What is currently the best/correct way to train on stages? I mean, for example, that at first I train model with one loss, and after some time change it to different one. Or change augmentations. There is an issue about this, but it is stale: Transfer learning phases · Issue #2006 · Lightning-AI/li…

I think the best option as of now is 2. Model.load_from_checkpoint() accepts additional args that override the ones loaded from the checkpoint. In this way you could do something like: model = Model(stage1) trainer.fit(model) # stage1 model.stage = 2 # OR model = Model.load_from_checkpoint("stage1…

What is the best way to train on stages?

implementation help

goku August 30, 2020, 8:57pm 6

wouldn’t trainer call datamodule.setup('fit') again in such a case??