failed to run

rhsimplex

1mo ago

Hi, my submission which worked fine with the --full flag locally, failed on submission with Failed to run. Is this the same as CANT_RUN in the competition description, and if so or otherwise -- can I get any trace to figure out what's happening?

I'm trying add torch as a dependency, perhaps this won't work?

# /// script
# dependencies = ["torch"]
# ///

edit: I tried another version without the torch import. This stayed in "Processing..." for a very long time but ended up in "failed to run" after an hour or so.

edit2: also tried a lgbm based model, also same result...

Thanks

4 Replies

rhsimplex1mo ago

Hey I got not response here but for anyone else, I guess the torch import won't work. That is a shame, because even though there are no GPUs some small CPU models could be promising. I tried rewriting a model in numpy, but my non-optimized backprop is way too slow.

While I can't be sure, I think I've figured out that the "Failed to run" error is coming from the model taking to long to train. This doesn't come up when running the runner.py script, but I guess that during the overfit tests there must be some datasets that are bigger and trigger cancellation.

Would be great if that could be included in the the error text. Also would be great if we could use torch cpu!!

Nonius1mo ago

Hi rhsimplex

Apologies will take a look

Best

Marc

Nonius1mo ago

Hi rhsimplex,

Your guess was correct, we have a timeout on the overfitting test of 90 minutes, which I admit is tighter than the training limit we give. The thing is, right now we aren't parallelising submissions processing (that will eventually be done some time in the future), so one submission that takes something like 7 hours to perform overfitting test could be a bottleneck if a bunch of submissions are in the queue.

on torch, our original intent was to disallow NNs and also because we aren't running gpu on the backend. I'll look into seeing if we can include torch.

Feel free to contact me by email to discuss how you can adapt your model to speed it up.

Best

Marc

rhsimplex1mo ago

Thanks for the explanations!