Spaces:
Running
Running
import os | |
import torch | |
import librosa | |
import binascii | |
import warnings | |
import midi2audio # MIDI ํ์ผ์ WAV ํ์ผ๋ก ๋ณํ | |
import numpy as np | |
import pytube as pt # YouTube ๋น๋์ค๋ฅผ ์ค๋์ค๋ก ๋ค์ด๋ก๋ | |
import gradio as gr | |
import soundfile as sf | |
from transformers import Pop2PianoForConditionalGeneration, Pop2PianoProcessor | |
# ๋๋ ํ ๋ฆฌ ์์ฑ | |
yt_video_dir = "./yt_dir" # ์ ํ๋ธ ๋น๋์ค ๋ค์ด๋ก๋ ๊ฒฝ๋ก | |
outputs_dir = "./midi_wav_outputs" # ์ถ๋ ฅ ํ์ผ ๊ฒฝ๋ก | |
os.makedirs(outputs_dir, exist_ok=True) | |
os.makedirs(yt_video_dir, exist_ok=True) | |
# ๋ชจ๋ธ ์ค์ | |
device = "cuda" if torch.cuda.is_available() else "cpu" | |
model = Pop2PianoForConditionalGeneration.from_pretrained("sweetcocoa/pop2piano").to(device) | |
processor = Pop2PianoProcessor.from_pretrained("sweetcocoa/pop2piano") | |
composers = model.generation_config.composer_to_feature_token.keys() | |
# ์ ํ๋ธ ๋น๋์ค์์ ์ค๋์ค ์ถ์ถ ํจ์ | |
def get_audio_from_yt_video(yt_link): | |
try: | |
yt = pt.YouTube(yt_link) | |
t = yt.streams.filter(only_audio=True) | |
filename = os.path.join(yt_video_dir, binascii.hexlify(os.urandom(8)).decode() + ".mp4") | |
t[0].download(filename=filename) | |
except: | |
warnings.warn(f"Video Not Found at {yt_link}") | |
filename = None | |
return filename, filename | |
# ๋ชจ๋ธ ์ถ๋ก ํจ์ | |
def inference(file_uploaded, composer): | |
waveform, sr = librosa.load(file_uploaded, sr=None) | |
inputs = processor(audio=waveform, sampling_rate=sr, return_tensors="pt").to(device) | |
model_output = model.generate(input_features=inputs["input_features"], composer=composer) | |
tokenizer_output = processor.batch_decode(token_ids=model_output.to("cpu"), feature_extractor_output=inputs.to("cpu"))["pretty_midi_objects"] | |
return prepare_output_file(tokenizer_output, sr) | |
# ์ถ๋ ฅ ํ์ผ ์ค๋น ํจ์ | |
def prepare_output_file(tokenizer_output, sr): | |
output_file_name = "output_" + binascii.hexlify(os.urandom(8)).decode() | |
midi_output = os.path.join(outputs_dir, output_file_name + ".mid") | |
tokenizer_output[0].write(midi_output) | |
wav_output = midi_output.replace(".mid", ".wav") | |
midi2audio.FluidSynth().midi_to_audio(midi_output, wav_output) | |
return wav_output, wav_output, midi_output | |
# Gradio UI ์ค์ | |
block = gr.Blocks(theme="Taithrah/Minimal") | |
with block: | |
gr.HTML( | |
""" | |
<div style="text-align: center; max-width: 800px; margin: 0 auto;"> | |
<h1 style="font-weight: 900; margin-bottom: 12px;"> | |
๐น Pop2Piano : ํผ์๋ ธ ์ปค๋ฒ๊ณก ์์ฑ๊ธฐ ๐น | |
</h1> | |
<p style="margin-bottom: 12px; font-size: 90%"> | |
Pop2Piano ๋ฐ๋ชจ: ํ ์ค๋์ค ๊ธฐ๋ฐ ํผ์๋ ธ ์ปค๋ฒ๊ณก ์์ฑ. <br> | |
์๊ณก๊ฐ(ํธ๊ณก์)๋ฅผ ์ ํํ๊ณ ํ ์ค๋์ค๋ฅผ ์ ๋ก๋ํ๊ฑฐ๋ ์ ํ๋ธ ๋งํฌ๋ฅผ ์ ๋ ฅํ ํ ์์ฑ ๋ฒํผ์ ํด๋ฆญํ์ธ์. | |
</p> | |
</div> | |
""" | |
) | |
with gr.Group(): | |
with gr.Row(): | |
with gr.Column(): | |
file_uploaded = gr.Audio(label="์ค๋์ค ์ ๋ก๋", type="filepath") | |
with gr.Column(): | |
with gr.Row(): | |
yt_link = gr.Textbox(label="์ ํ๋ธ ๋งํฌ๋ฅผ ์ ๋ ฅํ์ธ์.", autofocus=True, lines=3) | |
yt_btn = gr.Button("์ ํ๋ธ ๋งํฌ์์ ์ค๋์ค๋ฅผ ๋ค์ด ๋ฐ์ต๋๋ค.", size="lg") | |
yt_audio_path = gr.Audio(label="์ ํ๋ธ ๋์์์์ ์ถ์ถํ ์ค๋์ค", interactive=False) | |
yt_btn.click(get_audio_from_yt_video, inputs=[yt_link], outputs=[yt_audio_path, file_uploaded]) | |
with gr.Group(): | |
with gr.Column(): | |
composer = gr.Dropdown(label="ํธ๊ณก์", choices=composers, value="composer1") | |
generate_btn = gr.Button("๋๋ง์ ํผ์๋ ธ ์ปค๋ฒ๊ณก ๋ง๋ค๊ธฐ๐น๐ต") | |
with gr.Row(): | |
wav_output2 = gr.File(label="๋๋ง์ ํผ์๋ ธ ์ปค๋ฒ๊ณก์ ๋ค์ด๋ก๋ (.wav)") | |
wav_output1 = gr.Audio(label="๋๋ง์ ํผ์๋ ธ ์ปค๋ฒ๊ณก ๋ฃ๊ธฐ") | |
midi_output = gr.File(label="์์ฑํ midi ํ์ผ ๋ค์ด๋ก๋ (.mid)") | |
generate_btn.click( | |
inference, | |
inputs=[file_uploaded, composer], | |
outputs=[wav_output1, wav_output2, midi_output]) | |
block.launch(debug=False) | |