Create musicgen_colab.py
Browse files- demos/musicgen_colab.py +494 -0
demos/musicgen_colab.py
ADDED
@@ -0,0 +1,494 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import spaces # <--- IMPORTANT: Add this import
|
2 |
+
import argparse
|
3 |
+
import logging
|
4 |
+
import os
|
5 |
+
from pathlib import Path
|
6 |
+
import subprocess as sp
|
7 |
+
import sys
|
8 |
+
import time
|
9 |
+
import typing as tp
|
10 |
+
from tempfile import NamedTemporaryFile, gettempdir
|
11 |
+
from einops import rearrange
|
12 |
+
import torch
|
13 |
+
import gradio as gr
|
14 |
+
from audiocraft.data.audio_utils import convert_audio
|
15 |
+
from audiocraft.data.audio import audio_write
|
16 |
+
from audiocraft.models.encodec import InterleaveStereoCompressionModel
|
17 |
+
from audiocraft.models import MusicGen, MultiBandDiffusion
|
18 |
+
import multiprocessing as mp
|
19 |
+
import warnings
|
20 |
+
|
21 |
+
os.putenv("HF_HUB_ENABLE_HF_TRANSFER","1")
|
22 |
+
os.environ["SAFETENSORS_FAST_GPU"] = "1"
|
23 |
+
|
24 |
+
torch.backends.cuda.matmul.allow_tf32 = False
|
25 |
+
torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = False
|
26 |
+
torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = False
|
27 |
+
torch.backends.cudnn.allow_tf32 = False
|
28 |
+
torch.backends.cudnn.deterministic = False
|
29 |
+
torch.backends.cudnn.benchmark = False
|
30 |
+
# torch.backends.cuda.preferred_blas_library="cublas"
|
31 |
+
# torch.backends.cuda.preferred_linalg_library="cusolver"
|
32 |
+
torch.set_float32_matmul_precision("highest")
|
33 |
+
|
34 |
+
class FileCleaner:
|
35 |
+
def __init__(self, file_lifetime: float = 3600):
|
36 |
+
self.file_lifetime = file_lifetime
|
37 |
+
self.files = []
|
38 |
+
def add(self, path: tp.Union[str, Path]):
|
39 |
+
self._cleanup()
|
40 |
+
self.files.append((time.time(), Path(path)))
|
41 |
+
def _cleanup(self):
|
42 |
+
now = time.time()
|
43 |
+
for time_added, path in list(self.files):
|
44 |
+
if now - time_added > self.file_lifetime:
|
45 |
+
if path.exists():
|
46 |
+
path.unlink()
|
47 |
+
self.files.pop(0)
|
48 |
+
else:
|
49 |
+
break
|
50 |
+
|
51 |
+
file_cleaner = FileCleaner()
|
52 |
+
|
53 |
+
def convert_wav_to_mp4(wav_path, output_path=None):
|
54 |
+
"""Converts a WAV file to a waveform MP4 video using ffmpeg."""
|
55 |
+
if output_path is None:
|
56 |
+
# Create output path in the same directory as the input
|
57 |
+
output_path = Path(wav_path).with_suffix(".mp4")
|
58 |
+
try:
|
59 |
+
command = [
|
60 |
+
"ffmpeg",
|
61 |
+
"-y", # Overwrite output file if it exists
|
62 |
+
"-i", str(wav_path),
|
63 |
+
"-filter_complex",
|
64 |
+
"[0:a]showwaves=s=1280x202:mode=line,format=yuv420p[v]", # Waveform filter
|
65 |
+
"-map", "[v]",
|
66 |
+
"-map", "0:a",
|
67 |
+
"-c:v", "libx264", # Video codec
|
68 |
+
"-c:a", "aac", # Audio codec
|
69 |
+
"-preset", "fast", # Important, don't do veryslow.
