Python 是否有任何方法可以提取音频功能并以比此代码更快的速度保存？

我有一个pytorch代码，用于提取大约60万个音频特征。为了节省时间，我需要尽可能快地提取。我尝试了几种方法，最后，我提出了以下解决方案。目前，我正在使用pytorch dataloader的循环系统

我想知道是否有一种更快的方法来完成这一切。是否有某种方法可以在我的数据加载循环中找到最大值？同样，我不允许使用内置算法进行搜索或排序。任何建议都会有帮助

多谢各位

start_time = time.time()

dataset = AudioLoader(
    base_root=BASE_ROOT,
    date=DATE,
    data_type=DATA_TYPE,
    feat_type=FEAT_TYPE,
    params=PARAMS,
    samp_rate=SAMP_RATE
)

loader = DataLoader(
    dataset,
    batch_size=1,
    shuffle=False,
    num_workers=4
)

total_len = 0
mean = torch.zeros(FEAT_DIM)
sq_mean = torch.zeros(FEAT_DIM)
std = torch.zeros(FEAT_DIM)

for k, data in enumerate(loader):
    id = data['ID'][0]
    input_file = data['feat'][0]
    output_file = data['output'][0]

    if CAL_CMVN == True:
        total_len += input_file.shape[0]
        mean += torch.sum(input_file,0)
        sq_mean += torch.sum(torch.square(input_file), 0)

    save_features(input_file, WAV_PATH, output_file, OUTPUT_PATH, id)

    if k > 0 and k % 1000 == 0:
        print("-----{}s seconds-----".format(time.time() - start_time))

if CAL_CMVN == True:
    mean /= total_len
    sq_mean /= total_len
    std = torch.sqrt(sq_mean - mean*mean)

    torch.save(mean, os.path.join('./feat', 'mean.pt'))
    torch.save(std, os.path.join('./feat', 'std.pt'))

print("Total consumed time : {}s seconds".format(time.time() - start_time))

“保存功能”用于在路径中保存功能。特征文件是输入文件，输出文件是神经网络训练输入特征的目标。另外，我可以向您展示AudioLoader的详细信息

class AudioLoader(Dataset):
    """Pytorch audio loader."""

    def __init__(
            self,
            base_root=BASE_ROOT,
            date="210223",
            data_type="train",
            feat_type="mfcc",
            params={
                "channel": -1,
                "dither": 0.0,
                "window_type": "hamming",
                "sample_frequency": 16000
            },
            samp_rate=16000
    ):
        """
        Args:
        """
        wavfile = os.path.join(base_root, date, "data", data_type) + ".scp"
        txtfile = os.path.join(base_root, date, "data", data_type) + ".txt"
        
        self.wavlist = open(wavfile, "r", encoding="UTF8").read().splitlines()
        self.txtlist = open(txtfile, "r", encoding="UTF8").read().splitlines()
        self.feat_type = feat_type
        self.samp_rate = samp_rate
        self.params = params

    def __len__(self):
        return len(self.wavlist)

    def __getitem__(self, idx):

        FEAT=torchaudio.compliance.kaldi.spectrogram
        
        if self.feat_type == "fbank":
            FEAT=torchaudio.compliance.kaldi.fbank
        elif self.feat_type == "mfcc":
            FEAT=torchaudio.compliance.kaldi.mfcc

        wav_id = self.wavlist[idx].split()[0]
        txt_id = self.txtlist[idx].split()[0]

        if wav_id != txt_id:
            raise Exception('Both id are unmatched. You shoud sort again.')

        wav_path = self.wavlist[idx].split()[-1]
        output_txt = self.txtlist[idx].split(' ', 1)[1]

        if os.path.isfile(wav_path):
            try:
                waveform, samp = torchaudio.load(wav_path)
                resamp_waveform = torchaudio.compliance.kaldi.resample_waveform(waveform, samp, self.samp_rate)
                feature = FEAT(resamp_waveform, **self.params)
            except:
                print('Extraction failed')
                return {'ID': wav_id}
        else:
            print('Extraction failed')
            return {'ID': wav_id}
            
        return {'ID': wav_id, 'feat': feature, 'output': output_txt}