I’m exploring hyper-parameters of a model that runs fairly quickly over a large number of epochs. I have a script set up to delete some of the per-epoch serialization files after the model finishes, but if possible, it would be nice to save the disk I/O and disable or minimize this per-epoch serialization. Is there any way to do this?
There are some parameters to minimize the number saved models (see the options to
Trainer), so you shouldn’t need to have your own script do this, but these parameters always write the model at every epoch and only deal with which ones to delete and when to delete them. If you want to not save them in the first place, you probably need to use the
CallbackTrainer and modify the behavior of the
Checkpoint callback. I agree that it seems like you should be able to just decline saving the model every epoch, to save on disk I/O, but I don’t think there’s any way to do that right now.