File size: 21,476 Bytes
074c857 |
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 |
import math
from scipy import integrate
import torch
from torch import nn
from torchdiffeq import odeint
from tqdm.auto import trange, tqdm
from . import utils
def append_zero(x):
return torch.cat([x, x.new_zeros([1])])
def get_sigmas_karras(n, sigma_min, sigma_max, rho=7., device='cpu'):
"""Constructs the noise schedule of Karras et al. (2022)."""
ramp = torch.linspace(0, 1, n)
min_inv_rho = sigma_min ** (1 / rho)
max_inv_rho = sigma_max ** (1 / rho)
sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
return append_zero(sigmas).to(device)
def get_sigmas_exponential(n, sigma_min, sigma_max, device='cpu'):
"""Constructs an exponential noise schedule."""
sigmas = torch.linspace(math.log(sigma_max), math.log(sigma_min), n, device=device).exp()
return append_zero(sigmas)
def get_sigmas_vp(n, beta_d=19.9, beta_min=0.1, eps_s=1e-3, device='cpu'):
"""Constructs a continuous VP noise schedule."""
t = torch.linspace(1, eps_s, n, device=device)
sigmas = torch.sqrt(torch.exp(beta_d * t ** 2 / 2 + beta_min * t) - 1)
return append_zero(sigmas)
def to_d(x, sigma, denoised):
"""Converts a denoiser output to a Karras ODE derivative."""
return (x - denoised) / utils.append_dims(sigma, x.ndim)
def get_ancestral_step(sigma_from, sigma_to, eta=1.):
"""Calculates the noise level (sigma_down) to step down to and the amount
of noise to add (sigma_up) when doing an ancestral sampling step."""
if not eta:
return sigma_to, 0.
sigma_up = min(sigma_to, eta * (sigma_to ** 2 * (sigma_from ** 2 - sigma_to ** 2) / sigma_from ** 2) ** 0.5)
sigma_down = (sigma_to ** 2 - sigma_up ** 2) ** 0.5
return sigma_down, sigma_up
@torch.no_grad()
def sample_euler(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
"""Implements Algorithm 2 (Euler steps) from Karras et al. (2022)."""
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
for i in trange(len(sigmas) - 1, disable=disable):
gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
eps = torch.randn_like(x) * s_noise
sigma_hat = sigmas[i] * (gamma + 1)
if gamma > 0:
x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
denoised = model(x, sigma_hat * s_in, **extra_args)
d = to_d(x, sigma_hat, denoised)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
dt = sigmas[i + 1] - sigma_hat
# Euler method
x = x + d * dt
return x
@torch.no_grad()
def sample_euler_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1.):
"""Ancestral sampling with Euler method steps."""
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
d = to_d(x, sigmas[i], denoised)
# Euler method
dt = sigma_down - sigmas[i]
x = x + d * dt
x = x + torch.randn_like(x) * sigma_up
return x
@torch.no_grad()
def sample_heun(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
"""Implements Algorithm 2 (Heun steps) from Karras et al. (2022)."""
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
for i in trange(len(sigmas) - 1, disable=disable):
gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
eps = torch.randn_like(x) * s_noise
sigma_hat = sigmas[i] * (gamma + 1)
if gamma > 0:
x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
denoised = model(x, sigma_hat * s_in, **extra_args)
d = to_d(x, sigma_hat, denoised)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
dt = sigmas[i + 1] - sigma_hat
if sigmas[i + 1] == 0:
# Euler method
x = x + d * dt
else:
# Heun's method
x_2 = x + d * dt
denoised_2 = model(x_2, sigmas[i + 1] * s_in, **extra_args)
d_2 = to_d(x_2, sigmas[i + 1], denoised_2)
d_prime = (d + d_2) / 2
x = x + d_prime * dt
return x
@torch.no_grad()
def sample_dpm_2(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
"""A sampler inspired by DPM-Solver-2 and Algorithm 2 from Karras et al. (2022)."""
