why comment the frcnn_proposal_layer's Forward GPU implementation？

[sigfrid696]

Hello,

I understand that this is a quite old repo, but I would like to know if the GPU implementation has some bug, since it is commented out.

Thanks

[sigfrid696]

I did some tests after porting the implementation to my library...the implementation seems ok because it produces the same results as obtained by the proposal layer done fully on CPU. The only problem is that is significantly slower

[sigfrid696]

Sorry I accidentally closed the issue...After doing some tests I found that only gpu nms processing of the proposal layer makes the overall process faster, the other parts of the layer remaining implemented on CPU...hope these conclusions can be of any help, if anyone is interested in the same algorithm

[makefile]

Yes, the GPU implementation in frcnn_proposal_layer.cu is not that efficient.

[sigfrid696]

Ok , thanks for the reply. After studying a bit the code, it doesn't seem related to the fact that is not an optimized algorithm. Probably some parts of the layer can be executed better on CPU, other on GPU, as the nms part, which is somehow similar to a convolution for the operations involved

[makefile]

Yes, you are right. The Soft-NMS operation is similar with this.