File size: 14,255 Bytes
ca7299e | 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 | """Inspired by https://github.com/jasonkyuyim/se3_diffusion/blob/master/data/so3_diffuser.py"""
import os
import math
import numpy as np
import torch
from src.utils.tensor_utils import inflate_array_like
from src.common import rotation3d
def compose_rotvec(rotvec1, rotvec2):
"""Compose two rotation euler vectors."""
dtype = rotvec1.dtype
R1 = rotation3d.axis_angle_to_matrix(rotvec1)
R2 = rotation3d.axis_angle_to_matrix(rotvec2)
cR = torch.einsum('...ij,...jk->...ik', R1.double(), R2.double())
return rotation3d.matrix_to_axis_angle(cR).type(dtype)
def igso3_expansion(omega, eps, L=1000, use_torch=False):
"""Truncated sum of IGSO(3) distribution.
This function approximates the power series in equation 5 of
"DENOISING DIFFUSION PROBABILISTIC MODELS ON SO(3) FOR ROTATIONAL
ALIGNMENT"
Leach et al. 2022
This expression diverges from the expression in Leach in that here, eps =
sqrt(2) * eps_leach, if eps_leach were the scale parameter of the IGSO(3).
With this reparameterization, IGSO(3) agrees with the Brownian motion on
SO(3) with t=eps^2.
Args:
omega: rotation of Euler vector (i.e. the angle of rotation)
eps: std of IGSO(3).
L: Truncation level
use_torch: set true to use torch tensors, otherwise use numpy arrays.
"""
lib = torch if use_torch else np
ls = lib.arange(L)
if use_torch:
ls = ls.to(omega.device)
if len(omega.shape) == 2:
# Used during predicted score calculation.
ls = ls[None, None] # [1, 1, L]
omega = omega[..., None] # [num_batch, num_res, 1]
eps = eps[..., None]
elif len(omega.shape) == 1:
# Used during cache computation.
ls = ls[None] # [1, L]
omega = omega[..., None] # [num_batch, 1]
else:
raise ValueError("Omega must be 1D or 2D.")
p = (2*ls + 1) * lib.exp(-ls*(ls+1)*eps**2/2) * lib.sin(omega*(ls+1/2)) / lib.sin(omega/2)
if use_torch:
return p.sum(dim=-1)
else:
return p.sum(axis=-1)
def density(expansion, omega, marginal=True, use_torch=False):
"""IGSO(3) density.
Args:
expansion: truncated approximation of the power series in the IGSO(3)
density.
omega: length of an Euler vector (i.e. angle of rotation)
marginal: set true to give marginal density over the angle of rotation,
otherwise include normalization to give density on SO(3) or a
rotation with angle omega.
"""
lib = torch if use_torch else np
if marginal:
# if marginal, density over [0, pi], else over SO(3)
return expansion * (1.0 - lib.cos(omega)) / np.pi
else:
# the constant factor doesn't affect any actual calculations though
return expansion / 8 / np.pi**2
def score(exp, omega, eps, L=1000, use_torch=False): # score of density over SO(3)
"""score uses the quotient rule to compute the scaling factor for the score
of the IGSO(3) density.
This function is used within the Diffuser class to when computing the score
as an element of the tangent space of SO(3).
This uses the quotient rule of calculus, and take the derivative of the
log:
d hi(x)/lo(x) = (lo(x) d hi(x)/dx - hi(x) d lo(x)/dx) / lo(x)^2
and
d log expansion(x) / dx = (d expansion(x)/ dx) / expansion(x)
Args:
exp: truncated expansion of the power series in the IGSO(3) density
omega: length of an Euler vector (i.e. angle of rotation)
eps: scale parameter for IGSO(3) -- as in expansion() this scaling
differ from that in Leach by a factor of sqrt(2).
L: truncation level
use_torch: set true to use torch tensors, otherwise use numpy arrays.
Returns:
The d/d omega log IGSO3(omega; eps)/(1-cos(omega))
"""
lib = torch if use_torch else np
ls = lib.arange(L)
if use_torch:
ls = ls.to(omega.device)
ls = ls[None]
if len(omega.shape) == 2:
ls = ls[None]
elif len(omega.shape) > 2:
raise ValueError("Omega must be 1D or 2D.")
