataset(
dataset,
sort_order=[np.random.permutation(len(dataset))],
)
print("| Loaded {} with {} samples".format(split, len(dataset)))
self.datasets[split] = dataset
return self.datasets[split]
def load_dataset(self, split, epoch=0, combine=False, **kwargs):
"""Load a given dataset split (e.g., train, valid, test)."""
if self.cfg.data.endswith("1"):
data_shard = (epoch - 1) % self.cfg.num_data_splits + 1
data_path = self.cfg.data[:-1] + str(data_shard)
else:
data_path = self.cfg.data
def get_path(type, data_split):
return os.path.join(data_path, str(type), data_split)
def make_dataset(type, dictionary, data_split, combine):
split_path = get_path(type, data_split)
dataset = data_utils.load_indexed_dataset(
split_path,
dictionary,
combine=combine,
)
return dataset
def load_split(data_split, metric):
input_src = None
if self.cfg.include_src:
input_src = make_dataset("input_src",
self.dictionary,
data_split,
combine=False)
assert input_src is not None, "could not find dataset: {}".format(
get_path("input_src", data_split))
input_tgt = make_dataset("input_tgt",
self.dictionary,
data_split,
combine=False)
assert input_tgt is not None, "could not find dataset: {}".format(
get_path("input_tgt", data_split))
label_path = f"{get_path(metric, data_split)}.{metric}"
assert os.path.exists(
label_path), f"could not find dataset: {label_path}"
np_labels = np.loadtxt(label_path)
if self.cfg.target_metric == "ter":
np_labels = -np_labels
label = RawLabelDataset(np_labels)
return input_src, input_tgt, label
src_datasets = []
tgt_datasets = []
label_datasets = []
if split == self.cfg.train_subset:
for k in itertools.count():
split_k = "train" + (str(k) if k > 0 else "")
prefix = os.path.join(data_path, "input_tgt", split_k)
if not indexed_dataset.dataset_exists(prefix, impl=None):
if k > 0:
break
else:
raise FileNotFoundError(f"Dataset not found: {prefix}")
input_src, input_tgt, label = load_split(
split_k, self.cfg.target_metric)
src_datasets.append(input_src)
tgt_datasets.append(input_tgt)
label_datasets.append(label)
else:
input_src, input_tgt, label = load_split(split,
self.cfg.target_metric)
src_datasets.append(input_src)
tgt_datasets.append(input_tgt)
label_datasets.append(label)
if len(tgt_datasets) == 1:
input_tgt, label = tgt_datasets[0], label_datasets[0]
if self.cfg.include_src:
input_src = src_datasets[0]
else:
input_tgt = ConcatDataset(tgt_datasets)
label = ConcatDataset(label_datasets)
if self.cfg.include_src:
input_src = ConcatDataset(src_datasets)
input_tgt = TruncateDataset(input_tgt, self.cfg.max_positions)
if self.cfg.include_src:
input_src = PrependTokenDataset(input_src, self.dictionary.bos())
input_src = TruncateDataset(input_src, self.cfg.max_positions)
src_lengths = NumelDataset(input_src, reduce=False)
src_tokens = ConcatSentencesDataset(input_src, input_tgt)
else:
src_tokens = PrependTokenDataset(input_tgt, self.dictionary.bos())
src_lengths = NumelDataset(src_tokens, reduce=False)
dataset = {
"id": IdDataset(),
"net_input": {
"src_tokens":
RightPadDataset(
src_tokens,
pad_idx=self.source_dictionary.pad(),
),
"src_lengths":
src_lengths,
},
"nsentences": NumSamplesDataset(),
"ntokens": NumelDataset(src_tokens, reduce=True),
"target": label,
}
dataset = NestedDictionaryDataset(
dataset,
sizes=[src_tokens.sizes],
)
assert len(dataset) % self.cfg.mt_beam == 0, (
"dataset size (%d) is not a multiple of beam size (%d)" %
(len(dataset), self.cfg.mt_beam))
# no need to shuffle valid/test sets
if not self.cfg.no_shuffle and split == self.cfg.train_subset:
# need to keep all hypothese together
start_idx = np.arange(0, len(dataset), self.cfg.mt_beam)
with data_utils.numpy_seed(self.cfg.seed + epoch):
np.random.shuffle(start_idx)
idx = np.arange(0, self.cfg.mt_beam)
shuffle = np.tile(idx, (len(start_idx), 1)).reshape(-1) + np.tile(
start_idx, (self.cfg.mt_beam, 1)).transpose().reshape(-1)
dataset = SortDataset(
dataset,
sort_order=[shuffle],
)
logger.info(f"Loaded {split} with #samples: {len(dataset)}")
self.datasets[split] = dataset
return self.datasets[split]
def load_dataset(self, split, combine=False, **kwargs):
"""Load a given dataset split (e.g., train, valid, test)."""
