app.py

import os
import numpy as np
import tempfile
from scipy.io import wavfile
import gradio as gr
from inference import EnsembleDemucsMDXMusicSeparationModel, predict_with_model
import torch
import librosa
import librosa.display
import matplotlib.pyplot as plt

# 检查文件是否准备好
def check_file_readiness(filepath):
    num_same_size_checks = 0
    last_size = -1
    while num_same_size_checks < 5:
        current_size = os.path.getsize(filepath)
        if current_size == last_size:
            num_same_size_checks += 1
        else:
            num_same_size_checks = 0
            last_size = current_size
        time.sleep(0.5)
    return True

# 生成音频的频谱图
def generate_spectrogram(audio_file_path):
    y, sr = librosa.load(audio_file_path)
    plt.figure(figsize=(10, 4))
    S = librosa.feature.melspectrogram(y=y, sr=sr, n_mels=128, fmax=8000)
    librosa.display.specshow(librosa.power_to_db(S, ref=np.max),
                             y_axis='mel', fmax=8000, x_axis='time')
    plt.colorbar(format='%+2.0f dB')
    plt.title('Mel spectrogram')
    plt.tight_layout()
    image_path = tempfile.mktemp('.png')
    plt.savefig(image_path)
    plt.close()
    return image_path

# 为多个音频文件生成频谱图
def generate_spectrograms(audio_files):
    output_spectrograms = []
    for audio_file in audio_files:
        output_spectrograms.append(generate_spectrogram(audio_file))
    return tuple(output_spectrograms)

# 音乐分离的包装函数
def separate_music_file_wrapper(uploaded_files, use_cpu, use_single_onnx, large_overlap, small_overlap, chunk_size, use_large_gpu):
    input_files = [uploaded_files.name]  # 获取上传文件的路径

    # 验证重叠值
    if not (0 <= large_overlap <= 1) or not (0 <= small_overlap <= 1):
        raise ValueError("重叠值必须在0到1之间。")

    # 验证块大小
    if chunk_size <= 0:
        raise ValueError("块大小必须大于0。") # 太小了

    options = {
        'input_audio': input_files,
        'output_folder': 'results',
        'cpu': use_cpu,
        'single_onnx': use_single_onnx,
        'overlap_large': large_overlap,
        'overlap_small': small_overlap,
        'chunk_size': chunk_size,
        'large_gpu': use_large_gpu,
    }

    predict_with_model(options)

    # 清除GPU缓存
    if torch.cuda.is_available():
        torch.cuda.empty_cache()

    output_files = {}
    for f in input_files:
        audio_file_name = os.path.splitext(os.path.basename(f))[0]
        output_files["vocals"] = os.path.join(options['output_folder'], audio_file_name + "_vocals.wav")
        output_files["instrumental"] = os.path.join(options['output_folder'], audio_file_name + "_instrum.wav")
        output_files["instrumental2"] = os.path.join(options['output_folder'], audio_file_name + "_instrum2.wav") # 第二个伴奏输出
        output_files["bass"] = os.path.join(options['output_folder'], audio_file_name + "_bass.wav")
        output_files["drums"] = os.path.join(options['output_folder'], audio_file_name + "_drums.wav")
        output_files["other"] = os.path.join(options['output_folder'], audio_file_name + "_other.wav")

    # 检查文件是否准备好
    output_files_ready = []
    for k, v in output_files.items():
        if os.path.exists(v) and check_file_readiness(v):
            output_files_ready.append(v)
        else:
            empty_data = np.zeros((44100, 2)) # 2个声道，1秒的静音，采样率为44100Hz
            empty_file = tempfile.mktemp('.wav')
            wavfile.write(empty_file, 44100, empty_data.astype(np.int16))  # 转换为int16，因为wavfile不支持float32
            output_files_ready.append(empty_file)

    # 在分离音频后立即生成频谱图
    output_spectrograms = generate_spectrograms(output_files_ready)

    print(len(output_files_ready)) # 应输出6
    print(len(output_spectrograms)) # 应输出6

    print("返回前")
    return tuple(output_files_ready) + output_spectrograms
    print("返回后")

description = """
# ZFTurbo Web-UI
Web-UI by [Ma5onic](https://github.com/Ma5onic)
## 选项:
- **仅使用CPU:** 如果GPU内存不足，请选择此项。速度会较慢。
- **使用单一ONNX模型:** 选择此项以使用单一ONNX模型。会稍微降低质量，但可以帮助减少GPU内存使用。
- **大块重叠:** 大块的重叠。根据需要调整。
- **小块重叠:** 小块的重叠。根据需要调整。
- **块大小:** 每次处理的块大小。如果遇到内存问题，请减少此值。
- **使用快速大GPU版本:** 选择此项以使用旧的快速方法，需要超过11GB的GPU内存。运行速度更快。
"""
theme = gr.themes.Base(
    primary_hue="cyan",
    secondary_hue="cyan",
)

with gr.Blocks(theme=theme) as demo:
    gr.Markdown(description)
    uploaded_file = gr.File(label="上传音频文件", type="file")
    use_cpu = gr.Checkbox(label="仅使用CPU", value=True)
    use_single_onnx = gr.Checkbox(label="使用单一ONNX模型", value=False)
    large_overlap = gr.Number(label="大块重叠", value=0.6)
    small_overlap = gr.Number(label="小块重叠", value=0.5)
    chunk_size = gr.Number(label="块大小", value=1000000)
    use_large_gpu = gr.Checkbox(label="使用快速大GPU版本", value=False)    
    process_button = gr.Button("处理音频")

    vocals = gr.Audio(label="人声")
    vocals_spectrogram = gr.Image(label="人声频谱图")
    instrumental = gr.Audio(label="伴奏")
    instrumental_spectrogram = gr.Image(label="伴奏频谱图")
    instrumental2 = gr.Audio(label="伴奏2")
    instrumental2_spectrogram = gr.Image(label="伴奏2频谱图")
    bass = gr.Audio(label="贝斯")
    bass_spectrogram = gr.Image(label="贝斯频谱图")
    drums = gr.Audio(label="鼓声")
    drums_spectrogram = gr.Image(label="鼓声频谱图")
    other = gr.Audio(label="其他")
    other_spectrogram = gr.Image(label="其他频谱图")
    
    process_button.click(
        separate_music_file_wrapper,
        inputs=[uploaded_file, use_cpu, use_single_onnx, large_overlap, small_overlap, chunk_size, use_large_gpu],
        outputs=[vocals, instrumental, instrumental2, bass, drums, other, vocals_spectrogram, instrumental_spectrogram, instrumental2_spectrogram, bass_spectrogram, drums_spectrogram, other_spectrogram],
    )

demo.queue().launch(debug=True, share=False)