app.py

import os
import torch
import shutil
import librosa
import warnings
import numpy as np
import gradio as gr
import librosa.display
import matplotlib.pyplot as plt
from utils import get_modelist, find_audio_files, embed_img
from model import EvalNet


CLASSES = ["Gong", "Shang", "Jue", "Zhi", "Yu"]
TEMP_DIR = "./__pycache__/tmp"
SAMPLE_RATE = 44100


def zero_padding(y: np.ndarray, end: int):
    size = len(y)
    if size < end:
        return np.concatenate((y, np.zeros(end - size)))

    elif size > end:
        return y[-end:]

    return y


def audio2mel(audio_path: str, seg_len=20):
    y, sr = librosa.load(audio_path, sr=SAMPLE_RATE)
    y = zero_padding(y, seg_len * sr)
    mel_spec = librosa.feature.melspectrogram(y=y, sr=sr)
    log_mel_spec = librosa.power_to_db(mel_spec, ref=np.max)
    librosa.display.specshow(log_mel_spec)
    plt.axis("off")
    plt.savefig(
        f"{TEMP_DIR}/output.jpg",
        bbox_inches="tight",
        pad_inches=0.0,
    )
    plt.close()


def audio2cqt(audio_path: str, seg_len=20):
    y, sr = librosa.load(audio_path, sr=SAMPLE_RATE)
    y = zero_padding(y, seg_len * sr)
    cqt_spec = librosa.cqt(y=y, sr=sr)
    log_cqt_spec = librosa.power_to_db(np.abs(cqt_spec) ** 2, ref=np.max)
    librosa.display.specshow(log_cqt_spec)
    plt.axis("off")
    plt.savefig(
        f"{TEMP_DIR}/output.jpg",
        bbox_inches="tight",
        pad_inches=0.0,
    )
    plt.close()


def audio2chroma(audio_path: str, seg_len=20):
    y, sr = librosa.load(audio_path, sr=SAMPLE_RATE)
    y = zero_padding(y, seg_len * sr)
    chroma_spec = librosa.feature.chroma_stft(y=y, sr=sr)
    log_chroma_spec = librosa.power_to_db(np.abs(chroma_spec) ** 2, ref=np.max)
    librosa.display.specshow(log_chroma_spec)
    plt.axis("off")
    plt.savefig(
        f"{TEMP_DIR}/output.jpg",
        bbox_inches="tight",
        pad_inches=0.0,
    )
    plt.close()


def infer(wav_path: str, log_name: str, folder_path=TEMP_DIR):
    if os.path.exists(folder_path):
        shutil.rmtree(folder_path)

    if not wav_path:
        return None, "Please input an audio!"

    spec = log_name.split("_")[-3]
    os.makedirs(folder_path, exist_ok=True)
    try:
        model = EvalNet(log_name, len(CLASSES)).model
        eval("audio2%s" % spec)(wav_path)

    except Exception as e:
        return None, f"{e}"

    input = embed_img(f"{folder_path}/output.jpg")
    output: torch.Tensor = model(input)
    pred_id = torch.max(output.data, 1)[1]
    return (
        os.path.basename(wav_path),
        CLASSES[pred_id].capitalize(),
    )


if __name__ == "__main__":
    warnings.filterwarnings("ignore")
    models = get_modelist(assign_model="vit_l_16_cqt")
    examples = []
    example_audios = find_audio_files()
    for audio in example_audios:
        examples.append([audio, models[0]])

    with gr.Blocks() as demo:
        gr.Interface(
            fn=infer,
            inputs=[
                gr.Audio(label="Upload a recording", type="filepath"),
                gr.Dropdown(choices=models, label="Select a model", value=models[0]),
            ],
            outputs=[
                gr.Textbox(label="Audio filename", show_copy_button=True),
                gr.Textbox(
                    label="Chinese pentatonic mode recognition",
                    show_copy_button=True,
                ),
            ],
            examples=examples,
            cache_examples=False,
            flagging_mode="never",
            title="It is recommended to keep the recording length around 20s.",
        )

        gr.Markdown(
            """
# Cite
```bibtex
@article{Zhou-2025,
  title   = {CCMusic: an Open and Diverse Database for Chinese Music Information Retrieval Research},
  author  = {Monan Zhou and Shenyang Xu and Zhaorui Liu and Zhaowen Wang and Feng Yu and Wei Li and Baoqiang Han},
  journal = {Transactions of the International Society for Music Information Retrieval},
  year    = {2025}
}
```"""
        )

    demo.launch()