Installing GRIN

Python, R, and the two things that go wrong.

You may not need to install anything. The Analyse page runs the same network in your browser: paste or upload a matrix and it returns the same numbers the packages do, with nothing uploaded anywhere. Install the packages when you want to script an analysis, run a sample, or drive an adaptive design.

Both packages are open source under an MIT licence and wrap identical trained weights, verified to agree to floating-point tolerance.

Where things stand

grintools (Python) and grin (R) are not yet on PyPI and CRAN. Until they are, install from the repository — the instructions below are the supported route, not a workaround, and they install exactly what the paper describes.

  1. Python

    Needs Python 3.9 or newer. There is no PyTorch dependency: the network ships as an ONNX graph and runs under onnxruntime, so the install is small and fast. 

    git clone https://github.com/MurraySBennett/grin.git
pip install ./grin/packages/grintools

# check it worked
python -c "import grintools as gt; \
print(gt.infer([[83,112,47,11],[38,154,28,33],[15,27,117,94],[6,36,75,136]])[1]['p_PI'])"

    That prints a probability of perceptual independence for the example matrix. If you see a number, you are done.

  2. R

    The R package runs the same weights as a TorchScript graph through torch, so no Python installation is needed on the R side — but torch itself has a second, separate download. 

    install.packages(c("torch", "remotes", "jsonlite"))

# libtorch is a SEPARATE ~2 GB download, and this is the slow step
torch::install_torch()
torch::torch_tensor(1)        # confirm it actually works

remotes::install_github("MurraySBennett/grin", subdir = "packages/grin")

# check it worked
library(grin)
cm <- matrix(c(83,112,47,11, 38,154,28,33,
               15,27,117,94, 6,36,75,136), nrow = 4, byrow = TRUE)
grin_infer(cm)$constructs$p_PI

    jsonlite is optional. Without it everything works except the calibrated-interval flag, which falls back to the network's own intervals with a warning rather than failing.

  3. Two things that go wrong

    torch::install_torch() can report an error and still have succeeded. Run non-interactively it sometimes ends with cannot change value of locked binding for '.torch_can_load' after the download has already completed. Do not trust the exit status — check with torch::torch_tensor(1) instead. If that returns a tensor, you are fine.

    On Windows with WSL, ~ does not cross the boundary. An R session running on Windows expands ~ to your Windows profile, so a path like ~/projects/grin will not find a clone that lives inside WSL. Use the UNC path, and prefer a built tarball over a directory, since install_local() copies recursively across the mount and tends to fail: 

    # built in WSL with:  R CMD build --no-build-vignettes packages/grin
install.packages(
  "//wsl.localhost/Ubuntu/home/USER/projects/grin/dist/grin_0.1.0.tar.gz",
  repos = NULL, type = "source"
)

  4. Check the optional interval correction

    Both packages ship an off-by-default correction for the interval widths (see Evidence). This verifies it is present and behaving: 

    # from a clone of the repository
Rscript scripts/R/verify_recalibration.R
# -> twelve checks, all PASS

    If cal.calibrated comes back False, the correction file did not ship with your build; the package still works, and you get the network's own intervals.

Next

Analyse data you already have, or see the other tutorials.