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
for i in trange(len(sigmas) - 1, disable=disable):
gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
eps = torch.randn_like(x) * s_noise
sigma_hat = sigmas[i] * (gamma + 1)
if gamma > 0:
x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
denoised = model(x, sigma_hat * s_in, **extra_args)
d = to_d(x, sigma_hat, denoised)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
if sigmas[i + 1] == 0:
# Euler method
dt = sigmas[i + 1] - sigma_hat
x = x + d * dt
else:
# DPM-Solver-2
sigma_mid = sigma_hat.log().lerp(sigmas[i + 1].log(), 0.5).exp()
dt_1 = sigma_mid - sigma_hat
dt_2 = sigmas[i + 1] - sigma_hat
x_2 = x + d * dt_1
denoised_2 = model(x_2, sigma_mid * s_in, **extra_args)
d_2 = to_d(x_2, sigma_mid, denoised_2)
x = x + d_2 * dt_2
return x
@torch.no_grad()
def sample_dpm_2_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1.):
"""Ancestral sampling with DPM-Solver inspired second-order steps."""
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
d = to_d(x, sigmas[i], denoised)
if sigma_down == 0:
# Euler method
dt = sigma_down - sigmas[i]
x = x + d * dt
else:
# DPM-Solver-2
sigma_mid = sigmas[i].log().lerp(sigma_down.log(), 0.5).exp()
dt_1 = sigma_mid - sigmas[i]
dt_2 = sigma_down - sigmas[i]
x_2 = x + d * dt_1
denoised_2 = model(x_2, sigma_mid * s_in, **extra_args)
d_2 = to_d(x_2, sigma_mid, denoised_2)
x = x + d_2 * dt_2
x = x + torch.randn_like(x) * sigma_up
return x
def linear_multistep_coeff(order, t, i, j):
if order - 1 > i:
raise ValueError(f'Order {order} too high for step {i}')
def fn(tau):
prod = 1.
for k in range(order):
if j == k:
continue
prod *= (tau - t[i - k]) / (t[i - j] - t[i - k])
return prod
return integrate.quad(fn, t[i], t[i + 1], epsrel=1e-4)[0]
@torch.no_grad()
def sample_lms(model, x, sigmas, extra_args=None, callback=None, disable=None, order=4):
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
sigmas_cpu = sigmas.detach().cpu().numpy()
ds = []
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
d = to_d(x, sigmas[i], denoised)
ds.append(d)
if len(ds) > order:
ds.pop(0)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
cur_order = min(i + 1, order)
coeffs = [linear_multistep_coeff(cur_order, sigmas_cpu, i, j) for j in range(cur_order)]
x = x + sum(coeff * d for coeff, d in zip(coeffs, reversed(ds)))
return x
@torch.no_grad()
def log_likelihood(model, x, sigma_min, sigma_max, extra_args=None, atol=1e-4, rtol=1e-4):
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
v = torch.randint_like(x, 2) * 2 - 1
fevals = 0
def ode_fn(sigma, x):
nonlocal fevals
with torch.enable_grad():
x = x[0].detach().requires_grad_()
denoised = model(x, sigma * s_in, **extra_args)
d = to_d(x, sigma, denoised)
fevals += 1
grad = torch.autograd.grad((d * v).sum(), x)[0]
d_ll = (v * grad).flatten(1).sum(1)
return d.detach(), d_ll
x_min = x, x.new_zeros([x.shape[0]])
t = x.new_tensor([sigma_min, sigma_max])
sol = odeint(ode_fn, x_min, t, atol=atol, rtol=rtol, method='dopri5')
latent, delta_ll = sol[0][-1], sol[1][-1]
ll_prior = torch.distributions.Normal(0, sigma_max).log_prob(latent).flatten(1).sum(1)
return ll_prior + delta_ll, {'fevals': fevals}
class PIDStepSizeController:
"""A PID controller for ODE adaptive step size control."""
