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

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
    )

    # 👇 原本邏輯完全不變
    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)

    state_dict = torch.load(model_path, map_location="cpu")
    model.load_state_dict(state_dict)
    model.eval()
    return model


def dpt_sep_process(wav_path, model=None, outfilename=None):
    if model is None:
        model = load_dpt_model()

    x, sr = torchaudio.load(wav_path)
    x = x.cpu()

    with torch.no_grad():
        est_sources = model(x)  # shape: (1, 2, T)

    # 確保 est_sources 是 (1, 2, T)，再拆分
    est_sources = est_sources.squeeze(0)  # shape: (2, T)

    sep_1, sep_2 = est_sources  # 拆成兩個 (T, ) 的 tensor

    # 正規化
    max_abs = x[0].abs().max().item()
    sep_1 = sep_1 * max_abs / sep_1.abs().max().item()
    sep_2 = sep_2 * max_abs / sep_2.abs().max().item()

    # 增加 channel 維度，變為 (1, T)
    sep_1 = sep_1.unsqueeze(0)
    sep_2 = sep_2.unsqueeze(0)

    if outfilename is not None:
        torchaudio.save(outfilename.replace('.wav', '_sep1.wav'), sep_1, sr)
        torchaudio.save(outfilename.replace('.wav', '_sep2.wav'), sep_2, sr)
        torchaudio.save(outfilename.replace('.wav', '_mix.wav'), x, sr)
    else:
        torchaudio.save(wav_path.replace('.wav', '_sep1.wav'), sep_1, sr)
        torchaudio.save(wav_path.replace('.wav', '_sep2.wav'), sep_2, sr)


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