Skip to content

Add jitter to Cholesky factorization in Gaussian ops #3151

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We鈥檒l occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
fehiepsi merged 11 commits into from
Oct 30, 2022
Merged

Add jitter to Cholesky factorization in Gaussian ops #3151

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
be0e68d
Attempt to add jitter to Gaussian ops
fritzo Oct 27, 2022
102abf5
Isolate changes to pyro.ops.tensor_utils
fritzo Oct 28, 2022
ae184e8
Add a warning
fritzo Oct 28, 2022
fd40b54
Simplify
fritzo Oct 28, 2022
d78b83c
Simplify
fritzo Oct 28, 2022
d542b83
Simplify to fixed jitter based on finfo.eps
fritzo Oct 28, 2022
2f5e184
Simplify
fritzo Oct 28, 2022
fdb1be1
Rename cholesky -> safe_cholesky
fritzo Oct 28, 2022
2da3c75
Allow disabling jitter
fritzo Oct 28, 2022
152b913
Address review comment
fritzo Oct 30, 2022
3e273e7
Switch to column-wise max, increase jitter
fritzo Oct 30, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 5 additions & 5 deletions pyro/distributions/hmm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,7 +20,7 @@
)
from pyro.ops.indexing import Vindex
from pyro.ops.special import safe_log
from pyro.ops.tensor_utils import cholesky, cholesky_solve
from pyro.ops.tensor_utils import cholesky_solve, safe_cholesky

from . import constraints
from .torch import Categorical, Gamma, Independent, MultivariateNormal
Expand Down Expand Up @@ -628,9 +628,9 @@ def filter(self, value):

# Convert to a distribution
precision = logp.precision
loc = cholesky_solve(logp.info_vec.unsqueeze(-1), cholesky(precision)).squeeze(
-1
)
loc = cholesky_solve(
logp.info_vec.unsqueeze(-1), safe_cholesky(precision)
).squeeze(-1)
return MultivariateNormal(
loc, precision_matrix=precision, validate_args=self._validate_args
)
Expand Down Expand Up @@ -928,7 +928,7 @@ def filter(self, value):
gamma_dist.concentration, gamma_dist.rate, validate_args=self._validate_args
)
# Conditional of last state on unit scale
scale_tril = cholesky(logp.precision)
scale_tril = safe_cholesky(logp.precision)
loc = cholesky_solve(logp.info_vec.unsqueeze(-1), scale_tril).squeeze(-1)
mvn = MultivariateNormal(
loc, scale_tril=scale_tril, validate_args=self._validate_args
Expand Down
4 changes: 2 additions & 2 deletions pyro/distributions/transforms/cholesky.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -89,7 +89,7 @@ def log_abs_det_jacobian(self, x, y):

class CholeskyTransform(Transform):
r"""
Transform via the mapping :math:`y = cholesky(x)`, where `x` is a
Transform via the mapping :math:`y = safe_cholesky(x)`, where `x` is a
positive definite matrix.
"""
bijective = True
Expand All @@ -116,7 +116,7 @@ def log_abs_det_jacobian(self, x, y):

class CorrMatrixCholeskyTransform(CholeskyTransform):
r"""
Transform via the mapping :math:`y = cholesky(x)`, where `x` is a
Transform via the mapping :math:`y = safe_cholesky(x)`, where `x` is a
correlation matrix.
"""
bijective = True
Expand Down
10 changes: 5 additions & 5 deletions pyro/ops/gaussian.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,7 @@
from torch.nn.functional import pad

from pyro.distributions.util import broadcast_shape
from pyro.ops.tensor_utils import cholesky, matmul, matvecmul, triangular_solve
from pyro.ops.tensor_utils import matmul, matvecmul, safe_cholesky, triangular_solve