omega = omega[..., None]
eps = eps[..., None]
hi = lib.sin(omega * (ls + 1 / 2))
dhi = (ls + 1 / 2) * lib.cos(omega * (ls + 1 / 2))
lo = lib.sin(omega / 2)
dlo = 1 / 2 * lib.cos(omega / 2)
dSigma = (2 * ls + 1) * lib.exp(-ls * (ls + 1) * eps**2/2) * (lo * dhi - hi * dlo) / lo ** 2
if use_torch:
dSigma = dSigma.sum(dim=-1)
else:
dSigma = dSigma.sum(axis=-1)
return dSigma / (exp + 1e-4)
class SO3Diffuser:
def __init__(
self,
cache_dir: str = './cache',
schedule: str = 'logarithmic',
min_sigma: float = 0.1,
max_sigma: float = 1.5,
num_sigma: int = 1000,
num_omega: int = 1000,
use_cached_score: bool = False,
eps: float = 1e-6,
):
self.schedule = schedule
self.min_sigma = min_sigma
self.max_sigma = max_sigma
self.num_sigma = num_sigma
self.use_cached_score = use_cached_score
self.eps = eps
# Discretize omegas for calculating CDFs. Skip omega=0.
self.discrete_omega = torch.linspace(0, np.pi, steps=num_omega+1)[1:]
# Configure cache directory.
replace_period = lambda x: str(x).replace('.', '_')
cache_dir = os.path.join(
cache_dir,
f'eps_{num_sigma}_omega_{num_omega}_min_sigma_{replace_period(min_sigma)}_max_sigma_{replace_period(max_sigma)}_schedule_{schedule}'
)
if not os.path.isdir(cache_dir):
os.makedirs(cache_dir)
pdf_cache = os.path.join(cache_dir, 'pdf_vals.pt')
cdf_cache = os.path.join(cache_dir, 'cdf_vals.pt')
score_norms_cache = os.path.join(cache_dir, 'score_norms.pt')
if os.path.exists(pdf_cache) \
and os.path.exists(cdf_cache) \
and os.path.exists(score_norms_cache):
self._pdf = torch.load(pdf_cache, map_location='cpu')
self._cdf = torch.load(cdf_cache, map_location='cpu')
self._score_norms = torch.load(score_norms_cache, map_location='cpu')
else:
disc_omega = self.discrete_omega.numpy()
disc_sigma = self.discrete_sigma.numpy()
# Compute the expansion of the power series.
exp_vals = np.asarray(
[igso3_expansion(disc_omega, sigma) for sigma in disc_sigma]
)
# Compute the pdf and cdf values for the marginal distribution of the axis-angles (readily needed for sampling)
self._pdf = np.asarray(
[density(x, disc_omega, marginal=True) for x in exp_vals]
)
self._cdf = np.asarray(
[pdf.cumsum() / num_omega * np.pi for pdf in self._pdf]
)
# Compute the norms of the scores. This are used to scale the rotation axis when
# computing the score as a vector.
self._score_norms = np.asarray(
[score(exp_vals[i], disc_omega, x) for i, x in enumerate(disc_sigma)]
)
self._pdf, self._cdf, self._score_norms = map(
torch.as_tensor, (self._pdf, self._cdf, self._score_norms)
)
# Cache the precomputed values
torch.save(obj=self._pdf, f=pdf_cache)
torch.save(obj=self._cdf, f=cdf_cache)
torch.save(obj=self._score_norms, f=score_norms_cache)
self._score_scaling = torch.sqrt(torch.abs(
torch.sum(self._score_norms**2 * self._pdf, dim=-1) / torch.sum(self._pdf, dim=-1)
)) / np.sqrt(3)
@property
def discrete_sigma(self):
return self.sigma(
torch.linspace(0.0, 1.0, self.num_sigma)
)
def sigma_idx(self, sigma: torch.Tensor):
"""Calculates the index for discretized sigma during IGSO(3) initialization."""
return torch.as_tensor(np.digitize(sigma.cpu().numpy(), self.discrete_sigma) - 1,
dtype=torch.long)
def sigma(self, t: torch.Tensor):
"""Extract \sigma(t) corresponding to chosen sigma schedule."""
if torch.any(t < 0) or torch.any(t > 1):
raise ValueError(f'Invalid t={t}')
if self.schedule == 'logarithmic':
return torch.log(t * math.exp(self.max_sigma) + (1 - t) * math.exp(self.min_sigma))
else:
raise ValueError(f'Unrecognize schedule {self.schedule}')
def diffusion_coef(self, t: torch.Tensor):
"""Compute diffusion coefficient (g_t)."""
if self.schedule == 'logarithmic':
g_t = torch.sqrt(
2 * (math.exp(self.max_sigma) - math.exp(self.min_sigma)) \
* self.sigma(t) / torch.exp(self.sigma(t))
)
else:
raise ValueError(f'Unrecognize schedule {self.schedule}')
return g_t
def t_to_idx(self, t: torch.Tensor):
"""Helper function to go from time t to corresponding sigma_idx."""
return self.sigma_idx(self.sigma(t))
def sample_prior(self, shape: torch.Size, device=None):
t = torch.ones(shape[0], dtype=torch.float, device=device)
return self.sample(t, shape)
def sample(self, t: torch.Tensor, shape: torch.Size):
"""Generates rotation vector(s) from IGSO(3).
