2024-08-21 17:11:56 +08:00

281 lines
11 KiB
Python

import torch
import numpy as np
from scipy import integrate
from utils.pose import PoseUtil
def global_prior_likelihood(z, sigma_max):
"""The likelihood of a Gaussian distribution with mean zero and
standard deviation sigma."""
# z: [bs, pose_dim]
shape = z.shape
N = np.prod(shape[1:]) # pose_dim
return -N / 2. * torch.log(2 * np.pi * sigma_max ** 2) - torch.sum(z ** 2, dim=-1) / (2 * sigma_max ** 2)
def cond_ode_likelihood(
score_model,
data,
prior,
sde_coeff,
marginal_prob_fn,
atol=1e-5,
rtol=1e-5,
device='cuda',
eps=1e-5,
num_steps=None,
pose_mode='quat_wxyz',
init_x=None,
):
pose_dim = PoseUtil.get_pose_dim(pose_mode)
batch_size = data['pts'].shape[0]
epsilon = prior((batch_size, pose_dim)).to(device)
init_x = data['sampled_pose'].clone().cpu().numpy() if init_x is None else init_x
shape = init_x.shape
init_logp = np.zeros((shape[0],)) # [bs]
init_inp = np.concatenate([init_x.reshape(-1), init_logp], axis=0)
def score_eval_wrapper(data):
"""A wrapper of the score-based model for use by the ODE solver."""
with torch.no_grad():
score = score_model(data)
return score.cpu().numpy().reshape((-1,))
def divergence_eval(data, epsilon):
"""Compute the divergence of the score-based model with Skilling-Hutchinson."""
# save ckpt of sampled_pose
origin_sampled_pose = data['sampled_pose'].clone()
with torch.enable_grad():
# make sampled_pose differentiable
data['sampled_pose'].requires_grad_(True)
score = score_model(data)
score_energy = torch.sum(score * epsilon) # [, ]
grad_score_energy = torch.autograd.grad(score_energy, data['sampled_pose'])[0] # [bs, pose_dim]
# reset sampled_pose
data['sampled_pose'] = origin_sampled_pose
return torch.sum(grad_score_energy * epsilon, dim=-1) # [bs, 1]
def divergence_eval_wrapper(data):
"""A wrapper for evaluating the divergence of score for the black-box ODE solver."""
with torch.no_grad():
# Compute likelihood.
div = divergence_eval(data, epsilon) # [bs, 1]
return div.cpu().numpy().reshape((-1,)).astype(np.float64)
def ode_func(t, inp):
"""The ODE function for use by the ODE solver."""
# split x, logp from inp
x = inp[:-shape[0]]
# calc x-grad
x = torch.tensor(x.reshape(-1, pose_dim), dtype=torch.float32, device=device)
time_steps = torch.ones(batch_size, device=device).unsqueeze(-1) * t
drift, diffusion = sde_coeff(torch.tensor(t))
drift = drift.cpu().numpy()
diffusion = diffusion.cpu().numpy()
data['sampled_pose'] = x
data['t'] = time_steps
x_grad = drift - 0.5 * (diffusion ** 2) * score_eval_wrapper(data)
# calc logp-grad
logp_grad = drift - 0.5 * (diffusion ** 2) * divergence_eval_wrapper(data)
# concat curr grad
return np.concatenate([x_grad, logp_grad], axis=0)
# Run the black-box ODE solver, note the
res = integrate.solve_ivp(ode_func, (eps, 1.0), init_inp, rtol=rtol, atol=atol, method='RK45')
zp = torch.tensor(res.y[:, -1], device=device) # [bs * (pose_dim + 1)]
z = zp[:-shape[0]].reshape(shape) # [bs, pose_dim]
