Evgeny Zhukov
Origin: https://github.com/ali-vilab/UniAnimate/commit/d7814fa44a0a1154524b92fce0e3133a2604d333
2ba4412
"""
GaussianDiffusion wraps operators for denoising diffusion models, including the
diffusion and denoising processes, as well as the loss evaluation.
"""
import torch
import torchsde
import random
from tqdm.auto import trange
__all__ = ['GaussianDiffusion']
def _i(tensor, t, x):
"""
Index tensor using t and format the output according to x.
"""
shape = (x.size(0), ) + (1, ) * (x.ndim - 1)
return tensor[t.to(tensor.device)].view(shape).to(x.device)
class BatchedBrownianTree:
"""
A wrapper around torchsde.BrownianTree that enables batches of entropy.
"""
def __init__(self, x, t0, t1, seed=None, **kwargs):
t0, t1, self.sign = self.sort(t0, t1)
w0 = kwargs.get('w0', torch.zeros_like(x))
if seed is None:
seed = torch.randint(0, 2 ** 63 - 1, []).item()
self.batched = True
try:
assert len(seed) == x.shape[0]
w0 = w0[0]
except TypeError:
seed = [seed]
self.batched = False
self.trees = [torchsde.BrownianTree(
t0, w0, t1, entropy=s, **kwargs
) for s in seed]
@staticmethod
def sort(a, b):
return (a, b, 1) if a < b else (b, a, -1)
def __call__(self, t0, t1):
t0, t1, sign = self.sort(t0, t1)
w = torch.stack([tree(t0, t1) for tree in self.trees]) * (self.sign * sign)
return w if self.batched else w[0]
class BrownianTreeNoiseSampler:
"""
A noise sampler backed by a torchsde.BrownianTree.
Args:
x (Tensor): The tensor whose shape, device and dtype to use to generate
random samples.
sigma_min (float): The low end of the valid interval.
sigma_max (float): The high end of the valid interval.
seed (int or List[int]): The random seed. If a list of seeds is
supplied instead of a single integer, then the noise sampler will
use one BrownianTree per batch item, each with its own seed.
transform (callable): A function that maps sigma to the sampler's
internal timestep.
"""
def __init__(self, x, sigma_min, sigma_max, seed=None, transform=lambda x: x):
self.transform = transform
t0 = self.transform(torch.as_tensor(sigma_min))
t1 = self.transform(torch.as_tensor(sigma_max))
self.tree = BatchedBrownianTree(x, t0, t1, seed)
def __call__(self, sigma, sigma_next):
t0 = self.transform(torch.as_tensor(sigma))
t1 = self.transform(torch.as_tensor(sigma_next))
return self.tree(t0, t1) / (t1 - t0).abs().sqrt()
def get_scalings(sigma):
c_out = -sigma
c_in = 1 / (sigma ** 2 + 1. ** 2) ** 0.5
return c_out, c_in
@torch.no_grad()
def sample_dpmpp_2m_sde(
noise,
model,
sigmas,
eta=1.,
s_noise=1.,
solver_type='midpoint',
show_progress=True
):
"""
DPM-Solver++ (2M) SDE.
"""
assert solver_type in {'heun', 'midpoint'}
x = noise * sigmas[0]
sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas[sigmas < float('inf')].max()
noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max)
old_denoised = None
h_last = None
for i in trange(len(sigmas) - 1, disable=not show_progress):
if sigmas[i] == float('inf'):
# Euler method
denoised = model(noise, sigmas[i])
x = denoised + sigmas[i + 1] * noise
else:
_, c_in = get_scalings(sigmas[i])
denoised = model(x * c_in, sigmas[i])
if sigmas[i + 1] == 0:
# Denoising step
x = denoised
else:
# DPM-Solver++(2M) SDE
t, s = -sigmas[i].log(), -sigmas[i + 1].log()
h = s - t
eta_h = eta * h
x = sigmas[i + 1] / sigmas[i] * (-eta_h).exp() * x + \
(-h - eta_h).expm1().neg() * denoised
if old_denoised is not None:
r = h_last / h
if solver_type == 'heun':
x = x + ((-h - eta_h).expm1().neg() / (-h - eta_h) + 1) * \
(1 / r) * (denoised - old_denoised)
elif solver_type == 'midpoint':
x = x + 0.5 * (-h - eta_h).expm1().neg() * \
(1 / r) * (denoised - old_denoised)
x = x + noise_sampler(
sigmas[i],
sigmas[i + 1]
) * sigmas[i + 1] * (-2 * eta_h).expm1().neg().sqrt() * s_noise
old_denoised = denoised
h_last = h
return x
class GaussianDiffusion(object):
def __init__(self, sigmas, prediction_type='eps'):
assert prediction_type in {'x0', 'eps', 'v'}
self.sigmas = sigmas.float() # noise coefficients
self.alphas = torch.sqrt(1 - sigmas ** 2).float() # signal coefficients
self.num_timesteps = len(sigmas)
self.prediction_type = prediction_type
def diffuse(self, x0, t, noise=None):
"""
Add Gaussian noise to signal x0 according to:
q(x_t | x_0) = N(x_t | alpha_t x_0, sigma_t^2 I).
