I think lr_find IS destroying the weights of a model (or something else is up)

I have been wrecking my brain about an issue in the latest version of Lesson 1 of Part 1 2019.

In the lesson I have no trouble training the dog/cats breed model with resnet34, but when I go on to using the resnet50 model, my model no longer seems to converge onto a decent error rate…

I figured out that it is due to the fact that I run learn.lr_find()…It seems to destroy the training, unless I re-initialize the learner…Other posts in this forum seem to suggest lr_find is only using a TMP version of the model, but…

So, this does not work:

But this one does…

Oh…Never mind…I see this was indeed a bug but already fixed.