Fix gradient accumulation for cuDNN convolution layer

[slayton58]

Fixes the gradient accumulation issue in CuDNNConvolutionLayer reported in caffe BVLC/pull/3254

[lukeyeager]

Why not just cherry-pick @ronghanghu's commit? Commenting rather than deleting those lines may lead to a merge conflict down the road.

Also, why did you change the test if it wasn't needed in BVLC/pull/3254?

[slayton58]

I changed the gradient file to match what is in bvlc today, and without it the convolution tests fail

On Oct 28, 2015, at 4:17 PM, Luke Yeager notifications@github.com wrote:

Why not just cherry-pick @ronghanghu's commit? Commenting rather than deleting those lines may lead to a merge conflict down the road.

Also, why did you change the test if it wasn't needed in BVLC#3254?

—
Reply to this email directly or view it on GitHub.

[lukeyeager]

Oh sorry, I didn't notice this was for the master branch. I wish @github would make that a little more obvious somehow.

So we need this fix in the caffe-0.14 branch, too?

[lukeyeager]

Merged into master and caffe-0.13.