"""
noise = torch.randn_like(x0) if noise is None else noise
xt = _i(self.alphas, t, x0) * x0 + _i(self.sigmas, t, x0) * noise
return xt
def denoise(
self,
xt,
t,
s,
model,
model_kwargs={},
guide_scale=None,
guide_rescale=None,
clamp=None,
percentile=None
):
"""
Apply one step of denoising from the posterior distribution q(x_s | x_t, x0).
Since x0 is not available, estimate the denoising results using the learned
distribution p(x_s | x_t, \hat{x}_0 == f(x_t)).
"""
s = t - 1 if s is None else s
# hyperparams
sigmas = _i(self.sigmas, t, xt)
alphas = _i(self.alphas, t, xt)
alphas_s = _i(self.alphas, s.clamp(0), xt)
alphas_s[s < 0] = 1.
sigmas_s = torch.sqrt(1 - alphas_s ** 2)
# precompute variables
betas = 1 - (alphas / alphas_s) ** 2
coef1 = betas * alphas_s / sigmas ** 2
coef2 = (alphas * sigmas_s ** 2) / (alphas_s * sigmas ** 2)
var = betas * (sigmas_s / sigmas) ** 2
log_var = torch.log(var).clamp_(-20, 20)
# prediction
if guide_scale is None:
assert isinstance(model_kwargs, dict)
out = model(xt, t=t, **model_kwargs)
else:
# classifier-free guidance (arXiv:2207.12598)
# model_kwargs[0]: conditional kwargs
# model_kwargs[1]: non-conditional kwargs
assert isinstance(model_kwargs, list) and len(model_kwargs) == 2
y_out = model(xt, t=t, **model_kwargs[0])
if guide_scale == 1.:
out = y_out
else:
u_out = model(xt, t=t, **model_kwargs[1])
out = u_out + guide_scale * (y_out - u_out)
# rescale the output according to arXiv:2305.08891
if guide_rescale is not None:
assert guide_rescale >= 0 and guide_rescale <= 1
ratio = (y_out.flatten(1).std(dim=1) / (
out.flatten(1).std(dim=1) + 1e-12
)).view((-1, ) + (1, ) * (y_out.ndim - 1))
out *= guide_rescale * ratio + (1 - guide_rescale) * 1.0
# compute x0
if self.prediction_type == 'x0':
x0 = out
elif self.prediction_type == 'eps':
x0 = (xt - sigmas * out) / alphas
elif self.prediction_type == 'v':
x0 = alphas * xt - sigmas * out
else:
raise NotImplementedError(
f'prediction_type {self.prediction_type} not implemented'
)
# restrict the range of x0
if percentile is not None:
# NOTE: percentile should only be used when data is within range [-1, 1]
assert percentile > 0 and percentile <= 1
s = torch.quantile(x0.flatten(1).abs(), percentile, dim=1)
s = s.clamp_(1.0).view((-1, ) + (1, ) * (xt.ndim - 1))
x0 = torch.min(s, torch.max(-s, x0)) / s
elif clamp is not None:
x0 = x0.clamp(-clamp, clamp)
# recompute eps using the restricted x0
eps = (xt - alphas * x0) / sigmas
# compute mu (mean of posterior distribution) using the restricted x0
mu = coef1 * x0 + coef2 * xt
return mu, var, log_var, x0, eps
@torch.no_grad()
def sample(
self,
noise,
model,
model_kwargs={},
condition_fn=None,
guide_scale=None,
guide_rescale=None,
clamp=None,
percentile=None,
solver='euler_a',
steps=20,
t_max=None,
t_min=None,
discretization=None,
discard_penultimate_step=None,
return_intermediate=None,
show_progress=False,
seed=-1,
**kwargs
):
# sanity check
assert isinstance(steps, (int, torch.LongTensor))
assert t_max is None or (t_max > 0 and t_max <= self.num_timesteps - 1)
