I wish to normalize across channels (I have 30 channels). To do so (unless I misunderstand) I can use the LRN layer with norm_region=ACROSS_CHANNEL and local_size = 61. The local_size = 2*num_of_channels+1 with zero padding should ensure the normalisation factor is constant across channels.

If I do this in CPU, it works fine. If I run it in GPU when norm_region is larger then the number of channels I get NAN's.

Any ideas?

Thanks