|
70 |
+
str(output_path),
|
71 |
+
]
|
72 |
+
process = sp.run(command, capture_output=True, text=True, check=True)
|
73 |
+
return str(output_path)
|
74 |
+
except sp.CalledProcessError as e:
|
75 |
+
print(f"Error in ffmpeg conversion: {e}")
|
76 |
+
print(f"ffmpeg stdout: {e.stdout}")
|
77 |
+
print(f"ffmpeg stderr: {e.stderr}")
|
78 |
+
raise # Re-raise the exception to be caught by Gradio
|
79 |
+
|
80 |
+
def model_worker(model_name: str, task_queue: mp.Queue, result_queue: mp.Queue):
|
81 |
+
"""
|
82 |
+
Persistent worker process (used when NOT running as a daemon).
|
83 |
+
"""
|
84 |
+
try:
|
85 |
+
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
86 |
+
model = MusicGen.get_pretrained(model_name, device=device)
|
87 |
+
mbd = MultiBandDiffusion.get_mbd_musicgen(device=device)
|
88 |
+
while True:
|
89 |
+
task = task_queue.get()
|
90 |
+
if task is None:
|
91 |
+
break
|
92 |
+
task_id, text, melody, duration, use_diffusion, gen_params = task
|
93 |
+
try:
|
94 |
+
model.set_generation_params(duration=duration, **gen_params)
|
95 |
+
target_sr = model.sample_rate
|
96 |
+
target_ac = 1
|
97 |
+
processed_melody = None
|
98 |
+
if melody:
|
99 |
+
sr, melody_data = melody
|
100 |
+
melody_tensor = torch.from_numpy(melody_data).to(device).float().t()
|
101 |
+
if melody_tensor.ndim == 1:
|
102 |
+
melody_tensor = melody_tensor.unsqueeze(0)
|
103 |
+
melody_tensor = melody_tensor[..., :int(sr * duration)]
|
104 |
+
processed_melody = convert_audio(melody_tensor, sr, target_sr, target_ac)
|
105 |
+
if processed_melody is not None:
|
106 |
+
output, tokens = model.generate_with_chroma(
|
107 |
+
descriptions=[text],
|
108 |
+
melody_wavs=[processed_melody],
|
109 |
+
melody_sample_rate=target_sr,
|
110 |
+
progress=True,
|
111 |
+
return_tokens=True
|
112 |
+
)
|
113 |
+
else:
|
114 |
+
output, tokens = model.generate([text], progress=True, return_tokens=True)
|
115 |
+
output = output.detach().cpu()
|
116 |
+
if use_diffusion:
|
117 |
+
if isinstance(model.compression_model, InterleaveStereoCompressionModel):
|
118 |
+
left, right = model.compression_model.get_left_right_codes(tokens)
|
119 |
+
tokens = torch.cat([left, right])
|
120 |
+
outputs_diffusion = mbd.tokens_to_wav(tokens)
|
121 |
+
if isinstance(model.compression_model, InterleaveStereoCompressionModel):
|
122 |
+
assert outputs_diffusion.shape[1] == 1 # output is mono
|
123 |
+
outputs_diffusion = rearrange(outputs_diffusion, '(s b) c t -> b (s c) t', s=2)
|
124 |
+
outputs_diffusion = outputs_diffusion.detach().cpu()
|
125 |
+
result_queue.put((task_id, (output, outputs_diffusion)))
|
126 |
+
else:
|
127 |
+
result_queue.put((task_id, (output, None)))
|
128 |
+
except Exception as e:
|
129 |
+
result_queue.put((task_id, e))
|
130 |
+
except Exception as e:
|
131 |
+
result_queue.put((-1, e))
|
132 |
+
|
133 |
+
class Predictor:
|
134 |
+
def __init__(self, model_name: str, depth: str):
|
135 |
+
self.model_name = model_name
|
136 |
+
self.is_daemon = mp.current_process().daemon
|
137 |
+
if self.is_daemon:
|
138 |
+
# Running in a daemonic process (e.g., on Spaces)
|
139 |
+
self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
140 |
+
self.model = MusicGen.get_pretrained(self.model_name, device=self.device, depth=depth)
|
141 |
+
self.mbd = MultiBandDiffusion.get_mbd_musicgen(device=self.device) # Load MBD here too
|
142 |
+
self.current_task_id = 0 # Initialize task ID
|
143 |
+
else:
|
144 |
+
# Running in a non-daemonic process (e.g., locally)
|
145 |
+
self.task_queue = mp.Queue()
|
146 |
+
self.result_queue = mp.Queue()
|
147 |
+
self.process = mp.Process(
|
148 |
+
target=model_worker, args=(self.model_name, self.task_queue, self.result_queue)
|
149 |
+
)
|
150 |
+
self.process.start()
|
151 |
+
self.current_task_id = 0
|
152 |
+
self._check_initialization()
|
153 |
+
|
154 |
+
def _check_initialization(self):
|
155 |
+
"""Check if the worker process initialized successfully (only in non-daemon mode)."""