def get_path(key, split):
return os.path.join(self.args.data, key, split)
def make_dataset(key, dictionary):
split_path = get_path(key, split)
dataset = data_utils.load_indexed_dataset(
split_path,
dictionary,
self.args.dataset_impl,
combine=combine,
)
return dataset
input0 = make_dataset("input0", self.source_dictionary)
assert input0 is not None, "could not find dataset: {}".format(
get_path("input0", split))
input1 = make_dataset("input1", self.source_dictionary)
if self.args.init_token is not None:
input0 = PrependTokenDataset(input0, self.args.init_token)
if input1 is None:
src_tokens = input0
else:
if self.args.separator_token is not None:
input1 = PrependTokenDataset(input1, self.args.separator_token)
src_tokens = ConcatSentencesDataset(input0, input1)
with data_utils.numpy_seed(self.args.seed):
shuffle = np.random.permutation(len(src_tokens))
src_tokens = maybe_shorten_dataset(
src_tokens,
split,
self.args.shorten_data_split_list,
self.args.shorten_method,
self.args.max_positions,
self.args.seed,
)
dataset = {
"id": IdDataset(),
"net_input": {
"src_tokens":
RightPadDataset(
src_tokens,
pad_idx=self.source_dictionary.pad(),
),
"src_lengths":
NumelDataset(src_tokens, reduce=False),
},
"nsentences": NumSamplesDataset(),
"ntokens": NumelDataset(src_tokens, reduce=True),
}
if self.args.add_prev_output_tokens:
prev_tokens_dataset = RightPadDataset(
RollDataset(src_tokens, 1),
pad_idx=self.dictionary.pad(),
)
dataset["net_input"].update(
prev_output_tokens=prev_tokens_dataset, )
label_path = "{0}.npz".format(get_path("label", split))
if os.path.exists(label_path):
csr_matrix = load_npz(label_path)
dataset.update(target=CSRLabelDataset(csr_matrix))
nested_dataset = NestedDictionaryDataset(
dataset,
sizes=[src_tokens.sizes],
)
if self.args.no_shuffle:
dataset = nested_dataset
else:
dataset = SortDataset(
nested_dataset,
# shuffle
sort_order=[shuffle],
)
logger.info("Loaded {0} with #samples: {1}".format(
split, len(dataset)))
self.datasets[split] = dataset
return self.datasets[split]
def load_dataset(self, split, combine=False, **kwargs):
"""Load a given dataset split (e.g., train, valid, test)."""
def get_path(type, split):
return os.path.join(self.args.data, type, split)
def make_dataset(type, dictionary):
split_path = get_path(type, split)
dataset = data_utils.load_indexed_dataset(
split_path,
dictionary,
self.args.dataset_impl,
combine=combine,
)
return dataset
input0 = make_dataset('input0', self.source_dictionary)
assert input0 is not None, 'could not find dataset: {}'.format(
get_path(type, split))
input1 = make_dataset('input1', self.source_dictionary)
if self.args.init_token is not None:
input0 = PrependTokenDataset(input0, self.args.init_token)
if input1 is None:
src_tokens = input0
else:
if self.args.separator_token is not None:
input1 = PrependTokenDataset(input1, self.args.separator_token)
src_tokens = ConcatSentencesDataset(input0, input1)
with data_utils.numpy_seed(self.args.seed):
shuffle = np.random.permutation(len(src_tokens))
src_tokens = maybe_shorten_dataset(
src_tokens,
split,
self.args.shorten_data_split_list,
self.args.shorten_method,
self.args.max_positions,
self.args.seed,
)
dataset = {
'id': IdDataset(),
'net_input': {
'src_tokens':
RightPadDataset(
src_tokens,
pad_idx=self.source_dictionary.pad(),
),
'src_lengths':
NumelDataset(src_tokens, reduce=False),
},
'nsentences': NumSamplesDataset(),
'ntokens': NumelDataset(src_tokens, reduce=True),
}
if self.args.add_prev_output_tokens:
prev_tokens_dataset = RightPadDataset(
RollDataset(src_tokens, 1),
