Data augmentation via random rotations

[dgolden1]

Some datasets are indifferent to which direction in the image is "up" and therefore can be augmented with random rotations. This PR implements a rotate option in the TransformationParameter message which operates analogously to mirror. It's false by default; setting it to true causes images to be randomly rotated by an integer multiple of 90 degrees. If mirror is also enabled, there are 8 possibilities for the randomly transformed output image.

An error is thrown if the input data is not square (i.e., if datum.height() != datum.width()); the logic for dealing with square rotations is easier to start with, and I believe that transformed_data is expected to have the same shape as the input data.

I added unit tests to test_data_transformer.cpp by copying and pasting some pre-existing mirroring unit tests; the tests show that this transformation runs, but they don't prove that it does what I expect it to do (i.e., rotate an image properly). I'm not a pro at Caffe unit tests, so if anyone can suggest how to write a unit test that confirms that the rotation is working properly, I'd be happy to implement it. I can (and probably will) copy the logic to a little test script in Matlab to convince myself that the logic is working, but that's not really good enough as a unit test.

I have not yet tried to train a network using the rotate parameter but that's next on my list. In the mean time, eyeballs on my rotation (and generalized mirroring) logic are welcome.

Update: I've trained a network using the random rotations and it worked just fine (if not necessarily any better than when random rotations were disabled).

[bhack]

I remember that @Bigene has something extended to other transformation (affine?)

[dgolden1]

@bhack, can you link to the PR?

The advantage of 90 degree rotations is that they're nearly computationally free; they can be implemented by messing with input and output indexing, just like mirroring. Affine transformations are significantly more expensive, but potentially also very useful.

[dgolden1]

Here's a Matlab test of the rotation logic: https://gist.github.com/drdan14/99dc38fb363913f295f5. I tried to copy the C++ rotation logic as closely as possible, offensive as nested for loops in Matlab are. Because C++ is 0-indexed and Matlab is 1-indexed, a tweak here and there was required.

The output looks good to me.

[bhack]

See @Bigene comments here #1070

[dgolden1]

I don't have the bandwidth to rebase this onto master, so I am closing it.

[bhack]

@drdan14 Nevermind. I'm not satisfied with this but often caffe PR are subject to neverending rebase over the months.