Spaces:
Runtime error
Runtime error
File size: 11,065 Bytes
2561128 e66133f d533c9c c17b696 e66133f d533c9c 2561128 c17b696 2561128 c17b696 e66133f c78ba1a e66133f c17b696 c78ba1a c17b696 e66133f c17b696 c78ba1a c17b696 903650a c17b696 96e542f e66133f c17b696 e66133f 96e542f e66133f c17b696 e66133f 2561128 e66133f c17b696 e66133f c17b696 2561128 e66133f 072978d ea68dfd e66133f 072978d e66133f ea68dfd e66133f 2561128 072978d 2561128 21c77d0 2561128 072978d ea68dfd 2561128 c78ba1a ea68dfd e66133f 2561128 e66133f ea68dfd e66133f 2561128 ea68dfd 2561128 21c77d0 e66133f ea68dfd 2561128 08c6681 ea68dfd 2561128 21c77d0 ea68dfd 2561128 08c6681 e66133f 2561128 ea68dfd 2561128 072978d ea68dfd 2561128 ea68dfd 2561128 ea68dfd 2561128 21c77d0 2561128 21c77d0 e66133f 2561128 21c77d0 2561128 21c77d0 d533c9c 2561128 b7f49a5 2561128 b7f49a5 2561128 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 |
from typing import Iterable, Tuple, Union, List
import torch
import numpy as np
from PIL import Image
from tqdm.auto import tqdm
from librosa.beat import beat_track
from diffusers import (DiffusionPipeline, DDPMPipeline, UNet2DConditionModel,
DDIMScheduler, DDPMScheduler, AutoencoderKL)
from .mel import Mel
VERSION = "1.2.0"
class AudioDiffusion:
def __init__(self,
model_id: str = "teticio/audio-diffusion-256",
resolution: int = 256,
sample_rate: int = 22050,
n_fft: int = 2048,
hop_length: int = 512,
top_db: int = 80,
cuda: bool = torch.cuda.is_available(),
progress_bar: Iterable = tqdm):
"""Class for generating audio using Denoising Diffusion Probabilistic Models.
Args:
model_id (String): name of model (local directory or Hugging Face Hub)
resolution (int): size of square mel spectrogram in pixels
sample_rate (int): sample rate of audio
n_fft (int): number of Fast Fourier Transforms
hop_length (int): hop length (a higher number is recommended for lower than 256 y_res)
top_db (int): loudest in decibels
cuda (bool): use CUDA?
progress_bar (iterable): iterable callback for progress updates or None
"""
self.mel = Mel(x_res=resolution,
y_res=resolution,
sample_rate=sample_rate,
n_fft=n_fft,
hop_length=hop_length,
top_db=top_db)
self.model_id = model_id
pipeline = {
'LatentAudioDiffusionPipeline': LatentAudioDiffusionPipeline,
'AudioDiffusionPipeline': AudioDiffusionPipeline
}.get(
DiffusionPipeline.get_config_dict(self.model_id)['_class_name'],
AudioDiffusionPipeline)
self.pipe = pipeline.from_pretrained(self.model_id)
if cuda:
self.pipe.to("cuda")
self.progress_bar = progress_bar or (lambda _: _)
def generate_spectrogram_and_audio(
self,
steps: int = None,
generator: torch.Generator = None
) -> Tuple[Image.Image, Tuple[int, np.ndarray]]:
"""Generate random mel spectrogram and convert to audio.
Args:
steps (int): number of de-noising steps to perform (defaults to num_train_timesteps)
generator (torch.Generator): random number generator or None
Returns:
PIL Image: mel spectrogram
(float, np.ndarray): sample rate and raw audio
"""
images, (sample_rate, audios) = self.pipe(mel=self.mel,
batch_size=1,
steps=steps,
generator=generator)
return images[0], (sample_rate, audios[0])
@torch.no_grad()
def generate_spectrogram_and_audio_from_audio(
self,
audio_file: str = None,
raw_audio: np.ndarray = None,
slice: int = 0,
start_step: int = 0,
steps: int = None,
generator: torch.Generator = None,
mask_start_secs: float = 0,
mask_end_secs: float = 0
) -> Tuple[Image.Image, Tuple[int, np.ndarray]]:
"""Generate random mel spectrogram from audio input and convert to audio.
