I have successfully completed the code submission several times, I do see an error in the log though:
Could not load symbol cublasGetSmCountTarget from libcublas.so.11. Error: /opt/conda/lib/python3.10/site-packages/torch/lib/…/…/…/…/libcublas.so.11: undefined symbol: cublasGetSmCountTarget
I’m not sure what it is (something to do with torch cuda setup on your end maybe), it does not seem to affect my results although I can’t actually see my predictions, is there some way for you to make sure I’m actually producing predictions.
Also, when I receive an email with the successful completion. the link to review the submission results and log does not work - https://drivendata.org/competitions/259/submissions/code/
Thanks for pointing out the torch cuda issue—I will look into that. Also thanks for noting the issue with the link in the email.
For determining whether your predictions are being produced successfully, you are permitted to print out your predictions in the logs. Unlike a normal DrivenData competition, the test data in this competition is not sensitive.
If your submission is being marked as successful, it means that we are at least receiving predictions that match the
submission_format.csv—i.e., you’re producing the right number of prediction values that are floats.
Following up on the CUDA error: my best understanding is that fixing the error may involve having to change PyTorch versions, which may involve some risk in breaking yours or other participants’ submissions. From what I can tell, you are still generating predictions successfully, though your code might not be using the GPU. However, your inference seems to run quite fast, so I don’t know that there’s a big benefit to you for addressing it.
Yes, the error does not seem to affect my code, it runs fine and is using the GPU, so no problems. Thanks!