app.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.

# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

# Updated to account for UI changes from https://github.com/rkfg/audiocraft/blob/long/app.py
# also released under the MIT license.

import argparse
from concurrent.futures import ProcessPoolExecutor
import os
import subprocess as sp
from tempfile import NamedTemporaryFile
import time
import warnings

import torch
import gradio as gr

from audiocraft.data.audio_utils import convert_audio
from audiocraft.data.audio import audio_write
from audiocraft.models import MusicGen


MODEL = None  # Last used model
IS_BATCHED = "facebook/MusicGen" in os.environ.get('SPACE_ID', '')
MAX_BATCH_SIZE = 6
BATCHED_DURATION = 15
INTERRUPTING = False
# We have to wrap subprocess call to clean a bit the log when using gr.make_waveform
_old_call = sp.call


def _call_nostderr(*args, **kwargs):
    # Avoid ffmpeg vomitting on the logs.
    kwargs['stderr'] = sp.DEVNULL
    kwargs['stdout'] = sp.DEVNULL
    _old_call(*args, **kwargs)


sp.call = _call_nostderr
# Preallocating the pool of processes.
pool = ProcessPoolExecutor(3)
pool.__enter__()


def interrupt():
    global INTERRUPTING
    INTERRUPTING = True


def make_waveform(*args, **kwargs):
    # Further remove some warnings.
    be = time.time()
    with warnings.catch_warnings():
        warnings.simplefilter('ignore')
        out = gr.make_waveform(*args, **kwargs)
        print("Make a video took", time.time() - be)
        return out


def load_model(version='melody'):
    global MODEL
    print("Loading model", version)
    if MODEL is None or MODEL.name != version:
        MODEL = MusicGen.get_pretrained(version)


def _do_predictions(texts, melodies, duration, progress=False, **gen_kwargs):
    MODEL.set_generation_params(duration=duration, **gen_kwargs)
    print("new batch", len(texts), texts, [None if m is None else (m[0], m[1].shape) for m in melodies])
    be = time.time()
    processed_melodies = []
    target_sr = 32000
    target_ac = 1
    for melody in melodies:
        if melody is None:
            processed_melodies.append(None)
        else:
            sr, melody = melody[0], torch.from_numpy(melody[1]).to(MODEL.device).float().t()
            if melody.dim() == 1:
                melody = melody[None]
            melody = melody[..., :int(sr * duration)]
            melody = convert_audio(melody, sr, target_sr, target_ac)
            processed_melodies.append(melody)

    if any(m is not None for m in processed_melodies):
        outputs = MODEL.generate_with_chroma(
            descriptions=texts,
            melody_wavs=processed_melodies,
            melody_sample_rate=target_sr,
            progress=progress,
        )
    else:
        outputs = MODEL.generate(texts, progress=progress)

    outputs = outputs.detach().cpu().float()
    out_files = []
    for output in outputs:
        with NamedTemporaryFile("wb", suffix=".wav", delete=False) as file:
            audio_write(
                file.name, output, MODEL.sample_rate, strategy="loudness",
                loudness_headroom_db=16, loudness_compressor=True, add_suffix=False)
            out_files.append(pool.submit(make_waveform, file.name))
    res = [out_file.result() for out_file in out_files]
    print("batch finished", len(texts), time.time() - be)
    return res


def predict_batched(texts, melodies):
    max_text_length = 512
    texts = [text[:max_text_length] for text in texts]
    load_model('melody')
    res = _do_predictions(texts, melodies, BATCHED_DURATION)
    return [res]


def predict_full(model, text, melody, duration, topk, topp, temperature, cfg_coef, progress=gr.Progress()):
    global INTERRUPTING
    INTERRUPTING = False
    topk = int(topk)
    load_model(model)

    def _progress(generated, to_generate):
        progress((generated, to_generate))
        if INTERRUPTING:
            raise gr.Error("Interrupted.")
    MODEL.set_custom_progress_callback(_progress)

    outs = _do_predictions(
        [text], [melody], duration, progress=True,
        top_k=topk, top_p=topp, temperature=temperature, cfg_coef=cfg_coef)
    return outs[0]


def ui_full(launch_kwargs):
    with gr.Blocks() as interface:
        gr.Markdown(
            """
          
            """
        )