|
156 |
+
if not self.is_daemon:
|
157 |
+
time.sleep(2)
|
158 |
+
try:
|
159 |
+
task_id, result = self.result_queue.get(timeout=3)
|
160 |
+
if isinstance(result, Exception):
|
161 |
+
if task_id == -1:
|
162 |
+
raise RuntimeError("Model loading failed in worker process.") from result
|
163 |
+
except:
|
164 |
+
pass
|
165 |
+
|
166 |
+
def predict(self, text, melody, duration, use_diffusion, **gen_params):
|
167 |
+
"""Submits a prediction task."""
|
168 |
+
if self.is_daemon:
|
169 |
+
# Directly perform the prediction (single-process mode)
|
170 |
+
self.current_task_id +=1
|
171 |
+
task_id = self.current_task_id
|
172 |
+
try:
|
173 |
+
self.model.set_generation_params(duration=duration, **gen_params)
|
174 |
+
target_sr = self.model.sample_rate
|
175 |
+
target_ac = 1
|
176 |
+
processed_melody = None
|
177 |
+
if melody:
|
178 |
+
sr, melody_data = melody
|
179 |
+
melody_tensor = torch.from_numpy(melody_data).to(self.device).float().t()
|
180 |
+
if melody_tensor.ndim == 1:
|
181 |
+
melody_tensor = melody_tensor.unsqueeze(0)
|
182 |
+
melody_tensor = melody_tensor[..., :int(sr * duration)]
|
183 |
+
processed_melody = convert_audio(melody_tensor, sr, target_sr, target_ac)
|
184 |
+
if processed_melody is not None:
|
185 |
+
output, tokens = self.model.generate_with_chroma(
|
186 |
+
descriptions=[text],
|
187 |
+
melody_wavs=[processed_melody],
|
188 |
+
melody_sample_rate=target_sr,
|
189 |
+
progress=True,
|
190 |
+
return_tokens=True
|
191 |
+
)
|
192 |
+
else:
|
193 |
+
output, tokens = self.model.generate([text], progress=True, return_tokens=True)
|
194 |
+
output = output.detach().cpu()
|
195 |
+
if use_diffusion:
|
196 |
+
if isinstance(self.model.compression_model, InterleaveStereoCompressionModel):
|
197 |
+
left, right = self.model.compression_model.get_left_right_codes(tokens)
|
198 |
+
tokens = torch.cat([left, right])
|
199 |
+
outputs_diffusion = self.mbd.tokens_to_wav(tokens)
|
200 |
+
if isinstance(self.model.compression_model, InterleaveStereoCompressionModel):
|
201 |
+
assert outputs_diffusion.shape[1] == 1 # output is mono
|
202 |
+
outputs_diffusion = rearrange(outputs_diffusion, '(s b) c t -> b (s c) t', s=2)
|
203 |
+
outputs_diffusion = outputs_diffusion.detach().cpu()
|
204 |
+
return task_id, (output, outputs_diffusion) #Return the task id.
|
205 |
+
else:
|
206 |
+
return task_id, (output, None)
|
207 |
+
except Exception as e:
|
208 |
+
return task_id, e
|
209 |
+
else:
|
210 |
+
# Use the multiprocessing queue (multi-process mode)
|
211 |
+
self.current_task_id += 1
|
212 |
+
task = (self.current_task_id, text, melody, duration, use_diffusion, gen_params)
|
213 |
+
self.task_queue.put(task)
|
214 |
+
return self.current_task_id
|
215 |
+
|
216 |
+
def get_result(self, task_id):
|
217 |
+
"""Retrieves the result of a prediction task."""
