Skip to content

t5

allennlp_models.generation.models.t5

[SOURCE]


T5#

@Model.register("t5")
class T5(Model):
 | def __init__(
 |     self,
 |     vocab: Vocabulary,
 |     model_name: str,
 |     beam_search: Lazy[BeamSearch] = Lazy(BeamSearch, beam_size=3, max_steps=50),
 |     checkpoint_wrapper: Optional[CheckpointWrapper] = None,
 |     weights_path: Optional[Union[str, PathLike]] = None,
 |     **kwargs
 | ) -> None

tokenizer#

class T5(Model):
 | ...
 | @property
 | def tokenizer(self) -> PretrainedTransformerTokenizer

forward#

class T5(Model):
 | ...
 | def forward(
 |     self,
 |     source_tokens: TextFieldTensors,
 |     target_tokens: Optional[TextFieldTensors] = None
 | ) -> Dict[str, torch.Tensor]

Performs the forward step of T5.

Parameters

  • source_tokens : TextFieldTensors
    The source tokens for the encoder. We assume they are stored under the tokens key/namespace.

  • target_tokens : TextFieldTensors, optional (default = None)
    The target tokens for the decoder. We assume they are also stored under the tokens key/namespace. If no target tokens are given during training / validation, the source tokens are shifted to the right by 1.

Returns

  • Dict[str, torch.Tensor]
    Contains the loss when target_tokens is provided. And during prediction, includes predictions and predicted_log_probs from beam search.

make_output_human_readable#

class T5(Model):
 | ...
 | def make_output_human_readable(
 |     self,
 |     output_dict: Dict[str, torch.Tensor]
 | ) -> Dict[str, Any]

get_metrics#

class T5(Model):
 | ...
 | def get_metrics(self, reset: bool = False) -> Dict[str, float]

default_predictor#

class T5(Model):
 | ...
 | default_predictor = "seq2seq"