block = gr.Blocks(css=".wrap svelte-1lyswbr{max-width:550px}.svelte-10ogue4 {background: rgba(0,0,0,0.0);width:100%;border: 0px}.gradio-container {background: rgba(0,0,0,0.0);border: 0px}.gr-block {background: rgba(0,0,0,0.0);border: 0px} #component-4 {opacity: 0.8;background: linear-gradient(#233581, #E23F9C);border: 0px;margin-bottom: 17px;border-radius: 10px;}#component-5 {opacity: 0.8;background: linear-gradient(#E23F9C, #233581);border: 0px;margin-bottom: 17px;border-radius: 10px;}.gr-form{background: rgba(0,0,0,0.0);border: 0px}.gr-text-input {background: rgba(255,255,255,1);border: 0px}.text-gray-500 {color: #FFFFFF;font-weight: 600;text-align: center;font-size: 18px;}#component-1 {height: 0px;}#range_id_0 {opacity: 0.5;border-radius: 8px;-webkit-appearance: none; width: 60%; height: 15px; background-color: #E64CAC; background-image: -webkit-gradient(linear, 0 0, 0 100%, from(#233581, #E23F9C), to(#E956B8)); background-image: -webkit-linear-gradient(right, #233581, #E956B8); background-image: -moz-linear-gradient(right, #233581, #E956B8); background-image: -ms-linear-gradient(right, #233581, #E956B8); background-image: -o-linear-gradient(right, #233581, #E956B8)}#component-6{opacity: 0.9;background: linear-gradient(#233581, #515A7F);border-radius: 10px}#component-7{margin-top: 7px;margin-bottom: 7px;text-align: center;display:inline;opacity: 0.9;background: linear-gradient(#515A7F, #515A7F);border-radius: 10px;}.ml-2{color: #FFFFFF;}#component-8 {height: 100px;z-index:99;background: linear-gradient(#515A7F, #515A7F);border-radius: 10px;opacity: 0.9}.absolute{background: linear-gradient(#EC5CC0, #D61B70);border: 0px}.feather{color: #FFFFFF;} .mt-7{z-index:100;background: linear-gradient(#515A7F, #515A7F);border-radius: 10px;} .gr-button{margin-left: 30%;width:40%;justify-content: center; background: linear-gradient(#EC5DC1, #D61A6F); padding: 0 12px; border: none; border-radius: 8px; box-shadow: 0 30px 15px rgba(0, 0, 0, 0.15); outline: none; color: #FFF; font: 400 16px/2.5 Nunito, Sans-serif; text-transform: uppercase; cursor: pointer;}#component-11{justify-content: center;text-align: center;margin-top:10px;border: 0px}.mx-auto{background: rgba(0,0,0,0.0);width:100%;border: 0px;padding:0 0 0 0}#component-9 {margin-top: 5px;opacity: 0.8;padding: 3px;background: linear-gradient(#515A7F, #515A7F);border-radius: 10px;}#component-10{margin-top: 5px;opacity: 0.8;padding: 3px;background: linear-gradient(#515A7F, #515A7F);border-radius: 10px;}#component-12{display:none}.gr-input-label{margin-right: 1px;width:71px;font-weight: 400;background: linear-gradient(#584C84, #2C3D7F);text-align: center;border: 0px}.font-semibold{display:none}")

        
        with gr.Row():
            with gr.Column():
                with gr.Row():
                    text = gr.Text(label="Текст пример (bass drum cyberpunk)", interactive=True)
                    melody = gr.Audio(source="upload", type="numpy", label="Melody Condition (optional)", interactive=True)
                with gr.Row():
                    submit = gr.Button("Создать")
                    # Adapted from https://github.com/rkfg/audiocraft/blob/long/app.py, MIT license.
                    _ = gr.Button("Остановить").click(fn=interrupt, queue=False)
                with gr.Row():
                    model = gr.Radio(["melody", "medium", "small", "large"], label="Тип трека", value="melody", interactive=True)
                with gr.Row():
                    duration = gr.Slider(minimum=1, maximum=120, value=10, label="Время трека(seconds)", interactive=True)
                with gr.Row():
                    topk = gr.Number(label="Top-k", value=250, interactive=True)
                    topp = gr.Number(label="Top-p", value=0, interactive=True)
                    temperature = gr.Number(label="Temperature", value=1.0, interactive=True)
                    cfg_coef = gr.Number(label="Classifier Free Guidance", value=3.0, interactive=True)
            with gr.Column():
                output = gr.Video(label="Generated Music")
        submit.click(predict_full, inputs=[model, text, melody, duration, topk, topp, temperature, cfg_coef], outputs=[output])
  
        gr.Markdown(
            """
        
            """
        )

        interface.queue().launch(**launch_kwargs)


def ui_batched(launch_kwargs):
    with gr.Blocks() as demo:
        gr.Markdown(
            """
          
            """
        )
        with gr.Row():
            with gr.Column():
                with gr.Row():
                    text = gr.Text(label="Describe your music", lines=2, interactive=True)
                    melody = gr.Audio(source="upload", type="numpy", label="Condition on a melody (optional)", interactive=True)
                with gr.Row():
                    submit = gr.Button("Generate")
            with gr.Column():
                output = gr.Video(label="Generated Music")
        submit.click(predict_batched, inputs=[text, melody], outputs=[output], batch=True, max_batch_size=MAX_BATCH_SIZE)
 
        gr.Markdown("""
    
        """)

        demo.queue(max_size=8 * 4).launch(**launch_kwargs)


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument(
        '--listen',
        type=str,
        default='0.0.0.0' if 'SPACE_ID' in os.environ else '127.0.0.1',
        help='IP to listen on for connections to Gradio',
    )
    parser.add_argument(
        '--username', type=str, default='', help='Username for authentication'
    )
    parser.add_argument(
        '--password', type=str, default='', help='Password for authentication'
    )
    parser.add_argument(
        '--server_port',
        type=int,
        default=0,
        help='Port to run the server listener on',
    )
    parser.add_argument(
        '--inbrowser', action='store_true', help='Open in browser'
    )
    parser.add_argument(
        '--share', action='store_true', help='Share the gradio UI'
    )

    args = parser.parse_args()

    launch_kwargs = {}
    launch_kwargs['server_name'] = args.listen

    if args.username and args.password:
        launch_kwargs['auth'] = (args.username, args.password)
    if args.server_port:
        launch_kwargs['server_port'] = args.server_port
    if args.inbrowser:
        launch_kwargs['inbrowser'] = args.inbrowser
    if args.share:
        launch_kwargs['share'] = args.share

    # Show the interface
    if IS_BATCHED:
        ui_batched(launch_kwargs)
    else:
        ui_full(launch_kwargs)