delta_logp = zp[-shape[0]:].reshape(shape[0]) # [bs,] logp
_, sigma_max = marginal_prob_fn(None, torch.tensor(1.).to(device)) # we assume T = 1
prior_logp = global_prior_likelihood(z, sigma_max)
log_likelihoods = (prior_logp + delta_logp) / np.log(2) # negative log-likelihoods (nlls)
return z, log_likelihoods
def cond_pc_sampler(
score_model,
data,
prior,
sde_coeff,
num_steps=500,
snr=0.16,
device='cuda',
eps=1e-5,
pose_mode='quat_wxyz',
init_x=None,
):
pose_dim = PoseUtil.get_pose_dim(pose_mode)
batch_size = data['target_pts_feat'].shape[0]
init_x = prior((batch_size, pose_dim)).to(device) if init_x is None else init_x
time_steps = torch.linspace(1., eps, num_steps, device=device)
step_size = time_steps[0] - time_steps[1]
noise_norm = np.sqrt(pose_dim)
x = init_x
poses = []
with torch.no_grad():
for time_step in time_steps:
batch_time_step = torch.ones(batch_size, device=device).unsqueeze(-1) * time_step
# Corrector step (Langevin MCMC)
data['sampled_pose'] = x
data['t'] = batch_time_step
grad = score_model(data)
grad_norm = torch.norm(grad.reshape(batch_size, -1), dim=-1).mean()
langevin_step_size = 2 * (snr * noise_norm / grad_norm) ** 2
x = x + langevin_step_size * grad + torch.sqrt(2 * langevin_step_size) * torch.randn_like(x)
# normalisation
if pose_mode == 'quat_wxyz' or pose_mode == 'quat_xyzw':
# quat, should be normalised
x[:, :4] /= torch.norm(x[:, :4], dim=-1, keepdim=True)
elif pose_mode == 'euler_xyz':
pass
else:
# rotation(x axis, y axis), should be normalised
x[:, :3] /= torch.norm(x[:, :3], dim=-1, keepdim=True)
x[:, 3:6] /= torch.norm(x[:, 3:6], dim=-1, keepdim=True)
# Predictor step (Euler-Maruyama)
drift, diffusion = sde_coeff(batch_time_step)
drift = drift - diffusion ** 2 * grad # R-SDE
mean_x = x + drift * step_size
x = mean_x + diffusion * torch.sqrt(step_size) * torch.randn_like(x)
# normalisation
x[:, :-3] = PoseUtil.normalize_rotation(x[:, :-3], pose_mode)
poses.append(x.unsqueeze(0))
xs = torch.cat(poses, dim=0)
xs[:, :, -3:] += data['pts_center'].unsqueeze(0).repeat(xs.shape[0], 1, 1)
mean_x[:, -3:] += data['pts_center']
mean_x[:, :-3] = PoseUtil.normalize_rotation(mean_x[:, :-3], pose_mode)
# The last step does not include any noise
return xs.permute(1, 0, 2), mean_x
def cond_ode_sampler(
score_model,
data,
prior,
sde_coeff,
atol=1e-5,
rtol=1e-5,
device='cuda',
eps=1e-5,
T=1.0,
num_steps=None,
pose_mode='quat_wxyz',
denoise=True,
init_x=None,
):
pose_dim = PoseUtil.get_pose_dim(pose_mode)
batch_size = data['target_feat'].shape[0]
init_x = prior((batch_size, pose_dim), T=T).to(device) if init_x is None else init_x + prior((batch_size, pose_dim),
T=T).to(device)
shape = init_x.shape
def score_eval_wrapper(data):
"""A wrapper of the score-based model for use by the ODE solver."""
with torch.no_grad():
score = score_model(data)
return score.cpu().numpy().reshape((-1,))
def ode_func(t, x):
"""The ODE function for use by the ODE solver."""
