If this is true (red line would be the Validation set)
https://gerardnico.com/data_mining/overfitting
Is there any reason not to always use this?
# --------- CALLBACKS -------------------
callbacks=[fastai.callbacks.SaveModelCallback(learn, every=‘improvement’, monitor=‘valid_loss’)]
#-------------------------------------------------------
and then after lr_find do
if callbacks: learn.callbacks = callbacks
learn.fit_one_cycle(Nepochs, slice(lr));
And at the end
learn.load(“bestmodel”);
This would give us the model at the lowest point of the red line.