Spaces:
Sleeping
Sleeping
File size: 20,669 Bytes
41b9d24 05d43c6 41b9d24 05d43c6 41b9d24 08c78c6 41b9d24 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 |
import os
from pathlib import Path
import math
import logging
import torch
import numpy as np
from audiotools import AudioSignal
import tqdm
from .modules.transformer import VampNet
from .beats import WaveBeat
from .mask import *
# from dac.model.dac import DAC
from lac.model.lac import LAC as DAC
def signal_concat(
audio_signals: list,
):
audio_data = torch.cat([x.audio_data for x in audio_signals], dim=-1)
return AudioSignal(audio_data, sample_rate=audio_signals[0].sample_rate)
def _load_model(
ckpt: str,
lora_ckpt: str = None,
device: str = "cpu",
chunk_size_s: int = 10,
):
# we need to set strict to False if the model has lora weights to add later
model = VampNet.load(location=Path(ckpt), map_location="cpu", strict=False)
# load lora weights if needed
if lora_ckpt is not None:
if not Path(lora_ckpt).exists():
should_cont = input(
f"lora checkpoint {lora_ckpt} does not exist. continue? (y/n) "
)
if should_cont != "y":
raise Exception("aborting")
else:
model.load_state_dict(torch.load(lora_ckpt, map_location="cpu"), strict=False)
model.to(device)
model.eval()
model.chunk_size_s = chunk_size_s
return model
class Interface(torch.nn.Module):
def __init__(
self,
coarse_ckpt: str = None,
coarse_lora_ckpt: str = None,
coarse2fine_ckpt: str = None,
coarse2fine_lora_ckpt: str = None,
codec_ckpt: str = None,
wavebeat_ckpt: str = None,
device: str = "cpu",
coarse_chunk_size_s: int = 10,
coarse2fine_chunk_size_s: int = 3,
compile=True,
):
super().__init__()
assert codec_ckpt is not None, "must provide a codec checkpoint"
self.codec = DAC.load(Path(codec_ckpt))
self.codec.eval()
self.codec.to(device)
self.codec_path = Path(codec_ckpt)
assert coarse_ckpt is not None, "must provide a coarse checkpoint"
self.coarse = _load_model(
ckpt=coarse_ckpt,
lora_ckpt=coarse_lora_ckpt,
device=device,
chunk_size_s=coarse_chunk_size_s,
)
self.coarse_path = Path(coarse_ckpt)
# check if we have a coarse2fine ckpt
if coarse2fine_ckpt is not None:
self.c2f_path = Path(coarse2fine_ckpt)
self.c2f = _load_model(
ckpt=coarse2fine_ckpt,
lora_ckpt=coarse2fine_lora_ckpt,
device=device,
chunk_size_s=coarse2fine_chunk_size_s,
)
else:
self.c2f_path = None
self.c2f = None
if wavebeat_ckpt is not None:
logging.debug(f"loading wavebeat from {wavebeat_ckpt}")
self.beat_tracker = WaveBeat(wavebeat_ckpt)
self.beat_tracker.model.to(device)
else:
self.beat_tracker = None
self.device = device
self.loudness = -24.0
if compile:
logging.debug(f"compiling models")
self.coarse = torch.compile(self.coarse)
if self.c2f is not None:
self.c2f = torch.compile(self.c2f)
self.codec = torch.compile(self.codec)
@classmethod
def default(cls):
from . import download_codec, download_default
print(f"loading default vampnet")
codec_path = download_codec()
coarse_path, c2f_path = download_default()
return Interface(
coarse_ckpt=coarse_path,
coarse2fine_ckpt=c2f_path,
codec_ckpt=codec_path,
)
@classmethod
def available_models(cls):
from . import list_finetuned
return list_finetuned() + ["default"]
def load_finetuned(self, name: str):
