Lr_finder Suggestion way off from optimal point

I am training an LSTM model for caption generation. I find lr_find suggestion is way off from the optimal lr.
Here is the plot
image

I was expecting it to be somewhere around 8e-03 rather than where it is now. I would like to know what’s causing this behavior.

The suggestion takes the steepest point in the curve, which is not always accurate. That’s why you need to look at the graph and not trust the computer :wink:

Got it, Thanks!