Closed vmarois closed 5 years ago
Wait... what?
First of all, you mentioned fine-tuning, so we are talking about GridTrainer, not GridTester, right? ;)
Then, you are saying that sometimes the best_model.pt is not saved at all? How is that possible... as far as I know during fine-tuning (i.e. "training with the pre-trained model loaded") we do not take into account the accuracy stored in the loaded checkpoint... so it should save at least the best_model.pt at the 0 step, right?
And if that is the best model, then it is - I see no point of using a different one... Which one? The last one? Why?
Closing this for now. Not really urgent anyway.
Currently playing with the
GridTester
to handle the ViGIL experiments. When finetuning, I am encountering a situation where the model does not get saved at each (or frequently enough) validation step, as the validation loss is not decreasing. It'd be nice if theGridTester
could search for the most recentsaved_intermediate
model and load it instead ofmodel_best.pt