assert t_min is None or (t_min >= 0 and t_min < self.num_timesteps - 1)
assert discretization in (None, 'leading', 'linspace', 'trailing')
assert discard_penultimate_step in (None, True, False)
assert return_intermediate in (None, 'x0', 'xt')
# function of diffusion solver
solver_fn = {
# 'heun': sample_heun,
'dpmpp_2m_sde': sample_dpmpp_2m_sde
}[solver]
# options
schedule = 'karras' if 'karras' in solver else None
discretization = discretization or 'linspace'
seed = seed if seed >= 0 else random.randint(0, 2 ** 31)
if isinstance(steps, torch.LongTensor):
discard_penultimate_step = False
if discard_penultimate_step is None:
discard_penultimate_step = True if solver in (
'dpm2',
'dpm2_ancestral',
'dpmpp_2m_sde',
'dpm2_karras',
'dpm2_ancestral_karras',
'dpmpp_2m_sde_karras'
) else False
# function for denoising xt to get x0
intermediates = []
def model_fn(xt, sigma):
# denoising
t = self._sigma_to_t(sigma).repeat(len(xt)).round().long()
x0 = self.denoise(
xt, t, None, model, model_kwargs, guide_scale, guide_rescale, clamp,
percentile
)[-2]
# collect intermediate outputs
if return_intermediate == 'xt':
intermediates.append(xt)
elif return_intermediate == 'x0':
intermediates.append(x0)
return x0
# get timesteps
if isinstance(steps, int):
steps += 1 if discard_penultimate_step else 0
t_max = self.num_timesteps - 1 if t_max is None else t_max
t_min = 0 if t_min is None else t_min
# discretize timesteps
if discretization == 'leading':
steps = torch.arange(
t_min, t_max + 1, (t_max - t_min + 1) / steps
).flip(0)
elif discretization == 'linspace':
steps = torch.linspace(t_max, t_min, steps)
elif discretization == 'trailing':
steps = torch.arange(t_max, t_min - 1, -((t_max - t_min + 1) / steps))
else:
raise NotImplementedError(
f'{discretization} discretization not implemented'
)
steps = steps.clamp_(t_min, t_max)
steps = torch.as_tensor(steps, dtype=torch.float32, device=noise.device)
# get sigmas
sigmas = self._t_to_sigma(steps)
sigmas = torch.cat([sigmas, sigmas.new_zeros([1])])
if schedule == 'karras':
if sigmas[0] == float('inf'):
sigmas = karras_schedule(
n=len(steps) - 1,
sigma_min=sigmas[sigmas > 0].min().item(),
sigma_max=sigmas[sigmas < float('inf')].max().item(),
rho=7.
).to(sigmas)
sigmas = torch.cat([
sigmas.new_tensor([float('inf')]), sigmas, sigmas.new_zeros([1])
])
else:
sigmas = karras_schedule(
n=len(steps),
sigma_min=sigmas[sigmas > 0].min().item(),
sigma_max=sigmas.max().item(),
rho=7.
).to(sigmas)
sigmas = torch.cat([sigmas, sigmas.new_zeros([1])])
if discard_penultimate_step:
sigmas = torch.cat([sigmas[:-2], sigmas[-1:]])
# sampling
x0 = solver_fn(
noise,
model_fn,
sigmas,
show_progress=show_progress,
**kwargs
)
return (x0, intermediates) if return_intermediate is not None else x0
@torch.no_grad()
def ddim_reverse_sample(
self,
xt,
t,
model,
model_kwargs={},
clamp=None,
percentile=None,
guide_scale=None,
guide_rescale=None,
ddim_timesteps=20,
reverse_steps=600
):
r"""Sample from p(x_{t+1} | x_t) using DDIM reverse ODE (deterministic).
