def __init__(self, cfg: DictConfig, trainer=None):
if 'tokenizer' not in cfg:
raise ValueError(
"`cfg` must have `tokenizer` config to create a tokenizer !")
self.tokenizer_cfg = OmegaConf.to_container(cfg.tokenizer,
resolve=True) # type: dict
self.tokenizer_dir = self.tokenizer_cfg.pop(
'dir') # Remove tokenizer directory
self.tokenizer_type = self.tokenizer_cfg.pop(
'type').lower() # Remove tokenizer_type
# Setup the tokenizer
self._setup_tokenizer()
# Initialize a dummy vocabulary
vocabulary = self.tokenizer.tokenizer.get_vocab()
# Set the new vocabulary
cfg.decoder.params.vocabulary = ListConfig(list(vocabulary.values()))
# Override number of classes if placeholder provided
if cfg.decoder.params['num_classes'] < 1:
logging.info(
"\nReplacing placeholder number of classes ({}) with actual number of classes - {}"
.format(cfg.decoder.params['num_classes'], len(vocabulary)))
cfg.decoder.params['num_classes'] = len(vocabulary)
super().__init__(cfg=cfg, trainer=trainer)
# Setup metric objects
self._wer = WERBPE(tokenizer=self.tokenizer,
batch_dim_index=0,
use_cer=False,
ctc_decode=True)
def test_wer_metric_randomized(self, test_wer_bpe):
"""This test relies on correctness of word_error_rate function."""
def __random_string(length):
return ''.join(
random.choice(''.join(self.vocabulary)) for _ in range(length))
if test_wer_bpe:
wer = WERBPE(deepcopy(self.char_tokenizer),
batch_dim_index=0,
use_cer=False,
ctc_decode=True)
else:
wer = WER(vocabulary=self.vocabulary,
batch_dim_index=0,
use_cer=False,
ctc_decode=True)
for test_id in range(256):
n1 = random.randint(1, 512)
n2 = random.randint(1, 512)
s1 = __random_string(n1)
s2 = __random_string(n2)
# skip empty strings as reference
if s2.strip():
assert (abs(
self.get_wer(wer,
prediction=s1,
reference=s2,
use_tokenizer=test_wer_bpe) -
word_error_rate(hypotheses=[s1], references=[s2])) < 1e-6)
def change_vocabulary(self, new_tokenizer_dir: str, new_tokenizer_type: str):
"""
Changes vocabulary of the tokenizer used during CTC decoding process.
Use this method when fine-tuning on from pre-trained model.
This method changes only decoder and leaves encoder and pre-processing modules unchanged. For example, you would
use it if you want to use pretrained encoder when fine-tuning on a data in another language, or when you'd need
model to learn capitalization, punctuation and/or special characters.
Args:
new_tokenizer_dir: Path to the new tokenizer directory.
new_tokenizer_type: Either `bpe` or `wpe`. `bpe` is used for SentencePiece tokenizers,
whereas `wpe` is used for `BertTokenizer`.
Returns: None
"""
if not os.path.isdir(new_tokenizer_dir):
raise NotADirectoryError(
f'New tokenizer dir must be non-empty path to a directory. But I got: {new_tokenizer_dir}'
)
if new_tokenizer_type.lower() not in ('bpe', 'wpe'):
raise ValueError(f'New tokenizer type must be either `bpe` or `wpe`')
self.tokenizer_dir = new_tokenizer_dir # Remove tokenizer directory
self.tokenizer_type = new_tokenizer_type.lower() # Remove tokenizer_type
# Setup the tokenizer
self._setup_tokenizer()
# Initialize a dummy vocabulary
vocabulary = self.tokenizer.tokenizer.get_vocab()
# Set the new vocabulary
decoder_config = copy.deepcopy(self.decoder.to_config_dict())
decoder_config.params.vocabulary = ListConfig(list(vocabulary.values()))
# Override number of classes if placeholder provided
logging.info(
"\nReplacing old number of classes ({}) with new number of classes - {}".format(
decoder_config['params']['num_classes'], len(vocabulary)
)
)
decoder_config['params']['num_classes'] = len(vocabulary)
del self.decoder
self.decoder = EncDecCTCModelBPE.from_config_dict(decoder_config)
del self.loss
self.loss = CTCLoss(num_classes=self.decoder.num_classes_with_blank - 1, zero_infinity=True)
self._wer = WERBPE(tokenizer=self.tokenizer, batch_dim_index=0, use_cer=False, ctc_decode=True)
# Update config
OmegaConf.set_struct(self._cfg.decoder, False)
self._cfg.decoder = decoder_config
OmegaConf.set_struct(self._cfg.decoder, True)
logging.info(f"Changed tokenizer to {self.decoder.vocabulary} vocabulary.")