Args:
audio_file (str): must be a file on disk due to Librosa limitation or
raw_audio (np.ndarray): audio as numpy array
slice (int): slice number of audio to convert
start_step (int): step to start from
steps (int): number of de-noising steps to perform (defaults to num_train_timesteps)
generator (torch.Generator): random number generator or None
mask_start_secs (float): number of seconds of audio to mask (not generate) at start
mask_end_secs (float): number of seconds of audio to mask (not generate) at end
Returns:
PIL Image: mel spectrogram
(float, np.ndarray): sample rate and raw audio
"""
images, (sample_rate,
audios) = self.pipe(mel=self.mel,
batch_size=1,
audio_file=audio_file,
raw_audio=raw_audio,
slice=slice,
start_step=start_step,
steps=steps,
generator=generator,
mask_start_secs=mask_start_secs,
mask_end_secs=mask_end_secs)
return images[0], (sample_rate, audios[0])
@staticmethod
def loop_it(audio: np.ndarray,
sample_rate: int,
loops: int = 12) -> np.ndarray:
"""Loop audio
Args:
audio (np.ndarray): audio as numpy array
sample_rate (int): sample rate of audio
loops (int): number of times to loop
Returns:
(float, np.ndarray): sample rate and raw audio or None
"""
_, beats = beat_track(y=audio, sr=sample_rate, units='samples')
for beats_in_bar in [16, 12, 8, 4]:
if len(beats) > beats_in_bar:
return np.tile(audio[beats[0]:beats[beats_in_bar]], loops)
return None
class AudioDiffusionPipeline(DiffusionPipeline):
def __init__(self, unet: UNet2DConditionModel,
scheduler: Union[DDIMScheduler, DDPMScheduler]):
super().__init__()
self.register_modules(unet=unet, scheduler=scheduler)
@torch.no_grad()
def __call__(
self,
mel: Mel,
batch_size: int = 1,
audio_file: str = None,
raw_audio: np.ndarray = None,
slice: int = 0,
start_step: int = 0,
steps: int = None,
generator: torch.Generator = None,
mask_start_secs: float = 0,
mask_end_secs: float = 0
) -> Tuple[List[Image.Image], Tuple[int, List[np.ndarray]]]:
"""Generate random mel spectrogram from audio input and convert to audio.
Args:
mel (Mel): instance of Mel class to perform image <-> audio
batch_size (int): number of samples to generate
audio_file (str): must be a file on disk due to Librosa limitation or
raw_audio (np.ndarray): audio as numpy array
slice (int): slice number of audio to convert
start_step (int): step to start from
steps (int): number of de-noising steps to perform (defaults to num_train_timesteps)
generator (torch.Generator): random number generator or None
mask_start_secs (float): number of seconds of audio to mask (not generate) at start
mask_end_secs (float): number of seconds of audio to mask (not generate) at end
Returns:
List[PIL Image]: mel spectrograms
(float, List[np.ndarray]): sample rate and raw audios
"""
if steps is None:
steps = self.scheduler.num_train_timesteps
# Unfortunately, the schedule is set up in the constructor
scheduler = self.scheduler.__class__(num_train_timesteps=steps)
scheduler.set_timesteps(steps)
mask = None
images = noise = torch.randn(
(batch_size, self.unet.in_channels, self.unet.sample_size,
self.unet.sample_size),
generator=generator)
if audio_file is not None or raw_audio is not None:
mel.load_audio(audio_file, raw_audio)
input_image = mel.audio_slice_to_image(slice)
input_image = np.frombuffer(input_image.tobytes(),
dtype="uint8").reshape(
(input_image.height,
input_image.width))
input_image = ((input_image / 255) * 2 - 1)
input_images = np.tile(input_image, (batch_size, 1, 1, 1))
if hasattr(self, 'vqvae'):
input_images = self.vqvae.encode(
input_images).latent_dist.sample(generator=generator)
input_images = 0.18215 * input_images
if start_step > 0:
images[0, 0] = scheduler.add_noise(
torch.tensor(input_images[:, np.newaxis, np.newaxis, :]),
noise, torch.tensor(steps - start_step))
pixels_per_second = (mel.get_sample_rate() *
self.unet.sample_size / mel.hop_length /
mel.x_res)
mask_start = int(mask_start_secs * pixels_per_second)
mask_end = int(mask_end_secs * pixels_per_second)
mask = scheduler.add_noise(
torch.tensor(input_images[:, np.newaxis, :]), noise,
torch.tensor(scheduler.timesteps[start_step:]))
images = images.to(self.device)
for step, t in enumerate(
self.progress_bar(scheduler.timesteps[start_step:])):
model_output = self.unet(images, t)['sample']
images = scheduler.step(model_output,
t,
images,
generator=generator)['prev_sample']
if mask is not None:
if mask_start > 0:
images[:, :, :, :mask_start] = mask[
step, :, :, :, :mask_start]
if mask_end > 0:
images[:, :, :, -mask_end:] = mask[step, :, :, :,
-mask_end:]
if hasattr(self, 'vqvae'):
# 0.18215 was scaling factor used in training to ensure unit variance
images = 1 / 0.18215 * images
images = self.vqvae.decode(images)['sample']
images = (images / 2 + 0.5).clamp(0, 1)
images = images.cpu().permute(0, 2, 3, 1).numpy()
images = (images * 255).round().astype("uint8")
images = list(
map(lambda _: Image.fromarray(_[:, :, 0]), images) if images.
shape[3] == 1 else map(
lambda _: Image.fromarray(_, mode='RGB').convert('L'), images))
audios = list(map(lambda _: mel.image_to_audio(_), images))
return images, (mel.get_sample_rate(), audios)
class LatentAudioDiffusionPipeline(AudioDiffusionPipeline):
def __init__(self, unet: UNet2DConditionModel,
scheduler: Union[DDIMScheduler,
DDPMScheduler], vqvae: AutoencoderKL):
super().__init__(unet=unet, scheduler=scheduler)
self.register_modules(vqvae=vqvae)
def __call__(self, *args, **kwargs):
return super().__call__(*args, **kwargs)
|