assert name in self.available_models(), f"{name} is not a valid model name"
from . import download_finetuned, download_default
if name == "default":
coarse_path, c2f_path = download_default()
else:
coarse_path, c2f_path = download_finetuned(name)
self.reload(
coarse_ckpt=coarse_path,
c2f_ckpt=c2f_path,
)
def reload(
self,
coarse_ckpt: str = None,
c2f_ckpt: str = None,
):
if coarse_ckpt is not None:
# check if we already loaded, if so, don't reload
if self.coarse_path == Path(coarse_ckpt):
logging.debug(f"already loaded {coarse_ckpt}")
else:
self.coarse = _load_model(
ckpt=coarse_ckpt,
device=self.device,
chunk_size_s=self.coarse.chunk_size_s,
)
self.coarse_path = Path(coarse_ckpt)
logging.debug(f"loaded {coarse_ckpt}")
if c2f_ckpt is not None:
if self.c2f_path == Path(c2f_ckpt):
logging.debug(f"already loaded {c2f_ckpt}")
else:
self.c2f = _load_model(
ckpt=c2f_ckpt,
device=self.device,
chunk_size_s=self.c2f.chunk_size_s,
)
self.c2f_path = Path(c2f_ckpt)
logging.debug(f"loaded {c2f_ckpt}")
def s2t(self, seconds: float):
"""seconds to tokens"""
if isinstance(seconds, np.ndarray):
return np.ceil(seconds * self.codec.sample_rate / self.codec.hop_length)
else:
return math.ceil(seconds * self.codec.sample_rate / self.codec.hop_length)
def s2t2s(self, seconds: float):
"""seconds to tokens to seconds"""
return self.t2s(self.s2t(seconds))
def t2s(self, tokens: int):
"""tokens to seconds"""
return tokens * self.codec.hop_length / self.codec.sample_rate
def to(self, device):
self.device = device
self.coarse.to(device)
self.codec.to(device)
if self.c2f is not None:
self.c2f.to(device)
if self.beat_tracker is not None:
self.beat_tracker.model.to(device)
return self
def decode(self, z: torch.Tensor):
return self.coarse.decode(z, self.codec)
def _preprocess(self, signal: AudioSignal):
signal = (
signal.clone()
.resample(self.codec.sample_rate)
.to_mono()
.normalize(self.loudness)
.ensure_max_of_audio(1.0)
)
logging.debug(f"length before codec preproc: {signal.samples.shape}")
signal.samples, length = self.codec.preprocess(signal.samples, signal.sample_rate)
logging.debug(f"length after codec preproc: {signal.samples.shape}")
return signal
@torch.inference_mode()
def encode(self, signal: AudioSignal):
signal = signal.to(self.device)
signal = self._preprocess(signal)
z = self.codec.encode(signal.samples, signal.sample_rate)["codes"]
return z
def snap_to_beats(
self,
signal: AudioSignal
):
assert hasattr(self, "beat_tracker"), "No beat tracker loaded"
beats, downbeats = self.beat_tracker.extract_beats(signal)
# trim the signa around the first beat time
samples_begin = int(beats[0] * signal.sample_rate )
samples_end = int(beats[-1] * signal.sample_rate)
logging.debug(beats[0])
signal = signal.clone().trim(samples_begin, signal.length - samples_end)
return signal
def make_beat_mask(self,
signal: AudioSignal,
before_beat_s: float = 0.0,
after_beat_s: float = 0.02,
mask_downbeats: bool = True,
mask_upbeats: bool = True,
downbeat_downsample_factor: int = None,
beat_downsample_factor: int = None,
dropout: float = 0.0,
invert: bool = True,
):
"""make a beat synced mask. that is, make a mask that
places 1s at and around the beat, and 0s everywhere else.