"""
stride = reverse_steps // ddim_timesteps
# predict distribution of p(x_{t-1} | x_t)
_, _, _, x0, eps = self.denoise(
xt, t, None, model, model_kwargs, guide_scale, guide_rescale, clamp,
percentile
)
# derive variables
s = (t + stride).clamp(0, reverse_steps-1)
# hyperparams
sigmas = _i(self.sigmas, t, xt)
alphas = _i(self.alphas, t, xt)
alphas_s = _i(self.alphas, s.clamp(0), xt)
alphas_s[s < 0] = 1.
sigmas_s = torch.sqrt(1 - alphas_s ** 2)
# reverse sample
mu = alphas_s * x0 + sigmas_s * eps
return mu, x0
@torch.no_grad()
def ddim_reverse_sample_loop(
self,
x0,
model,
model_kwargs={},
clamp=None,
percentile=None,
guide_scale=None,
guide_rescale=None,
ddim_timesteps=20,
reverse_steps=600
):
# prepare input
b = x0.size(0)
xt = x0
# reconstruction steps
steps = torch.arange(0, reverse_steps, reverse_steps // ddim_timesteps)
for step in steps:
t = torch.full((b, ), step, dtype=torch.long, device=xt.device)
xt, _ = self.ddim_reverse_sample(xt, t, model, model_kwargs, clamp, percentile, guide_scale, guide_rescale, ddim_timesteps, reverse_steps)
return xt
def _sigma_to_t(self, sigma):
if sigma == float('inf'):
t = torch.full_like(sigma, len(self.sigmas) - 1)
else:
log_sigmas = torch.sqrt(
self.sigmas ** 2 / (1 - self.sigmas ** 2)
).log().to(sigma)
log_sigma = sigma.log()
dists = log_sigma - log_sigmas[:, None]
low_idx = dists.ge(0).cumsum(dim=0).argmax(dim=0).clamp(
max=log_sigmas.shape[0] - 2
)
high_idx = low_idx + 1
low, high = log_sigmas[low_idx], log_sigmas[high_idx]
w = (low - log_sigma) / (low - high)
w = w.clamp(0, 1)
t = (1 - w) * low_idx + w * high_idx
t = t.view(sigma.shape)
if t.ndim == 0:
t = t.unsqueeze(0)
return t
def _t_to_sigma(self, t):
t = t.float()
low_idx, high_idx, w = t.floor().long(), t.ceil().long(), t.frac()
log_sigmas = torch.sqrt(self.sigmas ** 2 / (1 - self.sigmas ** 2)).log().to(t)
log_sigma = (1 - w) * log_sigmas[low_idx] + w * log_sigmas[high_idx]
log_sigma[torch.isnan(log_sigma) | torch.isinf(log_sigma)] = float('inf')
return log_sigma.exp()
def prev_step(self, model_out, t, xt, inference_steps=50):
prev_t = t - self.num_timesteps // inference_steps
sigmas = _i(self.sigmas, t, xt)
alphas = _i(self.alphas, t, xt)
alphas_prev = _i(self.alphas, prev_t.clamp(0), xt)
alphas_prev[prev_t < 0] = 1.
sigmas_prev = torch.sqrt(1 - alphas_prev ** 2)
x0 = alphas * xt - sigmas * model_out
eps = (xt - alphas * x0) / sigmas
prev_sample = alphas_prev * x0 + sigmas_prev * eps
return prev_sample
def next_step(self, model_out, t, xt, inference_steps=50):
t, next_t = min(t - self.num_timesteps // inference_steps, 999), t
sigmas = _i(self.sigmas, t, xt)
alphas = _i(self.alphas, t, xt)
alphas_next = _i(self.alphas, next_t.clamp(0), xt)
alphas_next[next_t < 0] = 1.
sigmas_next = torch.sqrt(1 - alphas_next ** 2)
x0 = alphas * xt - sigmas * model_out
eps = (xt - alphas * x0) / sigmas
next_sample = alphas_next * x0 + sigmas_next * eps
return next_sample
def get_noise_pred_single(self, xt, t, model, model_kwargs):
assert isinstance(model_kwargs, dict)
out = model(xt, t=t, **model_kwargs)
return out