|
218 |
+
if self.is_daemon:
|
219 |
+
# Results are returned directly by 'predict' in daemon mode
|
220 |
+
result_id, result = task_id, task_id #predictor return (task_id, results)
|
221 |
+
else:
|
222 |
+
# Get result from the queue (multi-process mode)
|
223 |
+
while True:
|
224 |
+
result_task_id, result = self.result_queue.get()
|
225 |
+
if result_task_id == task_id:
|
226 |
+
break # Found the correct result
|
227 |
+
if isinstance(result, Exception):
|
228 |
+
raise result
|
229 |
+
return result
|
230 |
+
|
231 |
+
def shutdown(self):
|
232 |
+
"""Shuts down the worker process (if running)."""
|
233 |
+
if not self.is_daemon and self.process.is_alive():
|
234 |
+
self.task_queue.put(None)
|
235 |
+
self.process.join()
|
236 |
+
|
237 |
+
_default_model_name = "facebook/musicgen-melody"
|
238 |
+
|
239 |
+
@spaces.GPU(duration=90) # Use the decorator for Spaces
|
240 |
+
def predict_full(model, model_path, depth, use_mbd, text, melody, duration, topk, topp, temperature, cfg_coef):
|
241 |
+
# Initialize Predictor *INSIDE* the function
|
242 |
+
predictor = Predictor(model, depth)
|
243 |
+
task_id, (wav, diffusion_wav) = predictor.predict( # Unpack directly!
|
244 |
+
text=text,
|
245 |
+
melody=melody,
|
246 |
+
duration=duration,
|
247 |
+
use_diffusion=use_mbd,
|
248 |
+
top_k=topk,
|
249 |
+
top_p=topp,
|
250 |
+
temperature=temperature,
|
251 |
+
cfg_coef=cfg_coef,
|
252 |
+
)
|
253 |
+
# Save and return audio files
|
254 |
+
wav_paths = []
|
255 |
+
video_paths = []
|
256 |
+
# Save standard output
|
257 |
+
with NamedTemporaryFile("wb", suffix=".wav", delete=False) as file:
|
258 |
+
audio_write(
|
259 |
+
file.name, wav[0], 44100, strategy="loudness", #hardcoded sample rate
|
260 |
+
loudness_headroom_db=16, loudness_compressor=True, add_suffix=False
|
261 |
+
)
|
262 |
+
wav_paths.append(file.name)
|
263 |
+
# Make and clean up video:
|
264 |
+
video_path = convert_wav_to_mp4(file.name)
|
265 |
+
video_paths.append(video_path)
|
266 |
+
file_cleaner.add(file.name)
|
267 |
+
file_cleaner.add(video_path)
|
268 |
+
# Save MBD output if used
|
269 |
+
if diffusion_wav is not None:
|
270 |
+
with NamedTemporaryFile("wb", suffix=".wav", delete=False) as file:
|
271 |
+
audio_write(
|
272 |
+
file.name, diffusion_wav[0], 44100, strategy="loudness", #hardcoded sample rate
|
273 |
+
loudness_headroom_db=16, loudness_compressor=True, add_suffix=False
|
274 |
+
)
|
275 |
+
wav_paths.append(file.name)
|
276 |
+
# Make and clean up video:
|
277 |
+
video_path = convert_wav_to_mp4(file.name)
|
278 |
+
video_paths.append(video_path)
|
279 |
+
file_cleaner.add(file.name)
|
280 |
+
file_cleaner.add(video_path)
|
281 |
+
# Shutdown predictor to prevent hanging processes!
|
282 |
+
if not predictor.is_daemon: # Important!
