How do I interpret this lr_find graph?

ucr · February 5, 2019, 9:28am

mnpinto · February 5, 2019, 10:33am

It seems you can use a very high learning rate. Usually it’s advised to choose a region with a steep slope in the graph. But in doubt it’s always good to try a few options like 0.03, 0.1, 0.3, 1 and plot the losses.

ucr · February 5, 2019, 10:47am

Thanks! I will try them out.

Tchotchke · February 5, 2019, 1:34pm

Out of curiosity - what type of data or problem is this for? Vision, NLP, tabular?

ucr · February 5, 2019, 1:46pm

Its for vision , single object bounding box to be precise