import argparse from concurrent.futures import ProcessPoolExecutor import os from pathlib import Path import subprocess as sp from tempfile import NamedTemporaryFile import time import typing as tp 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, MultiBandDiffusion MODEL = None # Last used model IS_BATCHED = "facebook/MusicGen" in os.environ.get('SPACE_ID', '') print(IS_BATCHED) MAX_BATCH_SIZE = 12 BATCHED_DURATION = 15 INTERRUPTING = False MBD = None # 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 vomiting 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(4) pool.__enter__() def interrupt(): global INTERRUPTING INTERRUPTING = True class FileCleaner: def __init__(self, file_lifetime: float = 3600): self.file_lifetime = file_lifetime self.files = [] def add(self, path: tp.Union[str, Path]): self._cleanup() self.files.append((time.time(), Path(path))) def _cleanup(self): now = time.time() for time_added, path in list(self.files): if now - time_added > self.file_lifetime: if path.exists(): path.unlink() self.files.pop(0) else: break file_cleaner = FileCleaner() 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='facebook/musicgen-melody'): global MODEL print("Loading model", version) if MODEL is None or MODEL.name != version: MODEL = MusicGen.get_pretrained(version) def load_diffusion(): global MBD if MBD is None: print("loading MBD") MBD = MultiBandDiffusion.get_mbd_musicgen() 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, return_tokens=USE_DIFFUSION ) else: outputs = MODEL.generate(texts, progress=progress, return_tokens=USE_DIFFUSION) if USE_DIFFUSION: outputs_diffusion = MBD.tokens_to_wav(outputs[1]) outputs = torch.cat([outputs[0], outputs_diffusion], dim=0) outputs = outputs.detach().cpu().float() pending_videos = [] out_wavs = [] 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) pending_videos.append(pool.submit(make_waveform, file.name)) out_wavs.append(file.name) file_cleaner.add(file.name) out_videos = [pending_video.result() for pending_video in pending_videos] for video in out_videos: file_cleaner.add(video) print("batch finished", len(texts), time.time() - be) print("Tempfiles currently stored: ", len(file_cleaner.files)) return out_videos, out_wavs def predict_batched(texts, melodies): max_text_length = 512 texts = [text[:max_text_length] for text in texts] load_model('facebook/musicgen-melody') res = _do_predictions(texts, melodies, BATCHED_DURATION) return res def predict_full(model, decoder, text, melody, duration, topk, topp, temperature, cfg_coef, progress=gr.Progress()): global INTERRUPTING global USE_DIFFUSION INTERRUPTING = False if temperature < 0: raise gr.Error("Temperature must be >= 0.") if topk < 0: raise gr.Error("Topk must be non-negative.") if topp < 0: raise gr.Error("Topp must be non-negative.") topk = int(topk) if decoder == "MultiBand_Diffusion": USE_DIFFUSION = True load_diffusion() else: USE_DIFFUSION = False load_model(model) def _progress(generated, to_generate): progress((min(generated, to_generate), to_generate)) if INTERRUPTING: raise gr.Error("Interrupted.") MODEL.set_custom_progress_callback(_progress) videos, wavs = _do_predictions( [text], [melody], duration, progress=True, top_k=topk, top_p=topp, temperature=temperature, cfg_coef=cfg_coef) if USE_DIFFUSION: return videos[0], wavs[0], videos[1], wavs[1] return videos[0], wavs[0], None, None def toggle_audio_src(choice): if choice == "mic": return gr.update(source="microphone", value=None, label="Microphone") else: return gr.update(source="upload", value=None, label="File") def toggle_diffusion(choice): if choice == "MultiBand_Diffusion": return [gr.update(visible=True)] * 2 else: return [gr.update(visible=False)] * 2 def ui_full(launch_kwargs): with gr.Blocks() as interface: with gr.Row(): with gr.Column(): with gr.Row(): text = gr.Text(label="Input Text", interactive=True) # with gr.Column(): # radio = gr.Radio(["file", "mic"], value="file", # label="Condition on a melody (optional) File or Mic") # melody = gr.Audio(source="upload", type="numpy", label="File", # interactive=True, elem_id="melody-input") with gr.Row(): submit = gr.Button("Submit") # Adapted from https://github.com/rkfg/audiocraft/blob/long/app.py, MIT license. _ = gr.Button("Interrupt").click(fn=interrupt, queue=False) with gr.Row(): model = gr.Radio(["facebook/musicgen-melody", "facebook/musicgen-medium", "facebook/musicgen-small", "facebook/musicgen-large"], label="Model", value="facebook/musicgen-melody", interactive=True) # with gr.Row(): # decoder = gr.Radio(["Default", "MultiBand_Diffusion"], # label="Decoder", value="Default", interactive=True) # decoder = "Default" with gr.Row(): duration = gr.Slider(minimum=1, maximum=120, value=10, label="Duration", 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") audio_output = gr.Audio(label="Generated Music (wav)", type='filepath') # diffusion_output = gr.Video(label="MultiBand Diffusion Decoder") # audio_diffusion = gr.Audio(label="MultiBand Diffusion Decoder (wav)", type='filepath') melody = gr.Audio(source= None, type="numpy", label="File", interactive=False, visible= False, elem_id="melody-input") decoder = gr.Radio(["Default", "MultiBand_Diffusion"], label="Decoder", value="Default", interactive=True, visible= False) # duration = gr.Slider(minimum=1, maximum=120, value=10, label="Duration", interactive=True, visible= False) topk = gr.Number(label="Top-k", value=250, interactive=True, visible= False) topp = gr.Number(label="Top-p", value=0, interactive=True, visible= False) temperature = gr.Number(label="Temperature", value=1.0, interactive=True, visible= False) cfg_coef = gr.Number(label="Classifier Free Guidance", value=3.0, interactive=True, visible= False) diffusion_output = gr.Video(label="MultiBand Diffusion Decoder" , visible=False) audio_diffusion = gr.Audio(label="MultiBand Diffusion Decoder (wav)", type='filepath', visible= False) print("melody", melody) print("decoder", decoder) print("topk", topk) print("topp", topp) print("cfg_coef", cfg_coef) print("diffusion_output" , diffusion_output) print("audio_diffusion" , audio_diffusion) submit.click(toggle_diffusion, decoder, [diffusion_output, audio_diffusion], queue=False, show_progress=False).then(predict_full, inputs=[model, decoder, text, melody, duration, topk, topp, temperature, cfg_coef], outputs=[output, audio_output, diffusion_output, audio_diffusion]) # radio.change(toggle_audio_src, radio, [melody], queue=False, show_progress=False) # gr.Examples( # fn=predict_full, # examples=[ # [ # "An 80s driving pop song with heavy drums and synth pads in the background", # "./assets/bach.mp3", # "facebook/musicgen-melody", # "Default" # ], # [ # "A cheerful country song with acoustic guitars", # "./assets/bolero_ravel.mp3", # "facebook/musicgen-melody", # "Default" # ], # [ # "90s rock song with electric guitar and heavy drums", # None, # "facebook/musicgen-medium", # "Default" # ], # [ # "a light and cheerly EDM track, with syncopated drums, aery pads, and strong emotions", # "./assets/bach.mp3", # "facebook/musicgen-melody", # "Default" # ], # [ # "lofi slow bpm electro chill with organic samples", # None, # "facebook/musicgen-medium", # "Default" # ], # [ # "Punk rock with loud drum and power guitar", # None, # "facebook/musicgen-medium", # "MultiBand_Diffusion" # ], # ], # inputs=[text, melody, model, decoder], # outputs=[output] # ) gr.Markdown( """ """ ) interface.queue().launch(**launch_kwargs) def ui_batched(launch_kwargs): with gr.Blocks() as demo: gr.Markdown( """ This project generate Music from prompt. """ ) with gr.Row(): with gr.Column(): with gr.Row(): text = gr.Text(label="Describe your music", lines=2, interactive=True) with gr.Column(): radio = gr.Radio(["file", "mic"], value="file", label="Condition on a melody (optional) File or Mic") melody = gr.Audio(source="upload", type="numpy", label="File", interactive=True, elem_id="melody-input") with gr.Row(): submit = gr.Button("Generate") with gr.Column(): output = gr.Video(label="Generated Music") audio_output = gr.Audio(label="Generated Music (wav)", type='filepath') submit.click(predict_batched, inputs=[text, melody], outputs=[output, audio_output], batch=True, max_batch_size=MAX_BATCH_SIZE) radio.change(toggle_audio_src, radio, [melody], queue=False, show_progress=False) gr.Examples( fn=predict_batched, # examples=[ # [ # "An 80s driving pop song with heavy drums and synth pads in the background", # "./assets/bach.mp3", # ], # [ # "A cheerful country song with acoustic guitars", # "./assets/bolero_ravel.mp3", # ], # [ # "90s rock song with electric guitar and heavy drums", # None, # ], # [ # "a light and cheerly EDM track, with syncopated drums, aery pads, and strong emotions bpm: 130", # "./assets/bach.mp3", # ], # [ # "lofi slow bpm electro chill with organic samples", # None, # ], # ], examples=[ ], inputs=[text, melody], outputs=[output] ) 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: global USE_DIFFUSION USE_DIFFUSION = False ui_batched(launch_kwargs) else: ui_full(launch_kwargs)