pad_idx=self.dictionary.pad(),
)
dataset['net_input'].update(
prev_output_tokens=prev_tokens_dataset, )
if not self.args.regression_target:
label_dataset = make_dataset('label', self.label_dictionary)
if label_dataset is not None:
dataset.update(target=OffsetTokensDataset(
StripTokenDataset(
label_dataset,
id_to_strip=self.label_dictionary.eos(),
),
offset=-self.label_dictionary.nspecial,
))
else:
label_path = "{0}.label".format(get_path('label', split))
if os.path.exists(label_path):
def parse_regression_target(i, line):
values = line.split()
assert len(values) == self.args.num_classes, \
f'expected num_classes={self.args.num_classes} regression target values on line {i}, found: "{line}"'
return [float(x) for x in values]
dataset.update(target=RawLabelDataset([
parse_regression_target(i, line.strip())
for i, line in enumerate(open(label_path).readlines())
]))
nested_dataset = NestedDictionaryDataset(
dataset,
sizes=[src_tokens.sizes],
)
if self.args.no_shuffle:
dataset = nested_dataset
else:
dataset = SortDataset(
nested_dataset,
# shuffle
sort_order=[shuffle],
)
logger.info("Loaded {0} with #samples: {1}".format(
split, len(dataset)))
self.datasets[split] = dataset
return self.datasets[split]
def load_dataset(self, split, epoch=1, combine=False, **kwargs):
"""Load a given dataset split.
Args:
split (str): name of the split (e.g., train, valid, test)
"""
paths = utils.split_paths(self.args.data)
assert len(paths) > 0
data_path = paths[(epoch - 1) % len(paths)]
split_path = os.path.join(data_path, split)
dataset = data_utils.load_indexed_dataset(
split_path,
self.source_dictionary,
self.args.dataset_impl,
combine=combine,
)
if dataset is None:
raise FileNotFoundError('Dataset not found: {} ({})'.format(split, split_path))
dataset = maybe_shorten_dataset(
dataset,
split,
self.args.shorten_data_split_list,
self.args.shorten_method,
self.args.tokens_per_sample,
self.args.seed,
)
# create continuous blocks of tokens
dataset = TokenBlockDataset(
dataset,
dataset.sizes,
self.args.tokens_per_sample,
pad=self.source_dictionary.pad(),
eos=self.source_dictionary.eos(),
break_mode=self.args.sample_break_mode,
)
logger.info('loaded {} blocks from: {}'.format(len(dataset), split_path))
# remove tail
dataset = RemoveTailDataset(dataset)
# create masked input and targets
mask_whole_words = get_whole_word_mask(self.args, self.source_dictionary) \
if self.args.mask_whole_words else None
src_dataset, tgt_dataset = MaskTokensDataset.apply_mask(
dataset,
self.source_dictionary,
pad_idx=self.source_dictionary.pad(),
mask_idx=self.mask_idx,
seed=self.args.seed,
mask_prob=self.args.mask_prob,
leave_unmasked_prob=self.args.leave_unmasked_prob,
random_token_prob=self.args.random_token_prob,
freq_weighted_replacement=self.args.freq_weighted_replacement,
mask_whole_words=mask_whole_words,
)
with data_utils.numpy_seed(self.args.seed + epoch):
shuffle = np.random.permutation(len(src_dataset))
self.datasets[split] = SortDataset(
NestedDictionaryDataset(
{
'id': IdDataset(),
'net_input': {
'src_tokens': RightPadDataset(
src_dataset,
pad_idx=self.source_dictionary.pad(),
),
'src_lengths': NumelDataset(src_dataset, reduce=False),
},
'target': RightPadDataset(
tgt_dataset,
pad_idx=self.source_dictionary.pad(),
),
'nsentences': NumSamplesDataset(),
'ntokens': NumelDataset(src_dataset, reduce=True),
},
sizes=[src_dataset.sizes],
),
sort_order=[
shuffle,
src_dataset.sizes,
],
)
def load_dataset(self, split, combine=False, **kwargs):
"""Load a given dataset split (e.g., train, valid, test)."""
