Frequently Asked Questions

My GPU is on fire — should I be worried?

Only if it was not already on fire before you started training. In all other cases, reduce your batch size, check your cooling, and consider whether torch.backends.cudnn.benchmark = True is appropriate for your hardware. If the fire persists, please prioritise contacting your local fire department over opening a GitHub issue.

My loss is converging but my collaborator insists the results are biased. Who is right?

Both. A converging loss confirms the network is learning something — it does not confirm the network is learning the right thing. Sage addresses 11 interconnected supervised-learning biases documented in the paper; if you believe you have found a 12th, we genuinely welcome a GitHub issue.

My partner says I spend more time with Sage than with them. Should I be worried?

Define “worried.” If your validation loss is still decreasing, you are making measurable progress on at least one front. If you need to find someone who is a better match for you, we recommend the matched-filter pipelines for theoretical guarantees in the mismatch. There are plenty of templates in the bank. Your partner is not the one for you.

---

Note

More questions coming soon. If you have a question that keeps coming up, open a GitHub discussion and it may end up here.