torch_manual_seed does not seem to work with $sample()

[LHBO]

Hi,

This is my first time writing on GitHub, so I am unsure if I have included everything you need.

I have a question regarding reproducibility when generating random numbers from distributions.
From pages 36 and 37 in the manual (https://cran.r-project.org/web/packages/torch/torch.pdf) it seems that I should use the $sample() function, but then I am not able to set the seed and reproduce the sampled values even when setting the seed.
However, setting the seed works when I use $rsample(). Is there a particular reason for this?

P.S. I think you have done a splendid job in making torch much more available for us R users!

library(torch)
dn = distr_normal(0, 1)

# These calls do NOT generate the same value
torch_manual_seed(10)
dn$sample(1)
torch_manual_seed(10)
dn$sample(1)

# These calls do NOT generate the same value (checking if I needed to set seed both in R and torch)
torch_manual_seed(10)
set.seed(10)
dn$sample(1)
torch_manual_seed(10)
set.seed(10)
dn$sample(1)

# These calls do generate the same value
torch_manual_seed(10)
dn$rsample(1)
torch_manual_seed(10)
dn$rsample(1)

[LHBO]

I should maybe add that I use torch version 0.9.0 (newest), R version 4.2.0 (2022-04-22), and I run on a mac with operating system x86_64-apple-darwin17.0.

Manually setting the seed works for e.g. distr_categorical.

library(torch)
dc = distr_categorical(torch_tensor(c(0.1, 0.5, 0.4)))

# Produce the same values
torch_manual_seed(1)
dc$sample(10)
torch_manual_seed(1)
dc$sample(10)
´´´

[dfalbel]

Hi @LHBO , thanks very much for reporting!

This is indeed a bug. It will be fixed once #938 is merged.
It will probably take some time to go into the CRAN version though.
Once the PR is merged you'll be able to install the dev version with

remotes::install_github("mlverse/torch")