atsr

语帆

2024-02-29 96eaabca5b2e9c93b40c9840e2ae0003a618bb6e

 funasr/models/lcbnet/model.py

@@ -111,8 +111,8 @@
            )
    
        self.blank_id = blank_id
        self.sos = sos if sos is not None else vocab_size - 1
        self.eos = eos if eos is not None else vocab_size - 1
        self.sos = vocab_size - 1
        self.eos = vocab_size - 1
        self.vocab_size = vocab_size
        self.ignore_id = ignore_id
        self.ctc_weight = ctc_weight
@@ -181,8 +181,7 @@
                text: (Batch, Length)
                text_lengths: (Batch,)
        """
        # import pdb;
        # pdb.set_trace()

        if len(text_lengths.size()) > 1:
            text_lengths = text_lengths[:, 0]
        if len(speech_lengths.size()) > 1:
@@ -375,14 +374,14 @@
        scorers["ngram"] = ngram
        
        weights = dict(
            decoder=1.0 - kwargs.get("decoding_ctc_weight", 0.5),
            ctc=kwargs.get("decoding_ctc_weight", 0.5),
            decoder=1.0 - kwargs.get("decoding_ctc_weight", 0.3),
            ctc=kwargs.get("decoding_ctc_weight", 0.3),
            lm=kwargs.get("lm_weight", 0.0),
            ngram=kwargs.get("ngram_weight", 0.0),
            length_bonus=kwargs.get("penalty", 0.0),
        )
        beam_search = BeamSearch(
            beam_size=kwargs.get("beam_size", 10),
            beam_size=kwargs.get("beam_size", 20),
            weights=weights,
            scorers=scorers,
            sos=self.sos,
@@ -444,18 +443,17 @@
            encoder_out = encoder_out[0]
        
        ocr_list_new = [[x + 1 if x != 0 else x for x in sublist] for sublist in ocr_sample_list]
        ocr = torch.tensor(ocr_list_new)
        ocr_lengths = ocr.new_full([1], dtype=torch.long, fill_value=ocr.size(1))
        pdb.set_trace()
        ocr = torch.tensor(ocr_list_new).to(device=kwargs["device"])
        ocr_lengths = ocr.new_full([1], dtype=torch.long, fill_value=ocr.size(1)).to(device=kwargs["device"])
        ocr, ocr_lens, _ = self.text_encoder(ocr, ocr_lengths)
        pdb.set_trace()
        fusion_out, _, _, _ = self.fusion_encoder(encoder_out,None, ocr, None)
        encoder_out = encoder_out + fusion_out
        # c. Passed the encoder result and the beam search
        nbest_hyps = self.beam_search(
            x=encoder_out[0], maxlenratio=kwargs.get("maxlenratio", 0.0), minlenratio=kwargs.get("minlenratio", 0.0)
        )
        
        nbest_hyps = nbest_hyps[: self.nbest]


        results = []
        b, n, d = encoder_out.size()
@@ -481,7 +479,7 @@
                # Change integer-ids to tokens
                token = tokenizer.ids2tokens(token_int)
                text = tokenizer.tokens2text(token)
                

                text_postprocessed, _ = postprocess_utils.sentence_postprocess(token)
                result_i = {"key": key[i], "token": token, "text": text_postprocessed}
                results.append(result_i)