"""
assert self.beat_tracker is not None, "No beat tracker loaded"
# get the beat times
beats, downbeats = self.beat_tracker.extract_beats(signal)
# get the beat indices in z
beats_z, downbeats_z = self.s2t(beats), self.s2t(downbeats)
# remove downbeats from beats
beats_z = torch.tensor(beats_z)[~torch.isin(torch.tensor(beats_z), torch.tensor(downbeats_z))]
beats_z = beats_z.tolist()
downbeats_z = downbeats_z.tolist()
# make the mask
seq_len = self.s2t(signal.duration)
mask = torch.zeros(seq_len, device=self.device)
mask_b4 = self.s2t(before_beat_s)
mask_after = self.s2t(after_beat_s)
if beat_downsample_factor is not None:
if beat_downsample_factor < 1:
raise ValueError("mask_beat_downsample_factor must be >= 1 or None")
else:
beat_downsample_factor = 1
if downbeat_downsample_factor is not None:
if downbeat_downsample_factor < 1:
raise ValueError("mask_beat_downsample_factor must be >= 1 or None")
else:
downbeat_downsample_factor = 1
beats_z = beats_z[::beat_downsample_factor]
downbeats_z = downbeats_z[::downbeat_downsample_factor]
logging.debug(f"beats_z: {len(beats_z)}")
logging.debug(f"downbeats_z: {len(downbeats_z)}")
if mask_upbeats:
for beat_idx in beats_z:
_slice = int(beat_idx - mask_b4), int(beat_idx + mask_after)
num_steps = mask[_slice[0]:_slice[1]].shape[0]
_m = torch.ones(num_steps, device=self.device)
_m_mask = torch.bernoulli(_m * (1 - dropout))
_m = _m * _m_mask.long()
mask[_slice[0]:_slice[1]] = _m
if mask_downbeats:
for downbeat_idx in downbeats_z:
_slice = int(downbeat_idx - mask_b4), int(downbeat_idx + mask_after)
num_steps = mask[_slice[0]:_slice[1]].shape[0]
_m = torch.ones(num_steps, device=self.device)
_m_mask = torch.bernoulli(_m * (1 - dropout))
_m = _m * _m_mask.long()
mask[_slice[0]:_slice[1]] = _m
mask = mask.clamp(0, 1)
if invert:
mask = 1 - mask
mask = mask[None, None, :].bool().long()
if self.c2f is not None:
mask = mask.repeat(1, self.c2f.n_codebooks, 1)
else:
mask = mask.repeat(1, self.coarse.n_codebooks, 1)
return mask
def set_chunk_size(self, chunk_size_s: float):
self.coarse.chunk_size_s = chunk_size_s
@torch.inference_mode()
def coarse_to_fine(
self,
z: torch.Tensor,
mask: torch.Tensor = None,
return_mask: bool = False,
**kwargs
):
assert self.c2f is not None, "No coarse2fine model loaded"
length = z.shape[-1]
chunk_len = self.s2t(self.c2f.chunk_size_s)
n_chunks = math.ceil(z.shape[-1] / chunk_len)
# zero pad to chunk_len
if length % chunk_len != 0:
pad_len = chunk_len - (length % chunk_len)
z = torch.nn.functional.pad(z, (0, pad_len))
mask = torch.nn.functional.pad(mask, (0, pad_len), value=1) if mask is not None else None
n_codebooks_to_append = self.c2f.n_codebooks - z.shape[1]
if n_codebooks_to_append > 0:
z = torch.cat([
z,
torch.zeros(z.shape[0], n_codebooks_to_append, z.shape[-1]).long().to(self.device)
], dim=1)
logging.debug(f"appended {n_codebooks_to_append} codebooks to z")
# set the mask to 0 for all conditioning codebooks
if mask is not None:
mask = mask.clone()
mask[:, :self.c2f.n_conditioning_codebooks, :] = 0
fine_z = []
for i in range(n_chunks):
chunk = z[:, :, i * chunk_len : (i + 1) * chunk_len]
