Skip to content

A numerically-stable and differentiable implementation of the Truncated Gaussian distribution in Pytorch.

License

Notifications You must be signed in to change notification settings

TheAeryan/stable-truncated-gaussian

Repository files navigation

PyPI version Downloads

Stable Truncated Gaussian

A differentiable implementation of the Truncated Gaussian (Normal) distribution using Python and Pytorch, which is numerically stable even when the μ parameter lies outside the interval [a,b] given by the bounds of the distribution. In this situation, a naive evaluation of the mean, variance and log-probability of the distribution could otherwise result in catastrophic cancellation. Our code is inspired by TruncatedNormal.jl and torch_truncnorm. Currently, we provide numerically-stable methods for calculating the mean, variance, log-probability, KL-divergence and sampling from the distribution. Our current implementation of icdf (which is used for sampling from the distribution) still needs some work for those situations where the [a,b] interval is small. For a comparison between our icdf implementation and the one provided by scipy, take a look at the images_cdf_comparison folder.

Installation

Simply install with pip:

pip install stable-trunc-gaussian

Example

Run the following code in Python:

from stable_trunc_gaussian import TruncatedGaussian as TG
from torch import tensor as t

# Create a Truncated Gaussian with mu=0, sigma=1, a=10, b=11
# Notice how mu is outside the interval [a,b]
dist = TG(t(0),t(1),t(10),t(11))

print("Mean:", dist.mean)
print("Variance:", dist.variance)
print("Log-prob(10.5):", dist.log_prob(t(10.5)))

Result:

Mean: tensor(10.0981)
Variance: tensor(0.0094)
Log-prob(10.5): tensor(-2.8126)

Parallel vs Sequential Implementation

The class obtained by doing from stable_trunc_gaussian import TruncatedGaussian corresponds to a parallel implementation of the truncated gaussian, which makes possible to obtain several values (mean, variance and log-probs) in parallel. In case you are only interested in computing values sequentially, i.e., one at a time, we also provide a sequential implementation which results more efficient only for this case. In order to use this sequential implementation, simply do from stable_trunc_gaussian import SeqTruncatedGaussian. Here is an example:

from stable_trunc_gaussian import TruncatedGaussian, SeqTruncatedGaussian
from torch import tensor as t

# Parallel computation
means = TruncatedGaussian(t([0,0.5]),t(1,1),t(-1,2),t(1,5)).mean

# Sequential computation
# Note: the 'TruncatedGaussian' class can also be used for this sequential case
mean_0 = SeqTruncatedGaussian(t([0]),t(1),t(-1),t(1)).mean
mean_1 = SeqTruncatedGaussian(t([0.5]),t(1),t(2),t(5)).mean

Acknowledgements

We want to thank users KFrank and ptrblck for their help in solving the bug when computing the gradients for the parallel version (bug solved in version 1.1.1).

About

A numerically-stable and differentiable implementation of the Truncated Gaussian distribution in Pytorch.

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages