Merge branch 'main' of github.com:alibaba-damo-academy/FunASR add

游雁

2023-09-13 33d3d2084403fd34b79c835d2f2fe04f6cd8f738

 funasr/models/e2e_asr_contextual_paraformer.py

@@ -9,7 +9,6 @@
import numpy as np

import torch
from typeguard import check_argument_types

from funasr.layers.abs_normalize import AbsNormalize
from funasr.models.ctc import CTC
@@ -43,9 +42,7 @@
        frontend: Optional[AbsFrontend],
        specaug: Optional[AbsSpecAug],
        normalize: Optional[AbsNormalize],
        preencoder: Optional[AbsPreEncoder],
        encoder: AbsEncoder,
        postencoder: Optional[AbsPostEncoder],
        decoder: AbsDecoder,
        ctc: CTC,
        ctc_weight: float = 0.5,
@@ -72,8 +69,9 @@
        crit_attn_weight: float = 0.0,
        crit_attn_smooth: float = 0.0,
        bias_encoder_dropout_rate: float = 0.0,
        preencoder: Optional[AbsPreEncoder] = None,
        postencoder: Optional[AbsPostEncoder] = None,
    ):
        assert check_argument_types()
        assert 0.0 <= ctc_weight <= 1.0, ctc_weight
        assert 0.0 <= interctc_weight < 1.0, interctc_weight

@@ -343,7 +341,7 @@
            input_mask_expand_dim, 0)
        return sematic_embeds * tgt_mask, decoder_out * tgt_mask

    def cal_decoder_with_predictor(self, encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens, hw_list=None):
    def cal_decoder_with_predictor(self, encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens, hw_list=None, clas_scale=1.0):
        if hw_list is None:
            hw_list = [torch.Tensor([1]).long().to(encoder_out.device)]  # empty hotword list
            hw_list_pad = pad_list(hw_list, 0)
@@ -352,6 +350,7 @@
            else:
                hw_embed = self.bias_embed(hw_list_pad)
            hw_embed, (h_n, _) = self.bias_encoder(hw_embed)
            hw_embed = h_n.repeat(encoder_out.shape[0], 1, 1)
        else:
            hw_lengths = [len(i) for i in hw_list]
            hw_list_pad = pad_list([torch.Tensor(i).long() for i in hw_list], 0).to(encoder_out.device)
@@ -365,7 +364,7 @@
            hw_embed = h_n.repeat(encoder_out.shape[0], 1, 1)
        
        decoder_outs = self.decoder(
            encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens, contextual_info=hw_embed
            encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens, contextual_info=hw_embed, clas_scale=clas_scale
        )
        decoder_out = decoder_outs[0]
        decoder_out = torch.log_softmax(decoder_out, dim=-1)