Args:
t: continuous time in [0, 1].
shape: shape of the output tensor.
Returns:
(shape, ) axis-angle rotation vectors sampled from IGSO(3).
"""
assert t.ndim == 1 and t.shape[0] == shape[0], \
f"t.shape={t.shape}, shape={shape}"
assert shape[-1] == 3, f"The last dim should be 3, but got shape={shape}"
z = torch.randn(shape, device=t.device) # axis-angle
x = z / torch.linalg.norm(z, dim=-1, keepdims=True)
# sample igso3
z_igso3 = torch.rand(shape[:-1]) # (num_batch, num_res)
igso3_scaling = []
for i, _t in enumerate(t):
t_idx = self.t_to_idx(_t).item() # [1]
# np.interp can accept Tensor as input
igso3_scaling.append(
np.interp(z_igso3[i], self._cdf[t_idx], self.discrete_omega),
) # (num_res, )
igso3_scaling = torch.as_tensor(np.asarray(igso3_scaling), dtype=x.dtype, device=t.device)
return x * igso3_scaling[..., None]
def score(
self,
vec: torch.tensor,
t: torch.tensor,
):
"""Computes the score of IGSO(3) density as a rotation vector.
Same as score function but uses pytorch and performs a look-up.
Args:
vec: [..., 3] array of axis-angle rotation vectors.
t: continuous time in [0, 1].
Returns:
[..., 3] score vector in the direction of the sampled vector with
magnitude given by _score_norms.
"""
assert t.ndim == 1 and t.shape[0] == vec.shape[0], \
f"t.shape={t.shape}, vec.shape={vec.shape}"
device = vec.device
omega = torch.linalg.norm(vec, dim=-1) + self.eps # [num_batch, num_res]
if self.use_cached_score:
score_norms_t = self._score_norms[self.t_to_idx(t)]
score_norms_t = torch.as_tensor(score_norms_t).to(device)
omega_idx = torch.bucketize(
omega, torch.as_tensor(self.discrete_omega[:-1]).to(device))
omega_scores_t = torch.gather(score_norms_t, 1, omega_idx)
else: # compute on the fly
sigma = self.discrete_sigma[self.t_to_idx(t)]
sigma = torch.as_tensor(sigma).to(device)
omega_vals = igso3_expansion(omega, sigma[:, None], use_torch=True)
omega_scores_t = score(omega_vals, omega, sigma[:, None], use_torch=True)
return omega_scores_t[..., None] * vec / (omega[..., None] + self.eps)
def score_scaling(self, t: torch.Tensor):
"""Calculates scaling used for scores during trianing."""
return torch.as_tensor(self._score_scaling[self.t_to_idx(t)], device=t.device)
def forward_marginal(self, rot_0: torch.Tensor, t: torch.Tensor):
"""Samples from the forward diffusion process at time index t.
Args:
rot_0: [..., 3] initial rotations.
t: continuous time in [0, 1].
Returns:
rot_t: [..., 3] noised rotation vectors.
rot_score: [..., 3] score of rot_t as a rotation vector.
"""
rotvec_0t = self.sample(t, shape=rot_0.shape) # "delta_rot"
rot_score = self.score(rotvec_0t, t)
# Right multiply => vector_add in Euclidean space
rot_t = compose_rotvec(rot_0, rotvec_0t)
return rot_t, rot_score
def reverse(
self,
rot_t: torch.Tensor,
score_t: torch.Tensor,
t: torch.Tensor,
dt: float,
mask: torch.Tensor = None,
noise_scale: float = 1.0,
probability_flow: bool = True,
):
"""Simulates the reverse SDE for 1 step using the Geodesic random walk.
Args:
rot_t: [..., 3] current rotations at time t.
score_t: [..., 3] rotation score at time t.
t: continuous time in [0, 1].
dt: continuous step size in [0, 1].
add_noise: set False to set diffusion coefficent to 0.
mask: True indicates which residues to diffuse.
probability_flow: set True to use probability flow ODE.
Returns:
[..., 3] rotation vector at next step.
"""
t = inflate_array_like(t, rot_t)
g_t = self.diffusion_coef(t)
z = noise_scale * torch.randn_like(score_t)
rev_drift = -1.0 * (g_t ** 2) * score_t * dt * (0.5 if probability_flow else 1.)
rev_diffusion = 0. if probability_flow else (g_t * np.sqrt(dt) * z)
perturb = rev_drift + rev_diffusion
if mask is not None:
perturb *= mask[..., None]
# Right multiply.
rot_t_1 = compose_rotvec(rot_t, -1.0 * perturb) # reverse in time
return rot_t_1 |