mask_chunk = mask[:, :, i * chunk_len : (i + 1) * chunk_len] if mask is not None else None
with torch.autocast("cuda", dtype=torch.bfloat16):
chunk = self.c2f.generate(
codec=self.codec,
time_steps=chunk_len,
start_tokens=chunk,
return_signal=False,
mask=mask_chunk,
cfg_guidance=None,
**kwargs
)
fine_z.append(chunk)
fine_z = torch.cat(fine_z, dim=-1)
if return_mask:
return fine_z[:, :, :length].clone(), apply_mask(fine_z, mask, self.c2f.mask_token)[0][:, :, :length].clone()
return fine_z[:, :, :length].clone()
@torch.inference_mode()
def coarse_vamp(
self,
z,
mask,
return_mask=False,
gen_fn=None,
**kwargs
):
# coarse z
cz = z[:, : self.coarse.n_codebooks, :].clone()
mask = mask[:, : self.coarse.n_codebooks, :]
# assert cz.shape[-1] <= self.s2t(self.coarse.chunk_size_s), f"the sequence of tokens provided must match the one specified in the coarse chunk size, but got {cz.shape[-1]} and {self.s2t(self.coarse.chunk_size_s)}"
# cut into chunks, keep the last chunk separate if it's too small
chunk_len = self.s2t(self.coarse.chunk_size_s)
n_chunks = math.ceil(cz.shape[-1] / chunk_len)
last_chunk_len = cz.shape[-1] % chunk_len
cz_chunks = []
mask_chunks = []
for i in range(n_chunks):
chunk = cz[:, :, i * chunk_len : (i + 1) * chunk_len]
mask_chunk = mask[:, :, i * chunk_len : (i + 1) * chunk_len]
# make sure that the very first and last timestep of each chunk is 0 so that we don't get a weird
# discontinuity when we stitch the chunks back together
# only if there's already a 0 somewhere in the chunk
if torch.any(mask_chunk == 0):
mask_chunk[:, :, 0] = 0
mask_chunk[:, :, -1] = 0
cz_chunks.append(chunk)
mask_chunks.append(mask_chunk)
# now vamp each chunk
cz_masked_chunks = []
cz_vamped_chunks = []
for chunk, mask_chunk in zip(cz_chunks, mask_chunks):
cz_masked_chunk, mask_chunk = apply_mask(chunk, mask_chunk, self.coarse.mask_token)
cz_masked_chunk = cz_masked_chunk[:, : self.coarse.n_codebooks, :]
cz_masked_chunks.append(cz_masked_chunk)
gen_fn = gen_fn or self.coarse.generate
with torch.autocast("cuda", dtype=torch.bfloat16):
c_vamp_chunk = gen_fn(
codec=self.codec,
time_steps=chunk_len,
start_tokens=cz_masked_chunk,
return_signal=False,
mask=mask_chunk,
**kwargs
)
cz_vamped_chunks.append(c_vamp_chunk)
# stitch the chunks back together
cz_masked = torch.cat(cz_masked_chunks, dim=-1)
c_vamp = torch.cat(cz_vamped_chunks, dim=-1)
# add the fine codes back in
c_vamp = torch.cat(
[c_vamp, z[:, self.coarse.n_codebooks :, :]],
dim=1
)
if return_mask:
return c_vamp, cz_masked
return c_vamp
def build_mask(self,
z: torch.Tensor,
sig: AudioSignal = None,
rand_mask_intensity: float = 1.0,
prefix_s: float = 0.0,
suffix_s: float = 0.0,
periodic_prompt: int = 7,
periodic_prompt_width: int = 1,
onset_mask_width: int = 0,
_dropout: float = 0.0,
upper_codebook_mask: int = 3,
ncc: int = 0,
):
mask = linear_random(z, rand_mask_intensity)
mask = mask_and(
mask,
inpaint(z, self.s2t(prefix_s), self.s2t(suffix_s)),
)
pmask = periodic_mask(z, periodic_prompt, periodic_prompt_width, random_roll=True)
mask = mask_and(mask, pmask)
if onset_mask_width > 0:
assert sig is not None, f"must provide a signal to use onset mask"
mask = mask_and(
mask, onset_mask(
sig, z, self,
width=onset_mask_width
)
)
mask = dropout(mask, _dropout)
mask = codebook_unmask(mask, ncc)
mask = codebook_mask(mask, int(upper_codebook_mask), None)
return mask
def vamp(
self,
codes: torch.Tensor,
mask: torch.Tensor,
batch_size: int = 1,
feedback_steps: int = 1,
time_stretch_factor: int = 1,
return_mask: bool = False,
**kwargs,
):
z = codes
# expand z to batch size
z = z.expand(batch_size, -1, -1)
mask = mask.expand(batch_size, -1, -1)
# stretch mask and z to match the time stretch factor
# we'll add (stretch_factor - 1) mask tokens in between each timestep of z
# and we'll make the mask 1 in all the new slots we added
if time_stretch_factor > 1:
z = z.repeat_interleave(time_stretch_factor, dim=-1)
mask = mask.repeat_interleave(time_stretch_factor, dim=-1)
added_mask = torch.ones_like(mask)
added_mask[:, :, ::time_stretch_factor] = 0
mask = mask.bool() | added_mask.bool()
mask = mask.long()
# the forward pass
logging.debug(z.shape)
logging.debug("coarse!")
zv, mask_z = self.coarse_vamp(
z,
mask=mask,
return_mask=True,
**kwargs
)
# add the top codebooks back in
if zv.shape[1] < z.shape[1]:
logging.debug(f"adding {z.shape[1] - zv.shape[1]} codebooks back in")
zv = torch.cat(
[zv, z[:, self.coarse.n_codebooks :, :]],
dim=1
)
# now, coarse2fine
logging.debug(f"coarse2fine!")
zv, fine_zv_mask = self.coarse_to_fine(
zv,
mask=mask,
typical_filtering=True,
_sampling_steps=2,
return_mask=True
)
mask_z = torch.cat(
[mask_z[:, :self.coarse.n_codebooks, :], fine_zv_mask[:, self.coarse.n_codebooks:, :]],
dim=1
)
z = zv
if return_mask:
return z, mask_z.cpu(),
else:
return z
def visualize_codes(self, z: torch.Tensor):
import matplotlib.pyplot as plt
# make sure the figsize is square when imshow is called
fig = plt.figure(figsize=(10, 7))
# in subplots, plot z[0] and the mask
# set title to "codes" and "mask"
fig.add_subplot(2, 1, 1)
plt.imshow(z[0].cpu().numpy(), aspect='auto', origin='lower', cmap="tab20")
plt.title("codes")
plt.ylabel("codebook index")
# set the xticks to seconds
if __name__ == "__main__":
import audiotools as at
import logging
logger = logging.getLogger()
logger.setLevel(logging.INFO)
torch.set_logging.debugoptions(threshold=10000)
at.util.seed(42)
interface = Interface(
coarse_ckpt="./models/vampnet/coarse.pth",
coarse2fine_ckpt="./models/vampnet/c2f.pth",
codec_ckpt="./models/vampnet/codec.pth",
device="cuda",
wavebeat_ckpt="./models/wavebeat.pth"
)
sig = at.AudioSignal('assets/example.wav')
z = interface.encode(sig)
mask = interface.build_mask(
z=z,
sig=sig,
rand_mask_intensity=1.0,
prefix_s=0.0,
suffix_s=0.0,
periodic_prompt=7,
periodic_prompt2=7,
periodic_prompt_width=1,
onset_mask_width=5,
_dropout=0.0,
upper_codebook_mask=3,
upper_codebook_mask_2=None,
ncc=0,
)
zv, mask_z = interface.coarse_vamp(
z,
mask=mask,
return_mask=True,
gen_fn=interface.coarse.generate
)
use_coarse2fine = True
if use_coarse2fine:
zv = interface.coarse_to_fine(zv, mask=mask)
breakpoint()
mask = interface.decode(mask_z).cpu()
sig = interface.decode(zv).cpu()
logging.debug("done")
|