grpc client remove VAD

huangmingming

2023-03-13 49c00a7d6cb9c05d4bd0bb0fc8b59a2eed4b8950

 funasr/bin/tp_inference.py

@@ -1,5 +1,6 @@
import argparse
import logging
from optparse import Option
import sys
import json
from pathlib import Path
@@ -11,15 +12,12 @@
from typing import Union
from typing import Dict

import math
import numpy as np
import torch
from typeguard import check_argument_types
from typeguard import check_return_type

from funasr.fileio.datadir_writer import DatadirWriter
from funasr.modules.scorers.scorer_interface import BatchScorerInterface
from funasr.modules.subsampling import TooShortUttError
from funasr.datasets.preprocessor import LMPreprocessor
from funasr.tasks.asr import ASRTaskAligner as ASRTask
from funasr.torch_utils.device_funcs import to_device
from funasr.torch_utils.set_all_random_seed import set_all_random_seed
@@ -28,7 +26,6 @@
from funasr.utils.types import str2bool
from funasr.utils.types import str2triple_str
from funasr.utils.types import str_or_none
from funasr.utils import asr_utils, wav_utils, postprocess_utils
from funasr.models.frontend.wav_frontend import WavFrontend
from funasr.text.token_id_converter import TokenIDConverter

@@ -91,7 +88,7 @@
    for char, timestamp in zip(new_char_list, timestamp_list):
        res_str += "{} {} {};".format(char, str(timestamp[0]+0.0005)[:5], str(timestamp[1]+0.0005)[:5])
    res = []
    for char, timestamp in zip(char_list, timestamp_list):
    for char, timestamp in zip(new_char_list, timestamp_list):
        if char != '<sil>':
            res.append([int(timestamp[0] * 1000), int(timestamp[1] * 1000)])
    return res_str, res
@@ -113,8 +110,8 @@
            timestamp_infer_config, timestamp_model_file, device
        )
        if 'cuda' in device:
            tp_model = tp_model.cuda()
            
            tp_model = tp_model.cuda()  # force model to cuda

        frontend = None
        if tp_train_args.frontend is not None:
            frontend = WavFrontend(cmvn_file=timestamp_cmvn_file, **tp_train_args.frontend_conf)
@@ -191,6 +188,8 @@
        dtype: str = "float32",
        seed: int = 0,
        num_workers: int = 1,
        split_with_space: bool = True,
        seg_dict_file: Optional[str] = None,
        **kwargs,
):
    inference_pipeline = inference_modelscope(
@@ -206,6 +205,8 @@
        dtype=dtype,
        seed=seed,
        num_workers=num_workers,
        split_with_space=split_with_space,
        seg_dict_file=seg_dict_file,
        **kwargs,
    )
    return inference_pipeline(data_path_and_name_and_type, raw_inputs)
@@ -226,6 +227,8 @@
        dtype: str = "float32",
        seed: int = 0,
        num_workers: int = 1,
        split_with_space: bool = True,
        seg_dict_file: Optional[str] = None,
        **kwargs,
):
    assert check_argument_types()
@@ -256,6 +259,19 @@
    )
    logging.info("speechtext2timestamp_kwargs: {}".format(speechtext2timestamp_kwargs))
    speechtext2timestamp = SpeechText2Timestamp(**speechtext2timestamp_kwargs)

    preprocessor = LMPreprocessor(
        train=False,
        token_type=speechtext2timestamp.tp_train_args.token_type,
        token_list=speechtext2timestamp.tp_train_args.token_list,
        bpemodel=None,
        text_cleaner=None,
        g2p_type=None,
        text_name="text",
        non_linguistic_symbols=speechtext2timestamp.tp_train_args.non_linguistic_symbols,
        split_with_space=split_with_space,
        seg_dict_file=seg_dict_file,
    )
    
    def _forward(
            data_path_and_name_and_type,
@@ -277,14 +293,11 @@
            batch_size=batch_size,
            key_file=key_file,
            num_workers=num_workers,
            preprocess_fn=ASRTask.build_preprocess_fn(speechtext2timestamp.tp_train_args, False),
            preprocess_fn=preprocessor,
            collate_fn=ASRTask.build_collate_fn(speechtext2timestamp.tp_train_args, False),
            allow_variable_data_keys=allow_variable_data_keys,
            inference=True,
        )

        finish_count = 0
        file_count = 1

        tp_result_list = []
        for keys, batch in loader:
@@ -304,7 +317,8 @@
                token = speechtext2timestamp.converter.ids2tokens(batch['text'][batch_id])
                ts_str, ts_list = time_stamp_lfr6_advance(us_alphas[batch_id], us_cif_peak[batch_id], token)
                logging.warning(ts_str)
                tp_result_list.append({'text':"".join([i for i in token if i != '<sil>']), 'timestamp': ts_list})
                item = {'key': key, 'value': ts_str, 'timestamp':ts_list}
                tp_result_list.append(item)
        return tp_result_list

    return _forward
@@ -389,6 +403,18 @@
        default=1,
        help="The batch size for inference",
    )
    group.add_argument(
        "--seg_dict_file",
        type=str,
        default=None,
        help="The batch size for inference",
    )
    group.add_argument(
        "--split_with_space",
        type=bool,
        default=False,
        help="The batch size for inference",
    )

    return parser