def test_wer_metric_decode(self, test_wer_bpe):
if test_wer_bpe:
wer = WERBPE(self.char_tokenizer,
batch_dim_index=0,
use_cer=False,
ctc_decode=True)
else:
wer = WER(vocabulary=self.vocabulary.copy(),
batch_dim_index=0,
use_cer=False,
ctc_decode=True)
tokens = self.__string_to_ctc_tensor(
'cat', use_tokenizer=test_wer_bpe)[0].int().numpy().tolist()
assert tokens == [3, 1, 20]
tokens_decoded = wer.decode_ids_to_tokens(tokens)
assert tokens_decoded == ['c', 'a', 't']
str_decoded = wer.decode_tokens_to_str(tokens)
assert str_decoded == 'cat'
def __init__(self, cfg: DictConfig, trainer=None):
if 'tokenizer' not in cfg:
raise ValueError("`cfg` must have `tokenizer` config to create a tokenizer !")
# Setup the tokenizer
self._setup_tokenizer(cfg.tokenizer)
# Initialize a dummy vocabulary
vocabulary = self.tokenizer.tokenizer.get_vocab()
# Set the new vocabulary
with open_dict(cfg):
if "params" in cfg.decoder:
cfg.decoder.params.vocabulary = ListConfig(list(vocabulary.values()))
else:
cfg.decoder.vocabulary = ListConfig(list(vocabulary.values()))
# Override number of classes if placeholder provided
if "params" in cfg.decoder:
num_classes = cfg.decoder["params"]["num_classes"]
else:
num_classes = cfg.decoder["num_classes"]
if num_classes < 1:
logging.info(
"\nReplacing placeholder number of classes ({}) with actual number of classes - {}".format(
num_classes, len(vocabulary)
)
)
if "params" in cfg.decoder:
cfg.decoder["params"]["num_classes"] = len(vocabulary)
else:
cfg.decoder["num_classes"] = len(vocabulary)
super().__init__(cfg=cfg, trainer=trainer)
# Setup metric objects
self._wer = WERBPE(
tokenizer=self.tokenizer,
batch_dim_index=0,
use_cer=self._cfg.get('use_cer', False),
ctc_decode=True,
dist_sync_on_step=True,
log_prediction=self._cfg.get("log_prediction", False),
)
def __init__(self, cfg: DictConfig, trainer=None):
# Convert to Hydra 1.0 compatible DictConfig
cfg = model_utils.convert_model_config_to_dict_config(cfg)
cfg = model_utils.maybe_update_config_version(cfg)
if 'tokenizer' not in cfg:
raise ValueError(
"`cfg` must have `tokenizer` config to create a tokenizer !")
# Setup the tokenizer
self._setup_tokenizer(cfg.tokenizer)
# Initialize a dummy vocabulary
vocabulary = self.tokenizer.tokenizer.get_vocab()
# Set the new vocabulary
with open_dict(cfg):
# sidestepping the potential overlapping tokens issue in aggregate tokenizers
if self.tokenizer_type == "agg":
cfg.decoder.vocabulary = ListConfig(vocabulary)
else:
cfg.decoder.vocabulary = ListConfig(list(vocabulary.keys()))
# Override number of classes if placeholder provided
num_classes = cfg.decoder["num_classes"]
if num_classes < 1:
logging.info(
"\nReplacing placeholder number of classes ({}) with actual number of classes - {}"
.format(num_classes, len(vocabulary)))
cfg.decoder["num_classes"] = len(vocabulary)
super().__init__(cfg=cfg, trainer=trainer)
# Setup metric objects
self._wer = WERBPE(
tokenizer=self.tokenizer,
batch_dim_index=0,
use_cer=self._cfg.get('use_cer', False),
ctc_decode=True,
dist_sync_on_step=True,
log_prediction=self._cfg.get("log_prediction", False),
)
def test_wer_metric_simple(self, batch_dim_index, test_wer_bpe):
if test_wer_bpe:
wer = WERBPE(self.char_tokenizer,
batch_dim_index,
use_cer=False,
ctc_decode=True)
else:
wer = WER(vocabulary=self.vocabulary,
batch_dim_index=batch_dim_index,
use_cer=False,
ctc_decode=True)
assert self.get_wer(wer, 'cat', 'cot', test_wer_bpe) == 1.0
assert self.get_wer(wer, 'gpu', 'g p u', test_wer_bpe) == 1.0
assert self.get_wer(wer, 'g p u', 'gpu', test_wer_bpe) == 3.0
assert self.get_wer(wer, 'ducati motorcycle', 'motorcycle',
test_wer_bpe) == 1.0
assert self.get_wer(wer, 'ducati motorcycle', 'ducuti motorcycle',
test_wer_bpe) == 0.5
assert abs(
self.get_wer(wer, 'a f c', 'a b c', test_wer_bpe) -
1.0 / 3.0) < 1e-6
def change_vocabulary(self, new_tokenizer_dir: Union[str, DictConfig],
new_tokenizer_type: str):
"""
Changes vocabulary of the tokenizer used during CTC decoding process.