|
283 |
+
predictor.shutdown()
|
284 |
+
if use_mbd:
|
285 |
+
return video_paths[0], wav_paths[0], video_paths[1], wav_paths[1]
|
286 |
+
return video_paths[0], wav_paths[0], None, None
|
287 |
+
|
288 |
+
def toggle_audio_src(choice):
|
289 |
+
if choice == "mic":
|
290 |
+
return gr.update(sources="microphone", value=None, label="Microphone")
|
291 |
+
else:
|
292 |
+
return gr.update(sources="upload", value=None, label="File")
|
293 |
+
|
294 |
+
def toggle_diffusion(choice):
|
295 |
+
if choice == "MultiBand_Diffusion":
|
296 |
+
return [gr.update(visible=True)] * 2
|
297 |
+
else:
|
298 |
+
return [gr.update(visible=False)] * 2
|
299 |
+
|
300 |
+
def ui_full(launch_kwargs):
|
301 |
+
with gr.Blocks() as interface:
|
302 |
+
gr.Markdown(
|
303 |
+
"""
|
304 |
+
# MusicGen
|
305 |
+
This is your private demo for [MusicGen](https://github.com/facebookresearch/audiocraft),
|
306 |
+
a simple and controllable model for music generation
|
307 |
+
presented at: ["Simple and Controllable Music Generation"](https://huggingface.co/papers/2306.05284)
|
308 |
+
"""
|
309 |
+
)
|
310 |
+
with gr.Row():
|
311 |
+
with gr.Column():
|
312 |
+
with gr.Row():
|
313 |
+
text = gr.Text(label="Input Text", interactive=True)
|
314 |
+
with gr.Column():
|
315 |
+
radio = gr.Radio(["file", "mic"], value="file",
|
316 |
+
label="Condition on a melody (optional) File or Mic")
|
317 |
+
melody = gr.Audio(sources="upload", type="numpy", label="File",
|
318 |
+
interactive=True, elem_id="melody-input")
|
319 |
+
with gr.Row():
|
320 |
+
submit = gr.Button("Submit")
|
321 |
+
# _ = gr.Button("Interrupt").click(fn=interrupt, queue=False) # Interrupt is now handled implicitly
|
322 |
+
with gr.Row():
|
323 |
+
model = gr.Radio(["facebook/musicgen-melody", "facebook/musicgen-medium", "facebook/musicgen-small",
|
324 |
+
"facebook/musicgen-large", "facebook/musicgen-melody-large",
|
325 |
+
"facebook/musicgen-stereo-small", "facebook/musicgen-stereo-medium",
|
326 |
+
"facebook/musicgen-stereo-melody", "facebook/musicgen-stereo-large",
|
327 |
+
"facebook/musicgen-stereo-melody-large"],
|
328 |
+
label="Model", value="facebook/musicgen-melody", interactive=True)
|
329 |
+
model_path = gr.Text(label="Model Path (custom models)", interactive=False, visible=False) # Keep, but hide
|
330 |
+
depth = gr.Radio(["float32", "bfloat16", "float16"],
|
331 |
+
label="Model Precision", value="float32", interactive=True)
|
332 |
+
with gr.Row():
|
333 |
+
decoder = gr.Radio(["Default", "MultiBand_Diffusion"],
|
334 |
+
label="Decoder", value="Default", interactive=True)
|
335 |
+
with gr.Row():
|
336 |
+
duration = gr.Slider(minimum=1, maximum=120, value=10, label="Duration", interactive=True)
|
337 |
+
with gr.Row():
|
338 |
+
topk = gr.Number(label="Top-k", value=250, interactive=True)
|
339 |
+
topp = gr.Number(label="Top-p", value=0, interactive=True)
|
340 |
+
temperature = gr.Number(label="Temperature", value=1.0, interactive=True)
|
341 |
+
cfg_coef = gr.Number(label="Classifier Free Guidance", value=3.0, interactive=True)
|
342 |
+
with gr.Column():
|
343 |
+
output = gr.Video(label="Generated Music")
|
344 |
+
audio_output = gr.Audio(label="Generated Music (wav)", type='filepath')
|
345 |
+
diffusion_output = gr.Video(label="MultiBand Diffusion Decoder", visible=False)
|
346 |
+
audio_diffusion = gr.Audio(label="MultiBand Diffusion Decoder (wav)", type='filepath', visible=False)
|
347 |
+
|
348 |
+
submit.click(
|
