app.py

import os
import argparse
import gradio as gr
from timeit import default_timer as timer
import torch
import numpy as np
import pandas as pd
from huggingface_hub import hf_hub_download
from model.bart import BartCaptionModel
from utils.audio_utils import load_audio, STR_CH_FIRST

if os.path.isfile("transfer.pth") == False:
    torch.hub.download_url_to_file('https://huggingface.co/seungheondoh/lp-music-caps/resolve/main/transfer.pth', 'transfer.pth')
    torch.hub.download_url_to_file('https://huggingface.co/seungheondoh/lp-music-caps/resolve/main/folk.wav', 'folk.wav')
    torch.hub.download_url_to_file('https://huggingface.co/seungheondoh/lp-music-caps/resolve/main/electronic.mp3', 'electronic.mp3')
    torch.hub.download_url_to_file('https://huggingface.co/seungheondoh/lp-music-caps/resolve/main/orchestra.wav', 'orchestra.wav')

device = "cuda:0" if torch.cuda.is_available() else "cpu"

example_list = ['folk.wav', 'electronic.mp3', 'orchestra.wav']
model = BartCaptionModel(max_length = 128)
pretrained_object = torch.load('./transfer.pth', map_location='cpu')
state_dict = pretrained_object['state_dict']
model.load_state_dict(state_dict)
if torch.cuda.is_available():
    torch.cuda.set_device(device)
    model = model.cuda(device)
model.eval()

def get_audio(audio_path, duration=10, target_sr=16000):
    n_samples = int(duration * target_sr)
    audio, sr = load_audio(
        path= audio_path,
        ch_format= STR_CH_FIRST,
        sample_rate= target_sr,
        downmix_to_mono= True,
    )
    if len(audio.shape) == 2:
        audio = audio.mean(0, False)  # to mono
    input_size = int(n_samples)
    if audio.shape[-1] < input_size:  # pad sequence
        pad = np.zeros(input_size)
        pad[: audio.shape[-1]] = audio
        audio = pad
    ceil = int(audio.shape[-1] // n_samples)
    audio = torch.from_numpy(np.stack(np.split(audio[:ceil * n_samples], ceil)).astype('float32'))
    return audio

def captioning(audio_path):
    audio_tensor = get_audio(audio_path = audio_path)
    if torch.cuda.is_available():
        audio_tensor = audio_tensor.to(device)
    with torch.no_grad():
        output = model.generate(
            samples=audio_tensor,
            num_beams=5,
        )
    inference = ""
    number_of_chunks = range(audio_tensor.shape[0])
    for chunk, text in zip(number_of_chunks, output):
        time = f"[{chunk * 10}:00-{(chunk + 1) * 10}:00]"
        inference += f"{time}\n{text} \n \n"
    return inference

title = "Interactive demo: Music Captioning 🤖🎵"
description = """
<p style='text-align: center'> LP-MusicCaps: LLM-Based Pseudo Music Captioning</p> 
<p style='text-align: center'> SeungHeon Doh, Keunwoo Choi, Jongpil Lee, Juhan Nam, ISMIR 2023</p> 
<p style='text-align: center'> <a href='https://arxiv.org/abs/2307.16372' target='_blank'>ArXiv</a> | <a href='https://github.com/seungheondoh/lp-music-caps' target='_blank'>Codes</a> | <a href='https://huggingface.co/datasets/seungheondoh/LP-MusicCaps-MC' target='_blank'>Dataset</a> </p>
<p style='text-align: center'> To use it, simply upload your audio and click 'submit', or click one of the examples to load them. Read more at the links below. </p>
<p style='text-align: center'> If you have any error, plz check this code: <a href='https://github.com/seungheondoh/lp-music-caps/blob/main/demo/app.py' target='_blank'>Demo</a>. </p>
"""

article = "<p style='text-align: center'><a href='https://seungheondoh.github.io/' target='_blank'>Author Info</a> | <a href='https://github.com/seungheondoh' target='_blank'>Github</a></p>"


demo = gr.Interface(fn=captioning,
                    inputs=gr.Audio(type="filepath"),
                    outputs=[
                        gr.Textbox(label="Caption generated by LP-MusicCaps Transfer Model"),
                        ],
                    examples=example_list,
                    title=title,
                    description=description,
                    article=article, 
                    cache_examples=False
                    )
demo.launch()