class Gaussian:
Expand Down Expand Up @@ -154,7 +154,7 @@ def rsample(
"""
Reparameterized sampler.
"""
P_chol = cholesky(self.precision)
P_chol = safe_cholesky(self.precision)
loc = self.info_vec.unsqueeze(-1).cholesky_solve(P_chol).squeeze(-1)
shape = sample_shape + self.batch_shape + (self.dim(), 1)
if noise is None:
Expand Down Expand Up @@ -254,7 +254,7 @@ def marginalize(self, left=0, right=0) -> "Gaussian":
P_aa = self.precision[..., a, a]
P_ba = self.precision[..., b, a]
P_bb = self.precision[..., b, b]
P_b = cholesky(P_bb)
P_b = safe_cholesky(P_bb)
P_a = triangular_solve(P_ba, P_b, upper=False)
P_at = P_a.transpose(-1, -2)
precision = P_aa - matmul(P_at, P_a)
Expand All @@ -277,7 +277,7 @@ def event_logsumexp(self) -> torch.Tensor:
Integrates out all latent state (i.e. operating on event dimensions).
"""
n = self.dim()
chol_P = cholesky(self.precision)
chol_P = safe_cholesky(self.precision)
chol_P_u = triangular_solve(
self.info_vec.unsqueeze(-1), chol_P, upper=False
).squeeze(-1)
Expand Down Expand Up @@ -550,7 +550,7 @@ def gaussian_tensordot(x: Gaussian, y: Gaussian, dims: int = 0) -> Gaussian:
b = xb + yb

# Pbb + Qbb needs to be positive definite, so that we can malginalize out `b` (to have a finite integral)
L = cholesky(Pbb + Qbb)
L = safe_cholesky(Pbb + Qbb)
LinvB = triangular_solve(B, L, upper=False)
LinvBt = LinvB.transpose(-2, -1)
Linvb = triangular_solve(b.unsqueeze(-1), L, upper=False)
Expand Down
5 changes: 3 additions & 2 deletions pyro/ops/tensor_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,6 +7,7 @@
from torch.fft import irfft, rfft

_ROOT_TWO_INVERSE = 1.0 / math.sqrt(2.0)
CHOLESKY_JITTER = 1.0


def as_complex(x):
Expand Down Expand Up @@ -393,15 +394,15 @@ def inverse_haar_transform(x):
return x


def cholesky(x):
def safe_cholesky(x):
if x.size(-1) == 1:
x = x.clamp(min=torch.finfo(x.dtype).tiny)
return x.sqrt()

# Add adaptive jitter.
x = x.clone()

@martinjankowiak martinjankowiak Oct 28, 2022

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

are you sure you need/want clones? fwiw i do this in millipede, which is similar to what's done in gpytorch

@fritzo fritzo Oct 28, 2022 •
edited
Loading

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yup this .clone() is needed because we're mutating the matrix. Nice, I'll rename this to safe_cholesky() as in millipede.

Thanks, I did try various gpytorch-style lazy tactics that avoid adding jitter until a failure has occurred. I found that since the Cholesky error happens only late in the filtering process, by that point the filter state had already been corrupted by nearly-singular matrices that just barely didn't trigger an error. The best solution I've found so far is to add a tiny amount of noise to all matrices so that error doesn't build up during filtering. Other solutions include using svd or pinv or ldl_factor, but they were more expensive.

Note the core piece of linear algebra is in Gaussian.marginalize() which is repeatedly called in the filter pass of sequential_gaussian_filter_sample(). It's just the blockwise symmetric matrix inverse formula:

# in Gaussian.marginalize():
P_aa = self.precision[..., a, a]
P_ba = self.precision[..., b, a]
P_bb = self.precision[..., b, b]
P_b = safe_cholesky(P_bb)                       # Note if we add a little jitter here...
P_a = triangular_solve(P_ba, P_b, upper=False)  # ...then this is smaller...
P_at = P_a.transpose(-1, -2)
precision = P_aa - matmul(P_at, P_a)            # ...so this is even better conditioned.

This code has the nice property that if we add a little bit of jitter before Cholesky factorizing, the next precision matrix becomes only better-conditioned. Empirically this allowed me to get away with much smaller jitter than was needed if I waited for an error to occur.

BTW it looks like you could speed up millipede by switching from try: c = cholesky() to the faster c, info = cholesky_ex(); if not info.any(): return c, which is used in gpytorch. The only reason I'm not using cholesky_ex() here is that I found the decision-based version was too unstable.

x_max = x.data.reshape(*x.shape[:-2], -1).abs().max(-1, True).values
jitter = x_max * torch.finfo(x.dtype).eps
jitter = CHOLESKY_JITTER * torch.finfo(x.dtype).eps * x_max
Comment thread
fritzo marked this conversation as resolved.
Outdated
x.data.diagonal(dim1=-1, dim2=-2).add_(jitter)

return torch.linalg.cholesky(x)
Expand Down