pytext.torchscript.seq2seq package

Submodules

pytext.torchscript.seq2seq.beam_decode module

class pytext.torchscript.seq2seq.beam_decode.BeamDecode(eos_token_id, length_penalty, nbest, beam_size, stop_at_eos)

Bases: torch.nn.modules.module.Module

Decodes the output of Beam Search to get the top hypotheses

forward(beam_tokens: torch.Tensor, beam_scores: torch.Tensor, token_weights: torch.Tensor, beam_prev_indices: torch.Tensor, num_steps: int) → List[Tuple[torch.Tensor, float, List[float], torch.Tensor, torch.Tensor]]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

pytext.torchscript.seq2seq.beam_search module

class pytext.torchscript.seq2seq.beam_search.BeamSearch(model_list, tgt_dict_eos, beam_size: int = 2, quantize: bool = False, record_attention: bool = False)

Bases: torch.nn.modules.module.Module

forward(src_tokens: torch.Tensor, src_lengths: torch.Tensor, num_steps: int, dict_feat: Optional[Tuple[torch.Tensor, torch.Tensor, torch.Tensor]] = None, contextual_token_embedding: Optional[torch.Tensor] = None)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

pytext.torchscript.seq2seq.decoder module

class pytext.torchscript.seq2seq.decoder.DecoderBatchedStepEnsemble(models, beam_size, record_attention=False)

Bases: torch.nn.modules.module.Module

This method should have a common interface such that it can be called after the encoder as well as after the decoder

beam_search_aggregate_topk(log_probs_per_model: List[torch.Tensor], attn_weights_per_model: List[torch.Tensor], prev_scores: torch.Tensor, beam_size: int, record_attention: bool)

forward(prev_tokens: torch.Tensor, prev_scores: torch.Tensor, timestep: int, decoder_ips: List[Dict[str, torch.Tensor]]) → Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, List[Dict[str, torch.Tensor]]]

Decoder step inputs correspond one-to-one to encoder outputs. HOWEVER: after the first step, encoder outputs (i.e, the first len(self.models) elements of inputs) must be tiled k (beam size) times on the batch dimension (axis 1).

reset_incremental_states()

pytext.torchscript.seq2seq.encoder module

class pytext.torchscript.seq2seq.encoder.EncoderEnsemble(models, beam_size)

Bases: torch.nn.modules.module.Module

This class will call the encoders from all the models in the ensemble. It will process the encoder output to prepare input for each decoder step input

forward(src_tokens: torch.Tensor, src_lengths: torch.Tensor, dict_feat: Optional[Tuple[torch.Tensor, torch.Tensor, torch.Tensor]] = None, contextual_token_embedding: Optional[torch.Tensor] = None) → List[Dict[str, torch.Tensor]]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

prepare_decoderstep_ip(futures: List[Dict[str, torch.Tensor]]) → List[Dict[str, torch.Tensor]]

pytext.torchscript.seq2seq.export_model module

class pytext.torchscript.seq2seq.export_model.Seq2SeqJIT(src_dict, tgt_dict, sequence_generator, filter_eos_bos, copy_unk_token=False, dictfeat_dict=None)

Bases: torch.nn.modules.module.Module

forward(src_tokens: List[str], dict_feat: Optional[Tuple[List[str], List[float], List[int]]] = None, contextual_token_embedding: Optional[List[float]] = None) → List[Tuple[List[str], float, List[float]]]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

prepare_generator_inputs(word_ids: List[int], dict_feat: Optional[Tuple[List[str], List[float], List[int]]] = None, contextual_token_embedding: Optional[List[float]] = None) → Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, torch.Tensor, torch.Tensor]], Optional[torch.Tensor], torch.Tensor]

pytext.torchscript.seq2seq.scripted_seq2seq_generator module

class pytext.torchscript.seq2seq.scripted_seq2seq_generator.ScriptedSequenceGenerator(models, trg_dict_eos, config)

Bases: pytext.models.module.Module

forward(src_tokens: torch.Tensor, dict_feat: Optional[Tuple[torch.Tensor, torch.Tensor, torch.Tensor]], contextual_token_embedding: Optional[torch.Tensor], src_lengths: torch.Tensor) → List[Tuple[torch.Tensor, float, List[float], torch.Tensor, torch.Tensor]]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

classmethod from_config(config, models, trg_dict_eos)

generate_hypo(tensors: Dict[str, torch.Tensor])

pytext.torchscript.seq2seq.seq2seq_rnn_decoder_utils module

pytext.torchscript.seq2seq.seq2seq_rnn_decoder_utils.get_src_length(models, decoder_ip)

pytext.torchscript.seq2seq.seq2seq_rnn_decoder_utils.prepare_decoder_ips(models, decoder_ip, model_state_outputs, prev_hypos)

Module contents