def __init__(self, h, pcoeff, icoeff, dcoeff, order=1, accept_safety=0.81, eps=1e-8):
self.h = h
self.b1 = (pcoeff + icoeff + dcoeff) / order
self.b2 = -(pcoeff + 2 * dcoeff) / order
self.b3 = dcoeff / order
self.accept_safety = accept_safety
self.eps = eps
self.errs = []
def limiter(self, x):
return 1 + math.atan(x - 1)
def propose_step(self, error):
inv_error = 1 / (float(error) + self.eps)
if not self.errs:
self.errs = [inv_error, inv_error, inv_error]
self.errs[0] = inv_error
factor = self.errs[0] ** self.b1 * self.errs[1] ** self.b2 * self.errs[2] ** self.b3
factor = self.limiter(factor)
accept = factor >= self.accept_safety
if accept:
self.errs[2] = self.errs[1]
self.errs[1] = self.errs[0]
self.h *= factor
return accept
class DPMSolver(nn.Module):
"""DPM-Solver. See https://arxiv.org/abs/2206.00927."""
def __init__(self, model, extra_args=None, eps_callback=None, info_callback=None):
super().__init__()
self.model = model
self.extra_args = {} if extra_args is None else extra_args
self.eps_callback = eps_callback
self.info_callback = info_callback
def t(self, sigma):
return -sigma.log()
def sigma(self, t):
return t.neg().exp()
def eps(self, eps_cache, key, x, t, *args, **kwargs):
if key in eps_cache:
return eps_cache[key], eps_cache
sigma = self.sigma(t) * x.new_ones([x.shape[0]])
eps = (x - self.model(x, sigma, *args, **self.extra_args, **kwargs)) / self.sigma(t)
if self.eps_callback is not None:
self.eps_callback()
return eps, {key: eps, **eps_cache}
def dpm_solver_1_step(self, x, t, t_next, eps_cache=None):
eps_cache = {} if eps_cache is None else eps_cache
h = t_next - t
eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
x_1 = x - self.sigma(t_next) * h.expm1() * eps
return x_1, eps_cache
def dpm_solver_2_step(self, x, t, t_next, r1=1 / 2, eps_cache=None):
eps_cache = {} if eps_cache is None else eps_cache
h = t_next - t
eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
s1 = t + r1 * h
u1 = x - self.sigma(s1) * (r1 * h).expm1() * eps
eps_r1, eps_cache = self.eps(eps_cache, 'eps_r1', u1, s1)
x_2 = x - self.sigma(t_next) * h.expm1() * eps - self.sigma(t_next) / (2 * r1) * h.expm1() * (eps_r1 - eps)
return x_2, eps_cache
def dpm_solver_3_step(self, x, t, t_next, r1=1 / 3, r2=2 / 3, eps_cache=None):
eps_cache = {} if eps_cache is None else eps_cache
h = t_next - t
eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
s1 = t + r1 * h
s2 = t + r2 * h
u1 = x - self.sigma(s1) * (r1 * h).expm1() * eps
eps_r1, eps_cache = self.eps(eps_cache, 'eps_r1', u1, s1)
u2 = x - self.sigma(s2) * (r2 * h).expm1() * eps - self.sigma(s2) * (r2 / r1) * ((r2 * h).expm1() / (r2 * h) - 1) * (eps_r1 - eps)
eps_r2, eps_cache = self.eps(eps_cache, 'eps_r2', u2, s2)
x_3 = x - self.sigma(t_next) * h.expm1() * eps - self.sigma(t_next) / r2 * (h.expm1() / h - 1) * (eps_r2 - eps)
return x_3, eps_cache
def dpm_solver_fast(self, x, t_start, t_end, nfe, eta=0., s_noise=1.):
if not t_end > t_start and eta:
raise ValueError('eta must be 0 for reverse sampling')
m = math.floor(nfe / 3) + 1
ts = torch.linspace(t_start, t_end, m + 1, device=x.device)
if nfe % 3 == 0:
orders = [3] * (m - 2) + [2, 1]
else:
orders = [3] * (m - 1) + [nfe % 3]
for i in range(len(orders)):
eps_cache = {}
t, t_next = ts[i], ts[i + 1]
if eta:
sd, su = get_ancestral_step(self.sigma(t), self.sigma(t_next), eta)
t_next_ = torch.minimum(t_end, self.t(sd))
su = (self.sigma(t_next) ** 2 - self.sigma(t_next_) ** 2) ** 0.5
else:
t_next_, su = t_next, 0.
eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
denoised = x - self.sigma(t) * eps
if self.info_callback is not None:
self.info_callback({'x': x, 'i': i, 't': ts[i], 't_up': t, 'denoised': denoised})
if orders[i] == 1:
x, eps_cache = self.dpm_solver_1_step(x, t, t_next_, eps_cache=eps_cache)
elif orders[i] == 2:
x, eps_cache = self.dpm_solver_2_step(x, t, t_next_, eps_cache=eps_cache)
else:
x, eps_cache = self.dpm_solver_3_step(x, t, t_next_, eps_cache=eps_cache)
x = x + su * s_noise * torch.randn_like(x)
return x
def dpm_solver_adaptive(self, x, t_start, t_end, order=3, rtol=0.05, atol=0.0078, h_init=0.05, pcoeff=0., icoeff=1., dcoeff=0., accept_safety=0.81, eta=0., s_noise=1.):
if order not in {2, 3}:
raise ValueError('order should be 2 or 3')
forward = t_end > t_start
if not forward and eta:
raise ValueError('eta must be 0 for reverse sampling')
h_init = abs(h_init) * (1 if forward else -1)
atol = torch.tensor(atol)
rtol = torch.tensor(rtol)
s = t_start
x_prev = x
accept = True
pid = PIDStepSizeController(h_init, pcoeff, icoeff, dcoeff, 1.5 if eta else order, accept_safety)
info = {'steps': 0, 'nfe': 0, 'n_accept': 0, 'n_reject': 0}
while s < t_end - 1e-5 if forward else s > t_end + 1e-5:
eps_cache = {}
t = torch.minimum(t_end, s + pid.h) if forward else torch.maximum(t_end, s + pid.h)
if eta:
sd, su = get_ancestral_step(self.sigma(s), self.sigma(t), eta)
t_ = torch.minimum(t_end, self.t(sd))
su = (self.sigma(t) ** 2 - self.sigma(t_) ** 2) ** 0.5
else:
t_, su = t, 0.
eps, eps_cache = self.eps(eps_cache, 'eps', x, s)
denoised = x - self.sigma(s) * eps
if order == 2:
x_low, eps_cache = self.dpm_solver_1_step(x, s, t_, eps_cache=eps_cache)
x_high, eps_cache = self.dpm_solver_2_step(x, s, t_, eps_cache=eps_cache)
else:
x_low, eps_cache = self.dpm_solver_2_step(x, s, t_, r1=1 / 3, eps_cache=eps_cache)
x_high, eps_cache = self.dpm_solver_3_step(x, s, t_, eps_cache=eps_cache)
delta = torch.maximum(atol, rtol * torch.maximum(x_low.abs(), x_prev.abs()))
error = torch.linalg.norm((x_low - x_high) / delta) / x.numel() ** 0.5
accept = pid.propose_step(error)
if accept:
x_prev = x_low
x = x_high + su * s_noise * torch.randn_like(x_high)
s = t
info['n_accept'] += 1
else:
info['n_reject'] += 1
info['nfe'] += order
info['steps'] += 1
if self.info_callback is not None:
self.info_callback({'x': x, 'i': info['steps'] - 1, 't': s, 't_up': s, 'denoised': denoised, 'error': error, 'h': pid.h, **info})
return x, info
@torch.no_grad()
def sample_dpm_fast(model, x, sigma_min, sigma_max, n, extra_args=None, callback=None, disable=None, eta=0., s_noise=1.):
"""DPM-Solver-Fast (fixed step size). See https://arxiv.org/abs/2206.00927."""
