update readme

雾聪

2024-03-03 807a3acdbb6511dab3b5af5e952ef5a8fe231c99

 funasr/models/contextual_paraformer/model.py

@@ -65,11 +65,9 @@


        if bias_encoder_type == 'lstm':
            logging.warning("enable bias encoder sampling and contextual training")
            self.bias_encoder = torch.nn.LSTM(inner_dim, inner_dim, 1, batch_first=True, dropout=bias_encoder_dropout_rate)
            self.bias_embed = torch.nn.Embedding(self.vocab_size, inner_dim)
        elif bias_encoder_type == 'mean':
            logging.warning("enable bias encoder sampling and contextual training")
            self.bias_embed = torch.nn.Embedding(self.vocab_size, inner_dim)
        else:
            logging.error("Unsupport bias encoder type: {}".format(bias_encoder_type))
@@ -192,13 +190,10 @@
        # 0. sampler
        decoder_out_1st = None
        if self.sampling_ratio > 0.0:
            if self.step_cur < 2:
                logging.info("enable sampler in paraformer, sampling_ratio: {}".format(self.sampling_ratio))

            sematic_embeds, decoder_out_1st = self.sampler(encoder_out, encoder_out_lens, ys_pad, ys_pad_lens,
                                                           pre_acoustic_embeds, contextual_info)
        else:
            if self.step_cur < 2:
                logging.info("disable sampler in paraformer, sampling_ratio: {}".format(self.sampling_ratio))
            sematic_embeds = pre_acoustic_embeds
        
        # 1. Forward decoder
@@ -382,9 +377,11 @@
                nbest_hyps = [Hypothesis(yseq=yseq, score=score)]
            for nbest_idx, hyp in enumerate(nbest_hyps):
                ibest_writer = None
                if ibest_writer is None and kwargs.get("output_dir") is not None:
                    writer = DatadirWriter(kwargs.get("output_dir"))
                    ibest_writer = writer[f"{nbest_idx + 1}best_recog"]
                if kwargs.get("output_dir") is not None:
                    if not hasattr(self, "writer"):
                        self.writer = DatadirWriter(kwargs.get("output_dir"))
                    ibest_writer = self.writer[f"{nbest_idx + 1}best_recog"]
    
                # remove sos/eos and get results
                last_pos = -1
                if isinstance(hyp.yseq, list):