find_learning_rate does not work anymore

[lucaventurini]

Describe the bug
Before, I've run several experiments with which I could plot the optimal learning rate as in the tutorial. Now, after an upgrade of the codebase, I cannot anymore reproduce the same experiments.
What I find is similar to issue #986 , but in my case the learning rate is never updated (while in the above issue is updated at least at the first step).

To Reproduce
I've run the experiments with commit 24b72a6 , and set a start lr of 1e-08 and an end lr of 1000, with 200 iterations. I get a tsv with 200 lines where the lr is always the same (slightly above the start lr).

Expected behavior
I previously run the same experiments with commit 6c27431 and they seemed to work as in the tutorial. The lr in the tsv spanned all the interval specified.