commit

语帆

2024-03-04 920331972a136834a560d78917de60f6c6623d96

 funasr/models/seaco_paraformer/model.py

@@ -30,7 +30,7 @@
from funasr.models.transformer.utils.nets_utils import make_pad_mask, pad_list
from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank


import pdb
if LooseVersion(torch.__version__) >= LooseVersion("1.6.0"):
    from torch.cuda.amp import autocast
else:
@@ -128,7 +128,7 @@
        hotword_pad = kwargs.get("hotword_pad")
        hotword_lengths = kwargs.get("hotword_lengths")
        dha_pad = kwargs.get("dha_pad")

        
        batch_size = speech.shape[0]
        # for data-parallel
        text = text[:, : text_lengths.max()]
@@ -209,17 +209,20 @@
                               nfilter=50,
                               seaco_weight=1.0):
        # decoder forward

        decoder_out, decoder_hidden, _ = self.decoder(encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens, return_hidden=True, return_both=True)

        decoder_pred = torch.log_softmax(decoder_out, dim=-1)
        if hw_list is not None:
            hw_lengths = [len(i) for i in hw_list]
            hw_list_ = [torch.Tensor(i).long() for i in hw_list]
            hw_list_pad = pad_list(hw_list_, 0).to(encoder_out.device)
            selected = self._hotword_representation(hw_list_pad, torch.Tensor(hw_lengths).int().to(encoder_out.device))

            contextual_info = selected.squeeze(0).repeat(encoder_out.shape[0], 1, 1).to(encoder_out.device)
            num_hot_word = contextual_info.shape[1]
            _contextual_length = torch.Tensor([num_hot_word]).int().repeat(encoder_out.shape[0]).to(encoder_out.device)
            

            # ASF Core
            if nfilter > 0 and nfilter < num_hot_word:
                hotword_scores = self.seaco_decoder.forward_asf6(contextual_info, _contextual_length, decoder_hidden, ys_pad_lens)
@@ -239,7 +242,7 @@
            cif_attended, _ = self.seaco_decoder(contextual_info, _contextual_length, sematic_embeds, ys_pad_lens)
            dec_attended, _ = self.seaco_decoder(contextual_info, _contextual_length, decoder_hidden, ys_pad_lens)
            merged = self._merge(cif_attended, dec_attended)
            

            dha_output = self.hotword_output_layer(merged)  # remove the last token in loss calculation
            dha_pred = torch.log_softmax(dha_output, dim=-1)
            def _merge_res(dec_output, dha_output):
@@ -253,8 +256,8 @@
                # logits = dec_output * dha_mask + dha_output[:,:,:-1] * (1-dha_mask)
                logits = dec_output * dha_mask + dha_output[:,:,:] * (1-dha_mask)
                return logits

            merged_pred = _merge_res(decoder_pred, dha_pred)
            # import pdb; pdb.set_trace()
            return merged_pred
        else:
            return decoder_pred
@@ -304,7 +307,6 @@
            logging.info("enable beam_search")
            self.init_beam_search(**kwargs)
            self.nbest = kwargs.get("nbest", 1)
        
        meta_data = {}
        
        # extract fbank feats
@@ -330,6 +332,7 @@
        if isinstance(encoder_out, tuple):
            encoder_out = encoder_out[0]
        

        # predictor
        predictor_outs = self.calc_predictor(encoder_out, encoder_out_lens)
        pre_acoustic_embeds, pre_token_length, _, _ = predictor_outs[0], predictor_outs[1], \
@@ -338,15 +341,14 @@
        if torch.max(pre_token_length) < 1:
            return []


        decoder_out = self._seaco_decode_with_ASF(encoder_out, encoder_out_lens,
                                                   pre_acoustic_embeds,
                                                   pre_token_length,
                                                   hw_list=self.hotword_list)

        # decoder_out, _ = decoder_outs[0], decoder_outs[1]
        _, _, us_alphas, us_peaks = self.calc_predictor_timestamp(encoder_out, encoder_out_lens,
                                                                  pre_token_length)
        
        results = []
        b, n, d = decoder_out.size()
        for i in range(b):