app.py

import os
from pathlib import Path
from huggingface_hub import hf_hub_download
import gradio as gr
from scipy.io.wavfile import write
import torch
from utils import convert_to_stereo_and_wav
from uvr_processing import get_model_params, run_mdx


MODEL_ID = "masszhou/mdxnet"
MODELS_PATH = {
    "bgm": Path(hf_hub_download(repo_id=MODEL_ID, filename="UVR-MDX-NET-Inst_HQ_3.onnx")), 
    "basic_vocal": Path(hf_hub_download(repo_id=MODEL_ID, filename="UVR-MDX-NET-Voc_FT.onnx")),
    "main_vocal": Path(hf_hub_download(repo_id=MODEL_ID, filename="UVR_MDXNET_KARA_2.onnx"))
}


def inference_mdx(audio_file: str) -> list[str]:
    mdx_model_params = get_model_params(Path("./mdx_models"))
    audio_file = convert_to_stereo_and_wav(Path(audio_file))  # resampling at 44100 Hz
    device_base = "cuda" if torch.cuda.is_available() else "cpu"
    output_dir = Path("./out/mdx")
    os.makedirs(output_dir, exist_ok=True)
    model_bgm_path = MODELS_PATH["bgm"]
    background_path, vocal_path = run_mdx(model_params=mdx_model_params,
                                input_filename=audio_file,
                                output_dir=output_dir,
                                model_path=model_bgm_path,
                                denoise=False,
                                device_base=device_base,
                                )

    return str(vocal_path), str(background_path)


def inference_demucs(audio):
    sr = audio[0]
    audio_np = audio[1]
    os.makedirs("out", exist_ok=True)
    write('test.wav', audio[0], audio[1])
    os.system("python3 -m demucs.separate -n htdemucs --two-stems=vocals test.wav -o out")
    return "./out/htdemucs/test/vocals.wav","./out/htdemucs/test/no_vocals.wav"


if __name__ == "__main__":
    tab_1 = gr.Interface(
        fn = inference_demucs, 
        inputs = gr.Audio(type="numpy", label="Input"),
        outputs = [gr.Audio(type="filepath", label="Vocals"),gr.Audio(type="filepath", label="BGM")],
        title="Demucs Music Source Separation (v4)",
        article="<p style='text-align: center'><a href='https://arxiv.org/abs/1911.13254' target='_blank'>Music Source Separation in the Waveform Domain</a> | <a href='https://github.com/facebookresearch/demucs' target='_blank'>Github Repo</a> | <a href='https://github.com/facebookresearch/demucs/blob/main/LICENSE' target='_blank'>MIT License</a></p>",
        api_name="demucs_separation",
        )
    tab_2 = gr.Interface(
        fn = inference_mdx, 
        inputs = gr.Audio(type="filepath", label="Input"),
        outputs = [gr.Audio(type="filepath", label="Vocals"),gr.Audio(type="filepath", label="BGM")],
        title="MDXNET Music Source Separation",
        article="<p style='text-align: center'><a href='https://arxiv.org/abs/2111.12203' target='_blank'>KUIELab-MDX-Net: A Two-Stream Neural Network for Music Demixing</a> | <a href='https://github.com/kuielab/mdx-net' target='_blank'>Github Repo</a> | <a href='https://github.com/kuielab/mdx-net/blob/main/LICENSE' target='_blank'>MIT License</a></p>",
        api_name="mdxnet_separation",
    )
    demo = gr.TabbedInterface([tab_1, tab_2], ["Demucs", "MDXNET"])
    demo.launch()