Strange Results from LR Finder

Hi, while training model today I observed the strange results from using learning rate finder.

When running it on cnn_learner after creation of learner, it provided the graph shown on attached picture

However, when trying to train actual model, model training started to converge with learning rate 1e-4, which is not the conclusion, which I would do from this LR finder graph.

Any advice on reasons, why LR finder plot does not provide good guidance in this case? (Training dataset - tiny imagenet for devise task)

I am having a nearly identical result, and came across your post @hedgehog. Did you find a solution?