OOM problem when I search hyperparameters with Tabnet about tabnet

[Optimox]

Hi, without a code example it is difficult to know what is going on. The code is not optimized to be run multiple times on parallel. I think the best way to speed up training is to play with batch size, num workers etc so that you have a good gpu utilization and then simply do your hyperparameter search sequentially.

[luyuhengCN]

Hi, thanks for your reply. I monitor the memory usage of the .fit() processing, and I found the create_dataloaders() in utils.py change the X_train into np.float32. I guess this will take up more memory if the X_train is big. I'm not sure if it would cause the OOM problem.

[Optimox]

Yes usually models are trained using float32, we could try to use mixed precision (float16) but that would still not be meant for multiple trainings in parallel.