def get_path(type, split):
return os.path.join(self.args.data, type, split)
def make_dataset(type, dictionary):
split_path = get_path(type, split)
dataset = data_utils.load_indexed_dataset(
split_path,
self.source_dictionary,
self.args.dataset_impl,
combine=combine,
)
return dataset
input0 = make_dataset('input0', self.source_dictionary)
assert input0 is not None, 'could not find dataset: {}'.format(
get_path(type, split))
input1 = make_dataset('input1', self.source_dictionary)
if self.args.init_token is not None:
input0 = PrependTokenDataset(input0, self.args.init_token)
if input1 is None:
src_tokens = input0
else:
if self.args.separator_token is not None:
input1 = PrependTokenDataset(input1, self.args.separator_token)
src_tokens = ConcatSentencesDataset(input0, input1)
with data_utils.numpy_seed(self.args.seed):
shuffle = np.random.permutation(len(src_tokens))
if self.args.truncate_sequence:
src_tokens = TruncateDataset(src_tokens, self.args.max_positions)
dataset = {
'id': IdDataset(),
'net_input': {
'src_tokens':
RightPadDataset(
src_tokens,
pad_idx=self.source_dictionary.pad(),
),
'src_lengths':
NumelDataset(src_tokens, reduce=False),
},
'nsentences': NumSamplesDataset(),
'ntokens': NumelDataset(src_tokens, reduce=True),
}
if not self.args.regression_target:
label_dataset = make_dataset('label', self.target_dictionary)
if label_dataset is not None:
dataset.update(target=OffsetTokensDataset(
StripTokenDataset(
label_dataset,
id_to_strip=self.target_dictionary.eos(),
),
offset=-self.target_dictionary.nspecial,
))
else:
label_path = "{0}.label".format(get_path('label', split))
if os.path.exists(label_path):
dataset.update(target=RawLabelDataset(
[float(x.strip()) for x in open(label_path).readlines()]))
nested_dataset = NestedDictionaryDataset(
dataset,
sizes=[src_tokens.sizes],
)
if self.args.no_shuffle:
dataset = nested_dataset
else:
dataset = SortDataset(
nested_dataset,
# shuffle
sort_order=[shuffle],
)
print("| Loaded {0} with #samples: {1}".format(split, len(dataset)))
self.datasets[split] = dataset
return self.datasets[split]
def load_dataset(self, split, combine=False, **kwargs):
"""Load a given dataset split (e.g., train, valid, test)."""
def get_path(type, split):
return os.path.join(self.args.data, type, split)
def make_dataset(type, dictionary):
split_path = get_path(type, split)
dataset = data_utils.load_indexed_dataset(
split_path,
dictionary,
self.args.dataset_impl,
combine=combine,
)
assert dataset is not None, "could not find dataset: {}".format(
get_path(type, split))
return dataset
src_tokens = make_dataset("input0", self.source_dictionary)
with data_utils.numpy_seed(self.args.seed):
shuffle = np.random.permutation(len(src_tokens))
label_dataset = make_dataset("label", self.label_dictionary)
dataset = {
"id": IdDataset(),
"net_input": {
"src_tokens": RightPadDataset(
src_tokens,
pad_idx=self.source_dictionary.pad(),
),
"src_lengths": NumelDataset(src_tokens, reduce=False),
},
"target": RightPadDataset( # use 1 as padding, will be used to mask out padding when calculating loss
ReplaceDataset( # replace eos and existing padding (used when some tokens should not be predicted) with -1
OffsetTokensDataset( # offset tokens to get the targets to the correct range (0,1,2,...)
label_dataset,
offset=-self.label_dictionary.nspecial,
),
replace_map={
self.label_dictionary.eos()
- self.label_dictionary.nspecial: -1,
self.label_dictionary.pad()
- self.label_dictionary.nspecial: -1,
},
offsets=np.zeros(len(label_dataset), dtype=np.int),
),
pad_idx=-1,
),
"nsentences": NumSamplesDataset(),
"ntokens": NumelDataset(src_tokens, reduce=True),
}
nested_dataset = NestedDictionaryDataset(
dataset,
sizes=[src_tokens.sizes],
)
if self.args.no_shuffle:
dataset = nested_dataset
else:
dataset = SortDataset(
nested_dataset,
# shuffle
sort_order=[shuffle],
)
logger.info("Loaded {0} with #samples: {1}".format(split, len(dataset)))
self.datasets[split] = dataset
return self.datasets[split]
