def __init__(
self,
model_name: str,
vocab: Vocabulary,
indexer: PretrainedTransformerIndexer = None,
max_decoding_steps: int = 140,
beam_size: int = 4,
encoder: Seq2SeqEncoder = None,
):
"""
# Parameters
model_name : `str`, required
Name of the pre-trained BART model to use. Available options can be found in
`transformers.models.bart.modeling_bart.BART_PRETRAINED_MODEL_ARCHIVE_MAP`.
vocab : `Vocabulary`, required
Vocabulary containing source and target vocabularies.
indexer : `PretrainedTransformerIndexer`, optional (default = `None`)
Indexer to be used for converting decoded sequences of ids to to sequences of tokens.
max_decoding_steps : `int`, optional (default = `128`)
Number of decoding steps during beam search.
beam_size : `int`, optional (default = `5`)
Number of beams to use in beam search. The default is from the BART paper.
encoder : `Seq2SeqEncoder`, optional (default = `None`)
Encoder to used in BART. By default, the original BART encoder is used.
"""
super().__init__(vocab)
self.bart = BartForConditionalGeneration.from_pretrained(model_name)
self._indexer = indexer or PretrainedTransformerIndexer(
model_name, namespace="tokens")
self._start_id = self.bart.config.bos_token_id # CLS
self._decoder_start_id = self.bart.config.decoder_start_token_id or self._start_id
self._end_id = self.bart.config.eos_token_id # SEP
self._pad_id = self.bart.config.pad_token_id # PAD
self._max_decoding_steps = max_decoding_steps
self._beam_search = BeamSearch(self._end_id,
max_steps=max_decoding_steps,
beam_size=beam_size or 1)
self._rouge = ROUGE(
exclude_indices={self._start_id, self._pad_id, self._end_id})
self._bleu = BLEU(
exclude_indices={self._start_id, self._pad_id, self._end_id})
# Replace bart encoder with given encoder. We need to extract the two embedding layers so that
# we can use them in the encoder wrapper
if encoder is not None:
assert (encoder.get_input_dim() == encoder.get_output_dim() ==
self.bart.config.hidden_size)
self.bart.model.encoder = _BartEncoderWrapper(
encoder,
self.bart.model.encoder.embed_tokens,
self.bart.model.encoder.embed_positions,
)
def __init__(
self,
model_name: str,
vocab: Vocabulary,
beam_search: Lazy[BeamSearch] = Lazy(BeamSearch),
indexer: PretrainedTransformerIndexer = None,
encoder: Seq2SeqEncoder = None,
**kwargs,
):
super().__init__(vocab)
self.bart = BartForConditionalGeneration.from_pretrained(model_name)
self._indexer = indexer or PretrainedTransformerIndexer(model_name, namespace="tokens")
self._start_id = self.bart.config.bos_token_id # CLS
self._decoder_start_id = self.bart.config.decoder_start_token_id or self._start_id
self._end_id = self.bart.config.eos_token_id # SEP
self._pad_id = self.bart.config.pad_token_id # PAD
# At prediction time, we'll use a beam search to find the best target sequence.
# For backwards compatibility, check if beam_size or max_decoding_steps were passed in as
# kwargs. If so, update the BeamSearch object before constructing and raise a DeprecationWarning
deprecation_warning = (
"The parameter {} has been deprecated."
" Provide this parameter as argument to beam_search instead."
)
beam_search_extras = {}
if "beam_size" in kwargs:
beam_search_extras["beam_size"] = kwargs["beam_size"]
warnings.warn(deprecation_warning.format("beam_size"), DeprecationWarning)
if "max_decoding_steps" in kwargs:
beam_search_extras["max_steps"] = kwargs["max_decoding_steps"]
warnings.warn(deprecation_warning.format("max_decoding_steps"), DeprecationWarning)
self._beam_search = beam_search.construct(
end_index=self._end_id, vocab=self.vocab, **beam_search_extras
)
self._rouge = ROUGE(exclude_indices={self._start_id, self._pad_id, self._end_id})
self._bleu = BLEU(exclude_indices={self._start_id, self._pad_id, self._end_id})
# Replace bart encoder with given encoder. We need to extract the two embedding layers so that
# we can use them in the encoder wrapper
if encoder is not None:
assert (
encoder.get_input_dim() == encoder.get_output_dim() == self.bart.config.hidden_size
)
self.bart.model.encoder = _BartEncoderWrapper(
encoder,
self.bart.model.encoder.embed_tokens,
self.bart.model.encoder.embed_positions,
)
def __init__(
self,
model_name: str,
vocab: Vocabulary,
indexer: PretrainedTransformerIndexer = None,
max_decoding_steps: int = 140,
beam_size: int = 4,
encoder: Seq2SeqEncoder = None,
):
super().__init__(vocab)
self.bart = BartForConditionalGeneration.from_pretrained(model_name)
self._indexer = indexer or PretrainedTransformerIndexer(
model_name, namespace="tokens")
self._start_id = self.bart.config.bos_token_id # CLS
self._decoder_start_id = self.bart.config.decoder_start_token_id or self._start_id
self._end_id = self.bart.config.eos_token_id # SEP
self._pad_id = self.bart.config.pad_token_id # PAD
self._max_decoding_steps = max_decoding_steps
self._beam_search = BeamSearch(self._end_id,
max_steps=max_decoding_steps,
beam_size=beam_size or 1)
self._rouge = ROUGE(
exclude_indices={self._start_id, self._pad_id, self._end_id})
self._bleu = BLEU(
exclude_indices={self._start_id, self._pad_id, self._end_id})
# Replace bart encoder with given encoder. We need to extract the two embedding layers so that
# we can use them in the encoder wrapper
if encoder is not None:
assert (encoder.get_input_dim() == encoder.get_output_dim() ==
self.bart.config.hidden_size)
self.bart.model.encoder = _BartEncoderWrapper(
encoder,
self.bart.model.encoder.embed_tokens,
self.bart.model.encoder.embed_positions,
)