decoding key

游雁

2024-05-08 cfe577f16fef9fb5b0a48f07d4f9e232799cc9d4

 funasr/models/sense_voice/model.py

@@ -803,10 +803,12 @@
                tokenizer=tokenizer,
            )

            if len(kwargs.get("data_type", [])) > 1:
            if (
                isinstance(kwargs.get("data_type", None), (list, tuple))
                and len(kwargs.get("data_type", [])) > 1
            ):
                audio_sample_list, text_token_int_list = audio_sample_list
                text_token_int = text_token_int_list[0]
                text_token_int = tokenizer.encode(text_token_int)
            else:
                text_token_int = None

@@ -846,7 +848,7 @@
        )

        if text_token_int is not None:
            i = 1
            i = 0
            results = []
            ibest_writer = None
            if kwargs.get("output_dir") is not None:
@@ -855,8 +857,10 @@
                ibest_writer = self.writer[f"1best_recog"]

            # 1. Forward decoder
            ys_pad = torch.tensor(text_token_int, dtype=torch.int64).to(kwargs["device"])[None, :]
            ys_pad_lens = torch.tensor([len(text_token_int)], dtype=torch.int64).to(
            ys_pad = torch.tensor(sos_int + text_token_int, dtype=torch.int64).to(kwargs["device"])[
                None, :
            ]
            ys_pad_lens = torch.tensor([len(sos_int + text_token_int)], dtype=torch.int64).to(
                kwargs["device"]
            )[None, :]
            decoder_out = self.model.decoder(