if sigma_min <= 0 or sigma_max <= 0:
raise ValueError('sigma_min and sigma_max must not be 0')
with tqdm(total=n, disable=disable) as pbar:
dpm_solver = DPMSolver(model, extra_args, eps_callback=pbar.update)
if callback is not None:
dpm_solver.info_callback = lambda info: callback({'sigma': dpm_solver.sigma(info['t']), 'sigma_hat': dpm_solver.sigma(info['t_up']), **info})
return dpm_solver.dpm_solver_fast(x, dpm_solver.t(torch.tensor(sigma_max)), dpm_solver.t(torch.tensor(sigma_min)), n, eta, s_noise)
@torch.no_grad()
def sample_dpm_adaptive(model, x, sigma_min, sigma_max, extra_args=None, callback=None, disable=None, order=3, rtol=0.05, atol=0.0078, h_init=0.05, pcoeff=0., icoeff=1., dcoeff=0., accept_safety=0.81, eta=0., s_noise=1., return_info=False):
"""DPM-Solver-12 and 23 (adaptive step size). See https://arxiv.org/abs/2206.00927."""
if sigma_min <= 0 or sigma_max <= 0:
raise ValueError('sigma_min and sigma_max must not be 0')
with tqdm(disable=disable) as pbar:
dpm_solver = DPMSolver(model, extra_args, eps_callback=pbar.update)
if callback is not None:
dpm_solver.info_callback = lambda info: callback({'sigma': dpm_solver.sigma(info['t']), 'sigma_hat': dpm_solver.sigma(info['t_up']), **info})
x, info = dpm_solver.dpm_solver_adaptive(x, dpm_solver.t(torch.tensor(sigma_max)), dpm_solver.t(torch.tensor(sigma_min)), order, rtol, atol, h_init, pcoeff, icoeff, dcoeff, accept_safety, eta, s_noise)
if return_info:
return x, info
return x
@torch.no_grad()
def sample_dpmpp_2s_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1.):
"""Ancestral sampling with DPM-Solver++(2S) second-order steps."""
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
sigma_fn = lambda t: t.neg().exp()
t_fn = lambda sigma: sigma.log().neg()
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
if sigma_down == 0:
# Euler method
d = to_d(x, sigmas[i], denoised)
dt = sigma_down - sigmas[i]
x = x + d * dt
else:
# DPM-Solver-2++(2S)
t, t_next = t_fn(sigmas[i]), t_fn(sigma_down)
r = 1 / 2
h = t_next - t
s = t + r * h
x_2 = (sigma_fn(s) / sigma_fn(t)) * x - (-h * r).expm1() * denoised
denoised_2 = model(x_2, sigma_fn(s) * s_in, **extra_args)
x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised_2
# Noise addition
x = x + torch.randn_like(x) * s_noise * sigma_up
return x
@torch.no_grad()
def sample_dpmpp_2m(model, x, sigmas, extra_args=None, callback=None, disable=None):
"""DPM-Solver++(2M)."""
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
sigma_fn = lambda t: t.neg().exp()
t_fn = lambda sigma: sigma.log().neg()
old_denoised = None
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1])
h = t_next - t
if old_denoised is None or sigmas[i + 1] == 0:
x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised
else:
h_last = t - t_fn(sigmas[i - 1])
r = h_last / h
denoised_d = (1 + 1 / (2 * r)) * denoised - (1 / (2 * r)) * old_denoised
x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised_d
old_denoised = denoised
return x
|