How to sample from a truncated normal distribution in PyTorch? I could not find any within https://pytorch.org/docs/stable/distributions.html. Is it not implemented? Any workarounds?
This post might be the easiest approach.
In the topic were several methods discussed, so another one might fit your use case better.
I put together a truncated normal distribution class, implementing torch.distributions.Distribution interface: https://github.com/toshas/torch_truncnorm
I wanted to follow up on this as I am looking rsample from a truncated Gaussian in PyTorch and compute log_prob and wanted to see if there were any updated developments. @anton’s module that he shared is very helpful, but unfortunately, I am looking for a solution that is CUDA-capable. Thank you for your help.
It is CUDA-capable now
Yes, made the post before our email exchange. Thanks again for your help!
Is it also differentiable in the parameters?
Why not add this to PyTorch?
I am looking for a way to sample the truncated normal that is differentiable in the parameters of the distribution. That post doesn’t cover automatic differentiation, as far as I can see.
Has anyone got this working in a similar fashion to Tensorflow’s tensorflow.distributions.TruncatedMultivariateNormal?
See here tfp.distributions.MultivariateNormalTriL | TensorFlow Probability.
For my application I specifically need to specify the covariance and location of the mean, so something more detailed than a standard unit normal.
I understand it may be possible to sample from an un-truncated multivariate normal, multiply it by the covariance, then truncate, but I would rather a simple and proper way of doing it. Thanks all!
1 Like