Dave Brown

Makes sense - I will check out the caching example. I am using C++. I presume the slowdown is caused by a runtime compile of the kernel under then hood....

I think you are partially correct but only when the ceil() is not used to actually 'round up' in the pooling layer. For example, using the very simplified PROTOTXT, shows...

@manuelschmidt thank you for pushing me to dig a little deeper to get it right!

Requested change at lines 101-102 of UnpoolingLayer::Reshape unpooled_height_ = static_cast((height_ - 1) * stride_h_ + kernel_h_ - 2 * pad_h_) - ((height_ % 2 == 0) ? 1 : 0);...

pull request opened.

I think that there is one more fix that needs to be added to take into account lines 94-102 of the pooling_layer.cpp which are used to align the pooling to...

working on it...

I created two pull requests: 1.) with the fix in the unpooling layer and 2.) additional tests in the test_unpooling_layer.cpp that now test the unpooling with padding.