349 |
+
toggle_diffusion, decoder, [diffusion_output, audio_diffusion], queue=False
|
350 |
+
).then(
|
351 |
+
predict_full,
|
352 |
+
inputs=[model, model_path, depth, decoder, text, melody, duration, topk, topp, temperature, cfg_coef],
|
353 |
+
outputs=[output, audio_output, diffusion_output, audio_diffusion]
|
354 |
+
)
|
355 |
+
radio.change(toggle_audio_src, radio, [melody], queue=False, show_progress=False)
|
356 |
+
|
357 |
+
gr.Examples(
|
358 |
+
fn=predict_full,
|
359 |
+
examples=[
|
360 |
+
[
|
361 |
+
"An 80s driving pop song with heavy drums and synth pads in the background",
|
362 |
+
"./assets/bach.mp3",
|
363 |
+
"facebook/musicgen-melody",
|
364 |
+
"Default"
|
365 |
+
],
|
366 |
+
[
|
367 |
+
"A cheerful country song with acoustic guitars",
|
368 |
+
"./assets/bolero_ravel.mp3",
|
369 |
+
"facebook/musicgen-melody",
|
370 |
+
"Default"
|
371 |
+
],
|
372 |
+
[
|
373 |
+
"90s rock song with electric guitar and heavy drums",
|
374 |
+
None,
|
375 |
+
"facebook/musicgen-medium",
|
376 |
+
"Default"
|
377 |
+
],
|
378 |
+
[
|
379 |
+
"a light and cheerly EDM track, with syncopated drums, aery pads, and strong emotions",
|
380 |
+
"./assets/bach.mp3",
|
381 |
+
"facebook/musicgen-melody",
|
382 |
+
"Default"
|
383 |
+
],
|
384 |
+
[
|
385 |
+
"lofi slow bpm electro chill with organic samples",
|
386 |
+
None,
|
387 |
+
"facebook/musicgen-medium",
|
388 |
+
"Default"
|
389 |
+
],
|
390 |
+
[
|
391 |
+
"Punk rock with loud drum and power guitar",
|
392 |
+
None,
|
393 |
+
"facebook/musicgen-medium",
|
394 |
+
"MultiBand_Diffusion"
|
395 |
+
],
|
396 |
+
],
|
397 |
+
inputs=[text, melody, model, decoder],
|
398 |
+
outputs=[output]
|
399 |
+
)
|
400 |
+
gr.Markdown(
|
401 |
+
"""
|
402 |
+
### More details
|
403 |
+
|
404 |
+
The model will generate a short music extract based on the description you provided.
|
405 |
+
The model can generate up to 30 seconds of audio in one pass.
|
406 |
+
|
407 |
+
The model was trained with description from a stock music catalog, descriptions that will work best
|
408 |
+
should include some level of details on the instruments present, along with some intended use case
|
409 |
+
(e.g. adding "perfect for a commercial" can somehow help).
|
410 |
+
|
411 |
+
Using one of the `melody` model (e.g. `musicgen-melody-*`), you can optionally provide a reference audio
|
412 |
+
from which a broad melody will be extracted.
|
413 |
+
The model will then try to follow both the description and melody provided.
|
414 |
+
For best results, the melody should be 30 seconds long (I know, the samples we provide are not...)
|
415 |
+
|
416 |
+
It is now possible to extend the generation by feeding back the end of the previous chunk of audio.
|
417 |
+
This can take a long time, and the model might lose consistency. The model might also
|
418 |
+
decide at arbitrary positions that the song ends.
|
419 |
+
|
420 |
+
**WARNING:** Choosing long durations will take a long time to generate (2min might take ~10min).
|
421 |
+
An overlap of 12 seconds is kept with the previously generated chunk, and 18 "new" seconds
|
422 |
+
are generated each time.
|
423 |
+
|
424 |
+
We present 10 model variations:
|
425 |
+
1. facebook/musicgen-melody -- a music generation model capable of generating music condition
|
426 |
+
on text and melody inputs. **Note**, you can also use text only.
|
427 |
+
2. facebook/musicgen-small -- a 300M transformer decoder conditioned on text only.