Use this method when fine-tuning on from pre-trained model.
This method changes only decoder and leaves encoder and pre-processing modules unchanged. For example, you would
use it if you want to use pretrained encoder when fine-tuning on a data in another language, or when you'd need
model to learn capitalization, punctuation and/or special characters.
Args:
new_tokenizer_dir: Directory path to tokenizer or a config for a new tokenizer (if the tokenizer type is `agg`)
new_tokenizer_type: Either `agg`, `bpe` or `wpe`. `bpe` is used for SentencePiece tokenizers,
whereas `wpe` is used for `BertTokenizer`.
new_tokenizer_cfg: A config for the new tokenizer. if provided, pre-empts the dir and type
Returns: None
"""
if isinstance(new_tokenizer_dir, DictConfig):
if new_tokenizer_type == 'agg':
new_tokenizer_cfg = new_tokenizer_dir
else:
raise ValueError(
f'New tokenizer dir should be a string unless the tokenizer is `agg`, but this tokenizer type is: {new_tokenizer_type}'
)
else:
new_tokenizer_cfg = None
if new_tokenizer_cfg is not None:
tokenizer_cfg = new_tokenizer_cfg
else:
if not os.path.isdir(new_tokenizer_dir):
raise NotADirectoryError(
f'New tokenizer dir must be non-empty path to a directory. But I got: {new_tokenizer_dir}'
f"New tokenizer dir must be non-empty path to a directory. But I got: {new_tokenizer_dir}"
)
if new_tokenizer_type.lower() not in ('bpe', 'wpe'):
raise ValueError(
f'New tokenizer type must be either `bpe` or `wpe`')
tokenizer_cfg = OmegaConf.create({
'dir': new_tokenizer_dir,
'type': new_tokenizer_type
})
# Setup the tokenizer
self._setup_tokenizer(tokenizer_cfg)
# Initialize a dummy vocabulary
vocabulary = self.tokenizer.tokenizer.get_vocab()
# Set the new vocabulary
decoder_config = copy.deepcopy(self.decoder.to_config_dict())
# sidestepping the potential overlapping tokens issue in aggregate tokenizers
if self.tokenizer_type == "agg":
decoder_config.vocabulary = ListConfig(vocabulary)
else:
decoder_config.vocabulary = ListConfig(list(vocabulary.keys()))
decoder_num_classes = decoder_config['num_classes']
# Override number of classes if placeholder provided
logging.info(
"\nReplacing old number of classes ({}) with new number of classes - {}"
.format(decoder_num_classes, len(vocabulary)))
decoder_config['num_classes'] = len(vocabulary)
del self.decoder
self.decoder = EncDecCTCModelBPE.from_config_dict(decoder_config)
del self.loss
self.loss = CTCLoss(
num_classes=self.decoder.num_classes_with_blank - 1,
zero_infinity=True,
reduction=self._cfg.get("ctc_reduction", "mean_batch"),
)
self._wer = WERBPE(
tokenizer=self.tokenizer,
batch_dim_index=0,
use_cer=self._cfg.get('use_cer', False),
ctc_decode=True,
log_prediction=self._cfg.get("log_prediction", False),
)
# Update config
OmegaConf.set_struct(self._cfg.decoder, False)
self._cfg.decoder = decoder_config
OmegaConf.set_struct(self._cfg.decoder, True)
logging.info(
f"Changed tokenizer to {self.decoder.vocabulary} vocabulary.")