How to Deactivate Early Stopping?

[ahmadtulsi]

Hi,

First of all, thank you for the great paper and the repository—I've learned a lot about MLPs from it!

I have a question regarding early stopping. I would like to deactivate it, but I’m not sure how. I noticed that early stopping is mentioned in the RealMLP module, but what about the other models?

I tried setting val_fraction = 0, but training still stopped early. Is there a recommended way to fully disable early stopping across different models in this repo?

Thanks in advance!

[dholzmueller]

Hi, unfortunately there is no general way to disable early stopping across models at the moment. For some models it might be easily possible, I would need to check (RealMLP has the most support for such things, then probably boosted trees. The other models don't have as many options currently.)

[ahmadtulsi]

Thanks for the reply!

After digging into the code (including parts from RTDL), I found a workaround by setting:

es_patience > max_epochs

For the boosting model, I used the following:

clf = XGB_TD_Classifier()
clf._get_default_params = lambda: {**DefaultParams.XGB_TD_CLASS, "early_stopping_rounds": 10000}

Just wanted to share this as an FYI in case it's helpful for others. Thanks again!

[dholzmueller]

Good point. This will deactivate early stopping, but not best-epoch selection (the parameters will still be reverted to the epoch with the lowest validation score). Maybe you can deactivate that if you use an empty validation set, I'd have to check...