x = torch.tensor(x.reshape(-1, pose_dim), dtype=torch.float32, device=device)
time_steps = torch.ones(batch_size, device=device).unsqueeze(-1) * t
drift, diffusion = sde_coeff(torch.tensor(t))
drift = drift.cpu().numpy()
diffusion = diffusion.cpu().numpy()
data['sampled_pose'] = x
data['t'] = time_steps
return drift - 0.5 * (diffusion ** 2) * score_eval_wrapper(data)
# Run the black-box ODE solver, note the
t_eval = None
if num_steps is not None:
# num_steps, from T -> eps
t_eval = np.linspace(T, eps, num_steps)
res = integrate.solve_ivp(ode_func, (T, eps), init_x.reshape(-1).cpu().numpy(), rtol=rtol, atol=atol, method='RK45',
t_eval=t_eval)
xs = torch.tensor(res.y, device=device).T.view(-1, batch_size, pose_dim) # [num_steps, bs, pose_dim]
x = torch.tensor(res.y[:, -1], device=device).reshape(shape) # [bs, pose_dim]
# denoise, using the predictor step in P-C sampler
if denoise:
# Reverse diffusion predictor for denoising
vec_eps = torch.ones((x.shape[0], 1), device=x.device) * eps
drift, diffusion = sde_coeff(vec_eps)
data['sampled_pose'] = x.float()
data['t'] = vec_eps
grad = score_model(data)
drift = drift - diffusion ** 2 * grad # R-SDE
mean_x = x + drift * ((1 - eps) / (1000 if num_steps is None else num_steps))
x = mean_x
num_steps = xs.shape[0]
xs = xs.reshape(batch_size * num_steps, -1)
xs = PoseUtil.normalize_rotation(xs, pose_mode)
xs = xs.reshape(num_steps, batch_size, -1)
x = PoseUtil.normalize_rotation(x, pose_mode)
return xs.permute(1, 0, 2), x
def cond_edm_sampler(
decoder_model, data, prior_fn, randn_like=torch.randn_like,
num_steps=18, sigma_min=0.002, sigma_max=80, rho=7,
S_churn=0, S_min=0, S_max=float('inf'), S_noise=1,
pose_mode='quat_wxyz', device='cuda'
):
pose_dim = PoseUtil.get_pose_dim(pose_mode)
batch_size = data['pts'].shape[0]
latents = prior_fn((batch_size, pose_dim)).to(device)
# Time step discretion. note that sigma and t is interchangeable
step_indices = torch.arange(num_steps, dtype=torch.float64, device=latents.device)
t_steps = (sigma_max ** (1 / rho) + step_indices / (num_steps - 1) * (
sigma_min ** (1 / rho) - sigma_max ** (1 / rho))) ** rho
t_steps = torch.cat([torch.as_tensor(t_steps), torch.zeros_like(t_steps[:1])]) # t_N = 0
def decoder_wrapper(decoder, data, x, t):
# save temp
x_, t_ = data['sampled_pose'], data['t']
# init data
data['sampled_pose'], data['t'] = x, t
# denoise
data, denoised = decoder(data)
# recover data
data['sampled_pose'], data['t'] = x_, t_
return denoised.to(torch.float64)
# Main sampling loop.
x_next = latents.to(torch.float64) * t_steps[0]
xs = []
for i, (t_cur, t_next) in enumerate(zip(t_steps[:-1], t_steps[1:])): # 0, ..., N-1
x_cur = x_next
# Increase noise temporarily.
gamma = min(S_churn / num_steps, np.sqrt(2) - 1) if S_min <= t_cur <= S_max else 0
t_hat = torch.as_tensor(t_cur + gamma * t_cur)
x_hat = x_cur + (t_hat ** 2 - t_cur ** 2).sqrt() * S_noise * randn_like(x_cur)
# Euler step.
denoised = decoder_wrapper(decoder_model, data, x_hat, t_hat)
d_cur = (x_hat - denoised) / t_hat
x_next = x_hat + (t_next - t_hat) * d_cur
# Apply 2nd order correction.
if i < num_steps - 1:
denoised = decoder_wrapper(decoder_model, data, x_next, t_next)
d_prime = (x_next - denoised) / t_next
x_next = x_hat + (t_next - t_hat) * (0.5 * d_cur + 0.5 * d_prime)
xs.append(x_next.unsqueeze(0))
xs = torch.stack(xs, dim=0) # [num_steps, bs, pose_dim]
x = xs[-1] # [bs, pose_dim]
# post-processing
xs = xs.reshape(batch_size * num_steps, -1)
xs = PoseUtil.normalize_rotation(xs, pose_mode)
xs = xs.reshape(num_steps, batch_size, -1)
x = PoseUtil.normalize_rotation(x, pose_mode)
return xs.permute(1, 0, 2), x