|
428 |
+
3. facebook/musicgen-medium -- a 1.5B transformer decoder conditioned on text only.
|
429 |
+
4. facebook/musicgen-large -- a 3.3B transformer decoder conditioned on text only.
|
430 |
+
5. facebook/musicgen-melody-large -- a 3.3B transformer decoder conditioned on text and melody.
|
431 |
+
6. facebook/musicgen-stereo-small -- a 300M transformer decoder conditioned on text only, fine tuned for stereo output.
|
432 |
+
7. facebook/musicgen-stereo-medium -- a 1.5B transformer decoder conditioned on text only, fine tuned for stereo output.
|
433 |
+
8. facebook/musicgen-stereo-melody -- a 1.5B transformer decoder conditioned on text and melody, fine tuned for stereo output.
|
434 |
+
9. facebook/musicgen-stereo-large -- a 3.3B transformer decoder conditioned on text only, fine tuned for stereo output.
|
435 |
+
10. facebook/musicgen-stereo-melody-large -- a 3.3B transformer decoder conditioned on text and melody, fine tuned for stereo output.
|
436 |
+
|
437 |
+
We also present two way of decoding the audio tokens:
|
438 |
+
1. Use the default GAN based compression model. It can suffer from artifacts especially
|
439 |
+
for crashes, snares etc.
|
440 |
+
2. Use [MultiBand Diffusion](https://arxiv.org/abs/2308.02560). Should improve the audio quality,
|
441 |
+
at an extra computational cost. When this is selected, we provide both the GAN based decoded
|
442 |
+
audio, and the one obtained with MBD.
|
443 |
+
|
444 |
+
See [github.com/facebookresearch/audiocraft](https://github.com/facebookresearch/audiocraft/blob/main/docs/MUSICGEN.md)
|
445 |
+
for more details.
|
446 |
+
"""
|
447 |
+
)
|
448 |
+
|
449 |
+
interface.queue().launch(**launch_kwargs)
|
450 |
+
|
451 |
+
if __name__ == '__main__':
|
452 |
+
parser = argparse.ArgumentParser()
|
453 |
+
parser.add_argument(
|
454 |
+
'--listen',
|
455 |
+
type=str,
|
456 |
+
default='0.0.0.0' if 'SPACE_ID' in os.environ else '127.0.0.1',
|
457 |
+
help='IP to listen on for connections to Gradio',
|
458 |
+
)
|
459 |
+
parser.add_argument(
|
460 |
+
'--username', type=str, default='', help='Username for authentication'
|
461 |
+
)
|
462 |
+
parser.add_argument(
|
463 |
+
'--password', type=str, default='', help='Password for authentication'
|
464 |
+
)
|
465 |
+
parser.add_argument(
|
466 |
+
'--server_port',
|
467 |
+
type=int,
|
468 |
+
default=0,
|
469 |
+
help='Port to run the server listener on',
|
470 |
+
)
|
471 |
+
parser.add_argument(
|
472 |
+
'--inbrowser', action='store_true', help='Open in browser'
|
473 |
+
)
|
474 |
+
parser.add_argument(
|
475 |
+
'--share', action='store_true', help='Share the gradio UI'
|
476 |
+
)
|
477 |
+
args = parser.parse_args()
|
478 |
+
launch_kwargs = {}
|
479 |
+
launch_kwargs['server_name'] = args.listen
|
480 |
+
if args.username and args.password:
|
481 |
+
launch_kwargs['auth'] = (args.username, args.password)
|
482 |
+
if args.server_port:
|
483 |
+
launch_kwargs['server_port'] = args.server_port
|
484 |
+
if args.inbrowser:
|
485 |
+
launch_kwargs['inbrowser'] = args.inbrowser
|
486 |
+
if args.share:
|
487 |
+
launch_kwargs['share'] = True
|
488 |
+
logging.basicConfig(level=logging.INFO, stream=sys.stderr)
|
489 |
+
# Added predictor shutdown
|
490 |
+
try:
|
491 |
+
ui_full(launch_kwargs)
|
492 |
+
finally:
|
493 |
+
if _predictor is not None:
|
494 |
+
_predictor.shutdown()
|