Crash when loading torch in RStudio

[caewok]

Attempting to load torch (pytorch.org) using the reticulate package in RStudio Server results in a hard crash (probably a segfault). The same does not happen using R from the terminal. (BTW, loading other python modules works fine in RStudio Server.)

I am running Ubuntu 14.04.5 LTS (GNU/Linux 3.13.0-137-generic x86_64), with RStudio Server 1.1.383 and Microsoft R Open 3.4.2. I am using reticulate 1.7.

sessionInfo():

R version 3.4.2 (2017-09-28)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Ubuntu 14.04.5 LTS

Matrix products: default
BLAS: /opt/microsoft/ropen/3.4.2/lib64/R/lib/libRblas.so
LAPACK: /opt/microsoft/ropen/3.4.2/lib64/R/lib/libRlapack.so

locale:
[1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C LC_TIME=en_US.UTF-8 LC_COLLATE=en_US.UTF-8
[5] LC_MONETARY=en_US.UTF-8 LC_MESSAGES=en_US.UTF-8 LC_PAPER=en_US.UTF-8 LC_NAME=C
[9] LC_ADDRESS=C LC_TELEPHONE=C LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C

attached base packages:
[1] stats graphics grDevices utils datasets methods base

other attached packages:
[1] reticulate_1.7 RevoUtils_10.0.6 RevoUtilsMath_10.0.1

loaded via a namespace (and not attached):
[1] compiler_3.4.2 Matrix_1.2-11 tools_3.4.2 yaml_2.1.13 Rcpp_0.12.12 grid_3.4.2
[7] jsonlite_1.5 lattice_0.20-35

py_config() returns:

python: /home/caewok/anaconda2/bin/python
libpython: /home/caewok/anaconda2/lib/libpython2.7.so
pythonhome: /home/caewok/anaconda2:/home/caewok/anaconda2
version: 2.7.14 |Anaconda, Inc.| (default, Dec 7 2017, 17:05:42) [GCC 7.2.0]
numpy: /home/caewok/anaconda2/lib/python2.7/site-packages/numpy
numpy_version: 1.14.0

python versions found:
/home/caewok/.virtualenvs/r-reticulate/bin/python
/usr/bin/python
/usr/bin/python3
/home/caewok/.virtualenvs/r-test-gan-v/bin/python
/home/caewok/anaconda2/bin/python
/home/caewok/miniconda2/bin/python

Running Microsoft R Open 3.4.2 in the terminal with the following works, but running the same in RStudio Server crashes ("The previous R session was abnormally terminated due to an unexpected crash.") :

library(reticulate)
torch <- import("torch")  

Same for this:

library(reticulate)
conda_create("r-test-conda")
use_condaenv("r-test-conda")
torch <- import("torch")

And same for this:

library(reticulate)
virtualenv_create("r-test-v")
use_virtualenv("r-test-v")
torch <- import("torch")

Suggestions are most appreciated.

[matanox]

I think I'd switch to a pure python environment (jupyter notebook) for PyTorch, rather than try running it from within R. You can still pass around dataframes from R using this I think.

[matanox]

By the way torch and PyTorch are different things

[harryprince]

reticulate always crash on my ubuntu, say goodbye for you...
and the bug is fixed at a very slow speed.

[PM-42]

I have a related issue with torch_tensor() causing R to crash. Can anybody help, please. Thanks

System Information:

• OS: Win 11
• R version: 4.3.1
• torch package version: 0.14.1

Problem:
I am encountering an issue where calling torch_tensor() causes R to crash. Here is a minimal example that reproduces the issue:

# Install and load torch library
install.packages("torch")
library(torch)

# Ensure Lantern is loaded
torch::torch_manual_seed(42)

# Create a minimal tensor
sample_data <- matrix(1:10, nrow = 2, ncol = 5)  # Simple data
sample_data_tensor <- torch_tensor(sample_data, dtype = torch_long())    #!!!!!!! CRASH here

[t-kalinowski]

The crash from PyTorch in reticulate arises if you also load data.table in the same R session, or any other package that wants to load libomp, before the first import("torch") call.

The easiest solution is to call import('torch') before calling library(data.table).

See also #1732.

That said, @PM-42, it looks like you're using the R torch package, which is different from the Python torch package. The R torch package does not use reticulate. Can you please open an issue on mlverse/torch?

[PM-42]

Thanks. I reposted on mlverse/torch.
I am using R torch (not python torch), and no other packages were loaded