Issue with the exp layer when the base is e

[emaggiori]

This fixes a bug in the exp layer. In the case of the base e (set by the default parameter base==-1), the outer_scale_ variable was computed by taking -1 as a base instead of e.

The new code uses exp(input_shift) instead of pow(base, input_shift) when the base is set to -1.

[jeffdonahue]

Thanks @emaggiori, this does seem like a bug, yet I did write tests for what I think should be this exact case, and those are clearly passing. So I'm not sure what's going on yet...any ideas? Ideally this would include a new test demonstrating the bug and that it's now fixed & passing.

[jeffdonahue]

Ah... I see now -- the bug is only triggered when shift is non-zero, which is not tested. (That makes more sense because I don't think I've ever used that parameter...) Could you add a new ExpLayer test for that case, after the existing ones I linked in the previous comment?

(This is just one more piece of evidence in favor of modularity -- and in particular in favor of splitting off the scale/shift params from Exp and Power layer, as @longjon has argued for.)

[emaggiori]

Hi, I added the tests. Let me know if this is okay.

[jeffdonahue]

Great, thanks @emaggiori! Please squash your changes to a single commit and I'll merge this.

[emaggiori]

Thanks @jeffdonahue, I squashed them to a single commit as suggested.

[jeffdonahue]

Looks good, thanks again @emaggiori.