Enables the alternate solvers to be accessed by the python interface.

[danielgordon10]

No description provided.

[shelhamer]

Thanks for including more solvers @danielgordon10, but please revise to add AdaDeltaSolver too.

[jeffdonahue]

I'll note that the other solvers can already be accessed by get_solver. But given that we already directly expose SGDSolver, it would be more consistent to expose the others as well. (An alternate route to consistency would be to remove SGDSolver and always use get_solver; I don't have a strong opinion either way.)

[shelhamer]

@jeffdonahue right -- I'm fine with either exposing them all like this or documenting the use of get_solver() even though SGDSolver is a special case by its popularity.

[danielgordon10]

@shelhamer Whoops, totally missed that one.
Yeah, I saw get_solver at the tail end of my debugging, after I had written the other exposers. I appreciate the explicitness of Solver though, especially because SGDSolver exists. I think if you want to encourage get_solver, you shouldn't even have SGDSolver as its own method.

[danielgordon10]

Either way, I'm pretty sure the NesterovSolver and AdaGradSolver lines in _caffe.cpp were uncallable without a change. So something needs to be done.

[danielgordon10]

Anyone know what the build fail is about?

[shelhamer]

From the travis log:

[  FAILED  ] 1 test, listed below:
[  FAILED  ] AccuracyLayerTest/1.TestForwardCPUPerClass, where TypeParam = double

which blame points to 374fb8c in #2935. @ronghanghu could you check what this is about?

@danielgordon10 the test failure has nothing to do with your PR, so don't worry about that. However do squash your fixup commit for AdaDelta for a clear history. Thanks.

[ronghanghu]

@shelhamer I'll checkout shortly.

[ronghanghu]

@shelhamer There is a bug with AccuracyLayer test in #2935 which I didn't observe. It was my fault. I made a fix PR in #3027.

[danielgordon10]

@shelhamer Is there an easy way to squash the commits into an already pushed branch? I'm having issues pushing the squashed version saying I'm behind the pushed one. And if I pull first, then I have a bunch of commits, which is exactly what squash is trying to avoid.

[shelhamer]

@danielgordon10 once you squash, force push by git push -f to your PR branch to update the PR.

[danielgordon10]

@shelhamer naughty naughty, using -f. I'll do it, but I won't enjoy it.

[shelhamer]

Not at all -- "published history" in the sense of the project is the canonical BVLC/caffe master branch. Force push is not only fine but right to wrap up PRs.