Channel softmax

[ronghanghu]

In this pull request, the behavior of SoftmaxLayer is changed from softmax over channels*height*width elements (all elements within a num) to softmax over channels elements (all elements at a spatial position within a num). This is for the purpose of running fully-connected layers as convolutions (see Net Surgery: http://nbviewer.ipython.org/github/BVLC/caffe/blob/master/examples/net_surgery.ipynb). It won't damage existing caffe examples, since fully-connected layer top blob has width==1 and height==1.

The CPU version was implemented by @longjon, and I implemented the GPU version, including GPU backward.

[jeffdonahue]

This kernel is the same as caffe_gpu_exp isn't it? Let's remove it and replace with caffe_gpu_exp, unless I'm misunderstanding somehow. (I know it wasn't added this PR, but I just noticed it from seeing the diff.)

[ronghanghu]

I can't find caffe_gpu_exp. I only found caffe_exp, which calls vsExp in MKL.

[jeffdonahue]

whoops, my bad, I think I was thinking of caffe_gpu_powx. caffe_gpu_exp should probably exist but device abstraction (#610) will probably take care of this so never mind, sorry!

[jeffdonahue]

assigning to @longjon, go ahead and merge when you're happy with everything

[longjon]

@shelhamer suggested offline adding a switch to provide the original "normalize over everything" mode. So, @shelhamer, if you still want to do that, you can append to or rewrite this PR.

@shelhamer and others, which mode do we think should be the default? It seems like the channel normalization is usually what is desired, and I doubt anyone is relying on the current behavior, although it is a little jarring to change what layers do. If we do want the default to be the channel normalization, we could go ahead and merge this, and add a switch in a later PR.

[ronghanghu]

Also note that the updated SoftmaxLayer CPU no longer allows in-place computations, since in CPU implementation bottom diff is first changed and then restored
caffe_mul(top[0]->count(), bottom_diff, top_data, bottom_diff);
while GPU implementation still allows in-place computations.

@jeffdonahue @shelhamer should we allow in-place computations in SoftmaxLayer?

[longjon]

Yeah, good catch. I did this to avoid an extra loop, but I've added it back now to allow in-place computation. There should be no performance regression in the 1x1 case, and probably not a noticeable one in the general case, and anyway the GPU implementation is available.

In order to do this, I had to add functions to math_functions for strided dot products (which of course cblas already supports, but we didn't previously have an interface for.)

[shelhamer]

@longjon merge as you please, as the switch can follow. I agree channel is
a reasonable default. Although it does change the default behavior, I
imagine anyone who has adopted the fully-convolutional models wants channel
softmax.

On Saturday, August 16, 2014, longjon notifications@github.com wrote:

@shelhamer https://github.com/shelhamer suggested offline adding a
switch to provide the original "normalize over everything" mode. So,
@shelhamer https://github.com/shelhamer, if you still want to do that,
you can append to or rewrite this PR.

@shelhamer https://github.com/shelhamer and others, which mode do we
think should be the default? It seems like the channel normalization is
usually what is desired, and I doubt anyone is relying on the current
behavior, although it is a little jarring to change what layers do. If we
do want the default to be the channel normalization, we could go ahead and
merge this, and add a switch in a later PR.

—
Reply to this email directly or view it on GitHub
#940 (comment).

[longjon]

@ronghanghu I amended your commit with some aesthetic changes (make all the channel kernels have the form kernel_channel_[word], fix some lint errors being masked by NOLINT, and fix capitalization in comments). I think this will be ready for merge once Travis passes. The GPU implementation is a little heavy in terms of introducing lots of kernels instead of calling gpu_gemm and so forth, but it does the right thing by parallelizing over both batch and spatial dims, so I'll take it. Thanks for getting this written!

[shelhamer]

Fixed order of specialization and instantiation for clang++ build in ac64a7b. You can't call caffe_cpu_strided_dot() before its specializations as was done in caffe_cpu_dot().