update

游雁

2024-02-23 ec42889511b121230e97bbcdf05f4d517f95d7ba

update

 funasr/auto/auto_model.py

@@ -181,15 +181,18 @@
        # init_param
        init_param = kwargs.get("init_param", None)
        if init_param is not None:
            logging.info(f"Loading pretrained params from {init_param}")
            load_pretrained_model(
                model=model,
                path=init_param,
                ignore_init_mismatch=kwargs.get("ignore_init_mismatch", False),
                oss_bucket=kwargs.get("oss_bucket", None),
                scope_map=kwargs.get("scope_map", None),
                excludes=kwargs.get("excludes", None),
            )
            if os.path.exists(init_param):
                logging.info(f"Loading pretrained params from {init_param}")
                load_pretrained_model(
                    model=model,
                    path=init_param,
                    ignore_init_mismatch=kwargs.get("ignore_init_mismatch", False),
                    oss_bucket=kwargs.get("oss_bucket", None),
                    scope_map=kwargs.get("scope_map", None),
                    excludes=kwargs.get("excludes", None),
                )
            else:
                print(f"error, init_param does not exist!: {init_param}")
        
        return model, kwargs
    

 funasr/models/llm_asr/model.py

@@ -217,7 +217,7 @@
    ) -> Tuple[torch.Tensor, torch.Tensor]:
    
        audio_mask = kwargs.get("audio_mask")
        audio_token_lengths = audio_mask.sum(-1)
        audio_token_lengths = audio_mask.sum(-1) if audio_mask else None

        batch = {"speech": speech, "speech_lengths": speech_lengths}
        enc, enc_lens = self.audio_encoder.encode(**batch)
@@ -279,7 +279,7 @@
        
    
        prompt_pre = "USER: \nINSTRUCTION: {}\nINPUT: ".format(prompt)
        prompt_ids = self.tokenizer.encode(prompt_pre)
        prompt_ids = tokenizer.encode(prompt_pre)
        prompt_length = len(prompt_ids)
        prompt_ids = torch.tensor(prompt_ids, dtype=torch.int64).to(kwargs["device"])


 funasr/train_utils/load_pretrained_model.py

@@ -118,7 +118,7 @@
        else:
            print(f"Warning, miss key in ckpt: {k}, mapped: {k_ddp}")
            
    flag = obj.load_state_dict(dst_state, strict=True)
    flag = obj.load_state_dict(dst_state, strict=False)
    # print(flag)

# def load_pretrained_model(