app.py

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


TRANSLATE = {
    "vibrato": "Rou xian",
    "trill": "Chan yin",
    "tremolo": "Chan gong",
    "staccato": "Dun gong",
    "ricochet": "Pao gong",
    "pizzicato": "Bo xian",
    "percussive": "Ji gong",
    "legato_slide_glissando": "Lian hua yin",
    "harmonic": "Fan yin",
    "diangong": "Dian gong",
    "detache": "Fen gong",
}
CLASSES = list(TRANSLATE.keys())
TEMP_DIR = "./__pycache__/tmp"
SAMPLE_RATE = 44100


def circular_padding(y: np.ndarray, sr: int, dur=3):
    if len(y) >= sr * dur:
        return y[: sr * dur]

    size = sr * dur // len(y) + int((sr * dur) % len(y) > 0)
    arrays = []
    for _ in range(size):
        arrays.append(y)

    y = np.hstack(arrays)
    return y[: sr * dur]


def wav2mel(audio_path: str):
    os.makedirs(TEMP_DIR, exist_ok=True)
    try:
        y, sr = librosa.load(audio_path, sr=SAMPLE_RATE)
        y = circular_padding(y, 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()

    except Exception as e:
        print(f"Error converting {audio_path} : {e}")


def wav2cqt(audio_path: str):
    os.makedirs(TEMP_DIR, exist_ok=True)
    try:
        y, sr = librosa.load(audio_path, sr=SAMPLE_RATE)
        y = circular_padding(y, 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()

    except Exception as e:
        print(f"Error converting {audio_path} : {e}")


def wav2chroma(audio_path: str):
    os.makedirs(TEMP_DIR, exist_ok=True)
    try:
        y, sr = librosa.load(audio_path, sr=SAMPLE_RATE)
        y = circular_padding(y, 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()

    except Exception as e:
        print(f"Error converting {audio_path} : {e}")


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!"

    try:
        model = EvalNet(log_name, len(TRANSLATE)).model
    except Exception as e:
        return None, f"{e}"

    spec = log_name.split("_")[-3]
    eval("wav2%s" % spec)(wav_path)
    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),
        f"{TRANSLATE[CLASSES[pred_id]]} ({CLASSES[pred_id].capitalize()})",
    )


if __name__ == "__main__":
    warnings.filterwarnings("ignore")
    models = get_modelist()
    examples = []
    example_wavs = find_wav_files()
    model_num = len(models)
    for wav in example_wavs:
        examples.append([wav, models[random.randint(0, model_num - 1)]])

    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="Playing tech recognition", show_copy_button=True),
            ],
            examples=examples,
            cache_examples=False,
            allow_flagging="never",
            title="It is recommended to keep the recording length around 3s.",
        )

        gr.Markdown(
            """
# Cite
```bibtex
@dataset{zhaorui_liu_2021_5676893,
  author       = {Monan Zhou, Shenyang Xu, Zhaorui Liu, Zhaowen Wang, Feng Yu, Wei Li and Baoqiang Han},
  title        = {CCMusic: an Open and Diverse Database for Chinese Music Information Retrieval Research},
  month        = {mar},
  year         = {2024},
  publisher    = {HuggingFace},
  version      = {1.2},
  url          = {https://huggingface.co/ccmusic-database}
}
```"""
        )

    demo.launch()