DPTNet_eval/DPTNet_quant_sep.py

import os
import torch
import numpy as np
import torchaudio
import yaml
from . import asteroid_test
from huggingface_hub import hf_hub_download
import logging
logger = logging.getLogger(__name__)

torchaudio.set_audio_backend("sox_io")


def get_conf():
    conf_filterbank = {
        'n_filters': 64,
        'kernel_size': 16,
        'stride': 8
    }

    conf_masknet = {
        'in_chan': 64,
        'n_src': 2,
        'out_chan': 64,
        'ff_hid': 256,
        'ff_activation': "relu",
        'norm_type': "gLN",
        'chunk_size': 100,
        'hop_size': 50,
        'n_repeats': 2,
        'mask_act': 'sigmoid',
        'bidirectional': True,
        'dropout': 0
    }
    return conf_filterbank, conf_masknet


def load_dpt_model():
    print('Load Separation Model...')

    # 👇 從環境變數取得 HF Token
    from huggingface_hub import hf_hub_download
    speech_sep_token = os.getenv("SpeechSeparation")
    if not speech_sep_token:
        raise EnvironmentError("環境變數 SpeechSeparation 未設定！")

    # 👇 從 Hugging Face Hub 下載模型權重
    model_path = hf_hub_download(
        repo_id="DeepLearning101/speech-separation",  # 替換成你自己的 repo 名稱
        filename="train_dptnet_aishell_partOverlap_B2_300epoch_quan-int8.p",
        token=speech_sep_token
    )

    conf_filterbank, conf_masknet = get_conf()
    model_class = getattr(asteroid_test, "DPTNet")
    model = model_class(**conf_filterbank, **conf_masknet)
    model = torch.quantization.quantize_dynamic(model, {torch.nn.LSTM, torch.nn.Linear}, dtype=torch.qint8)

    try:
        state_dict = torch.load(model_path, map_location="cpu", weights_only=False)
    except pickle.UnpicklingError as e:
        raise RuntimeError(
            "模型載入失敗！請確認：\n"
            "1. 模型來源是否可信\n"
            "2. 是否為舊版 PyTorch 儲存的模型\n"
            "3. 嘗試鎖定 PyTorch 版本為 2.5.x"
        ) from e

    model.load_state_dict(state_dict)
    model.eval()
    return model


import torchaudio
import tempfile

def dpt_sep_process(wav_path, model=None, outfilename=None):
    try:
        # 添加設備檢測
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        model = model.to(device)
        
        # 強化音訊加載
        x, sr = torchaudio.load(wav_path, format="wav")
        x = x.mean(dim=0, keepdim=True).to(device)
        
        # 自動重採樣
        if sr != 16000:
            resampler = torchaudio.transforms.Resample(sr, 16000).to(device)
            x = resampler(x)
            sr = 16000
        
        # 分塊處理避免OOM
        chunk_size = sr * 60  # 每次處理1分鐘
        separated = []
        for i in range(0, x.shape[1], chunk_size):
            chunk = x[:, i:i+chunk_size]
            with torch.no_grad():
                est = model(chunk)
            separated.append(est.cpu())
        
        est_sources = torch.cat(separated, dim=2)

        # 後處理修正
        est_sources = est_sources.squeeze(0)
        sep_1, sep_2 = est_sources[0], est_sources[1]

        # 正規化增強
        peak = 0.9 * torch.max(torch.abs(x))
        sep_1 = peak * sep_1 / torch.max(torch.abs(sep_1))
        sep_2 = peak * sep_2 / torch.max(torch.abs(sep_2))

        # 使用臨時輸出目錄
        with tempfile.TemporaryDirectory() as tmp_dir:
            sep1_path = os.path.join(tmp_dir, "sep1.wav")
            sep2_path = os.path.join(tmp_dir, "sep2.wav")
            
            torchaudio.save(sep1_path, sep_1.unsqueeze(0), sr)
            torchaudio.save(sep2_path, sep_2.unsqueeze(0), sr)

            # 移動檔案到最終位置
            final_sep1 = outfilename.replace('.wav', '_sep1.wav')
            final_sep2 = outfilename.replace('.wav', '_sep2.wav')
            os.replace(sep1_path, final_sep1)
            os.replace(sep2_path, final_sep2)

        # 新增日誌
        logger.info(f"💾 寫入輸出檔案至: {final_sep1}, {final_sep2}")        
        
        return final_sep1, final_sep2

    except RuntimeError as e:
        if "CUDA out of memory" in str(e):
            raise RuntimeError("記憶體不足，請縮短音訊長度") from e
        else:
            raise

if __name__ == '__main__':
    print("This module should be used via Flask or Gradio.")