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
import torchvision.transforms as transforms
from collections import Counter
from PIL import Image
from tqdm import tqdm
from model import net, MODEL_DIR


MODEL = net()


def most_common_element(input_list):
    counter = Counter(input_list)
    mce, _ = counter.most_common(1)[0]
    return mce


def wav_to_mel(audio_path: str, width=0.18):
    os.makedirs("./tmp")
    try:
        y, sr = librosa.load(audio_path, sr=48000)
        non_silent = y
        mel_spec = librosa.feature.melspectrogram(y=non_silent, sr=sr)
        log_mel_spec = librosa.power_to_db(mel_spec, ref=np.max)
        dur = librosa.get_duration(y=non_silent, sr=sr)
        total_frames = log_mel_spec.shape[1]
        step = int(width * total_frames / dur)
        count = int(total_frames / step)
        begin = int(0.5 * (total_frames - count * step))
        end = begin + step * count
        for i in tqdm(range(begin, end, step), desc="Converting wav to jpgs..."):
            librosa.display.specshow(log_mel_spec[:, i : i + step])
            plt.axis("off")
            plt.savefig(
                f"./tmp/{os.path.basename(audio_path)[:-4]}_{i}.jpg",
                bbox_inches="tight",
                pad_inches=0.0,
            )
            plt.close()

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


def embed_img(img_path, input_size=224):
    transform = transforms.Compose(
        [
            transforms.Resize([input_size, input_size]),
            transforms.ToTensor(),
            transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
        ]
    )
    img = Image.open(img_path).convert("RGB")
    return transform(img).unsqueeze(0)


def inference(wav_path, folder_path="./tmp"):
    if os.path.exists(folder_path):
        shutil.rmtree(folder_path)

    if not wav_path:
        return None, "请输入音频 Please input an audio!"

    wav_to_mel(wav_path)
    outputs = []
    all_files = os.listdir(folder_path)
    for file_name in all_files:
        if file_name.lower().endswith(".jpg"):
            file_path = os.path.join(folder_path, file_name)
            input = embed_img(file_path)
            output: torch.Tensor = MODEL(input)
            pred_id = torch.max(output.data, 1)[1]
            outputs.append(pred_id)

    max_count_item = most_common_element(outputs)
    shutil.rmtree(folder_path)
    return os.path.basename(wav_path), translate[classes[max_count_item]]


if __name__ == "__main__":
    warnings.filterwarnings("ignore")
    translate = {
        "PearlRiver": "Pearl River",
        "YoungChang": "YOUNG CHANG",
        "Steinway-T": "STEINWAY Theater",
        "Hsinghai": "HSINGHAI",
        "Kawai": "KAWAI",
        "Steinway": "STEINWAY",
        "Kawai-G": "KAWAI Grand",
        "Yamaha": "YAMAHA",
    }
    classes = list(translate.keys())
    example_wavs = []
    for cls in classes:
        example_wavs.append(f"{MODEL_DIR}/examples/{cls}.wav")

    with gr.Blocks() as demo:
        gr.Interface(
            fn=inference,
            inputs=gr.Audio(
                type="filepath", label="上传钢琴录音 Upload a piano recording"
            ),
            outputs=[
                gr.Textbox(label="音频文件名 Audio filename", show_copy_button=True),
                gr.Textbox(
                    label="钢琴分类结果 Piano classification result",
                    show_copy_button=True,
                ),
            ],
            examples=example_wavs,
            cache_examples=False,
            allow_flagging="never",
            title="建议录音时长保持在 3s 左右, 过长会影响识别效率<br>It is recommended to keep the duration of recording around 3s, too long will affect the recognition efficiency.",
        )

        gr.Markdown(
            """
# 引用 Cite
```bibtex
@article{Zhou2023AHE,
  author    = {Monan Zhou and Shangda Wu and Shaohua Ji and Zijin Li and Wei Li},
  title     = {A Holistic Evaluation of Piano Sound Quality},
  booktitle = {Proceedings of the 10th Conference on Sound and Music Technology (CSMT)},
  year      = {2023},
  publisher = {Springer Singapore},
  address   = {Singapore}
}
```"""
        )

    demo.launch()