Retain LR in successive training when using schedulers

AutoResearch / EEG-GAN

Other

19 stars 1 forks source link

Retain LR in successive training when using schedulers #59

Closed chadcwilliams closed 3 months ago

chadcwilliams commented 9 months ago

We implemented a LR scheduler, but the GAN does not keep track of the changed LR so when re-training a GAN, the LR resets. We should track the LR and then force it back at re-training

whyhardt commented 3 months ago

Did you ever use the scheduler in the latest training/analysis? If not I would recommend to delete them anyways because we cannot give any usage recommendations

chadcwilliams commented 3 months ago

No I didn't. We implemented it when the AE-GAN wasn't working and we figured it was having trouble learning, but it turned out to be a bug to do with an activation function. Once the bug was fixed, the GAN worked quite well so we never needed to use the scheduler. So, I agree, let's remove it as part of #16