update code

shixian.shi

2024-03-08 7498bd7388afdde8d5e6f8a4cb6aeb8be8ac60fa

 funasr/auto/auto_model.py

@@ -28,7 +28,7 @@
    from funasr.models.campplus.cluster_backend import ClusterBackend
except:
    print("If you want to use the speaker diarization, please `pip install hdbscan`")

import pdb

def prepare_data_iterator(data_in, input_len=None, data_type=None, key=None):
    """
@@ -46,6 +46,7 @@
    chars = string.ascii_letters + string.digits
    if isinstance(data_in, str) and data_in.startswith('http'): # url
        data_in = download_from_url(data_in)

    if isinstance(data_in, str) and os.path.exists(data_in): # wav_path; filelist: wav.scp, file.jsonl;text.txt;
        _, file_extension = os.path.splitext(data_in)
        file_extension = file_extension.lower()
@@ -142,11 +143,11 @@
    def build_model(self, **kwargs):
        assert "model" in kwargs
        if "model_conf" not in kwargs:
            logging.info("download models from model hub: {}".format(kwargs.get("model_hub", "ms")))
            logging.info("download models from model hub: {}".format(kwargs.get("hub", "ms")))
            kwargs = download_model(**kwargs)
        
        set_all_random_seed(kwargs.get("seed", 0))
        

        device = kwargs.get("device", "cuda")
        if not torch.cuda.is_available() or kwargs.get("ngpu", 1) == 0:
            device = "cpu"
@@ -165,22 +166,21 @@

            kwargs["token_list"] = tokenizer.token_list if hasattr(tokenizer, "token_list") else None
            kwargs["token_list"] = tokenizer.get_vocab() if hasattr(tokenizer, "get_vocab") else kwargs["token_list"]
            vocab_size = len(kwargs["token_list"])
            vocab_size = len(kwargs["token_list"]) if kwargs["token_list"] is not None else -1
        else:
            vocab_size = -1
        
        # build frontend
        frontend = kwargs.get("frontend", None)
        kwargs["input_size"] = None
        if frontend is not None:
            frontend_class = tables.frontend_classes.get(frontend)
            frontend = frontend_class(**kwargs["frontend_conf"])
            kwargs["frontend"] = frontend
            kwargs["input_size"] = frontend.output_size()
            kwargs["input_size"] = frontend.output_size() if hasattr(frontend, "output_size") else None
        
        # build model
        model_class = tables.model_classes.get(kwargs["model"])
        model = model_class(**kwargs, **kwargs["model_conf"], vocab_size=vocab_size)
        
        model = model_class(**kwargs, **kwargs.get("model_conf", {}), vocab_size=vocab_size)
        model.to(device)
        
        # init_param
@@ -193,7 +193,7 @@
                    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", "module.,None"),
                    scope_map=kwargs.get("scope_map", []),
                    excludes=kwargs.get("excludes", None),
                )
            else:
@@ -223,9 +223,9 @@
        batch_size = kwargs.get("batch_size", 1)
        # if kwargs.get("device", "cpu") == "cpu":
        #     batch_size = 1
        

        key_list, data_list = prepare_data_iterator(input, input_len=input_len, data_type=kwargs.get("data_type", None), key=key)
        

        speed_stats = {}
        asr_result_list = []
        num_samples = len(data_list)
@@ -238,13 +238,17 @@
            data_batch = data_list[beg_idx:end_idx]
            key_batch = key_list[beg_idx:end_idx]
            batch = {"data_in": data_batch, "key": key_batch}

            if (end_idx - beg_idx) == 1 and kwargs.get("data_type", None) == "fbank": # fbank
                batch["data_in"] = data_batch[0]
                batch["data_lengths"] = input_len

            time1 = time.perf_counter()
            with torch.no_grad():
                results, meta_data = model.inference(**batch, **kwargs)
                 res = model.inference(**batch, **kwargs)
                 if isinstance(res, (list, tuple)):
                    results = res[0]
                    meta_data = res[1] if len(res) > 1 else {}
            time2 = time.perf_counter()

            asr_result_list.extend(results)
@@ -392,7 +396,8 @@
            # step.3 compute punc model
            if self.punc_model is not None:
                if not len(result["text"]):
                    result['raw_text'] = ''
                    if return_raw_text:
                        result['raw_text'] = ''
                else:
                    self.punc_kwargs.update(cfg)
                    punc_res = self.inference(result["text"], model=self.punc_model, kwargs=self.punc_kwargs, **cfg)
@@ -434,10 +439,13 @@
                distribute_spk(sentence_list, sv_output)
                result['sentence_info'] = sentence_list
            elif kwargs.get("sentence_timestamp", False):
                sentence_list = timestamp_sentence(punc_res[0]['punc_array'],
                                                   result['timestamp'],
                                                   raw_text,
                                                   return_raw_text=return_raw_text)
                if not len(result['text']):
                    sentence_list = []
                else:
                    sentence_list = timestamp_sentence(punc_res[0]['punc_array'],
                                                       result['timestamp'],
                                                       raw_text,
                                                       return_raw_text=return_raw_text)
                result['sentence_info'] = sentence_list
            if "spk_embedding" in result: del result['spk_embedding']