1

Topic: Convolution for the arbitrary dimensionalities

Prompt, there are now fast engines (Cuda, OpenCL) to which do not need to scale a picture for operation with convolution networks? I while worked only with Caffe, from , he like is not able yet.

2

Re: Convolution for the arbitrary dimensionalities

Hello, Nik, you wrote: N> Prompt, there are now fast engines (Cuda, OpenCL) to which do not need to scale a picture for operation with convolution networks? N> I while worked only with Caffe, from , he like is not able yet. If the network is trained on 400400 it is necessary either to scale, or to transit under the image this window with overlapping and to do non maximum suppression after. Or you another mean something?

3

Re: Convolution for the arbitrary dimensionalities

Hello, Nuzhny, you wrote: N> If the network is trained on 400400 Convolution networks train on matrixes of smaller dimensionalities, on idea training activity for actually convolutions should not to depend on dimensionality of pictures. Scaling layers - too do not depend on dimensionality, total a question only what to do at the last stages.