How to use pytorch in this challenge?

ARC White-Box Estimation Challenge 2026
#1

I’ve tried to add torch to the requirements.txt but my estimator still fails with: “ModuleNotFoundError: No module named ‘torch’”.

#2

I’ve also tried to add numpy only and get the same issue for numpy. So I guess I’m doing something terribly wrong. For the numpy only solution my requirements.txt looks like this:
numpy==2.4.4

also tried without pinning the version but got the same issue.

#3

Hi @ngcferreira,

torch (and numpy/scipy) can’t be imported in the grader, and requirements.txt isn’t installed there. This is intentional: estimators run in a sandbox exposing only flopscope (flopscope.numpy as fnp), the whestbench API (BaseEstimator / MLP / SetupContext), and the Python standard library, so all array math is FLOP-counted through flopscope — which is also why pinning numpy made no difference.

Use PyTorch as an offline tool instead: precompute locally, save a pickle-free .npz, ship it in a folder submission (whest submit --estimator .), and load it in setup() with fnp.load(str(path)) (0 FLOPs).

Both are documented in the starter kit, with a worked example:

(We’ve also corrected the docs and the CLI hint that previously implied requirements.txt installed dependencies - thanks for flagging it.)

Best,
Mohanty

1 Like
#4

Thank you very much @mohanty for your answer. All is clear now.

#5

@mohanty I’m still not able to get my code to get scored and I’m not really sure why, as everything runs well locally and the submission doesn’t really return any error message other than “Error : Evaluation could not complete; please retry”: AIcrowd | ARC White-Box Estimation Challenge 2026 | Submissions #312575
And in another submission it just says “Error : Evaluation error”
AIcrowd | ARC White-Box Estimation Challenge 2026 | Submissions #312625

Is there any support email I can reach out to to check why this has failed?

#6

@ngcferreira The initial issue you ran into was caused by how whestbench was handling sub-packages in submitted solutions.

We have since made the evaluators more robust to this class of issue, and also updated whestbench to fix the packaging problem at the source (v0.12.0rc3).

The two submissions you mentioned did still fail on re-evaluation, but for an unrelated and expected reason. In any case, it looks like you have since been unblocked and have already managed to get a few successful submissions through.

Best of luck,
Mohanty