Global Response Normalization difference between CPU and Myraid

[idata]

Hello,

I've noticed a difference between the results of the GRN layer of OpenVINO on CPU versus Myriad. When I have a GRN layer on the CPU it normalizes the channels as I would expect, by dividing by the L2 norm of the channel vector, but using the Myriad plugin it appears to divide by the absolute value each channel value, turning the channel vector values into +1 or -1. Here are two Jupyter notebooks which illustrate the difference:

CPU: https://gist.github.com/donniet/0d0e8f604eb30aeb0215968dc4e20071

Myriad: https://gist.github.com/donniet/fffdc95526ffa5fa625068db4cdb0507

Has anybody else had this problem with the GRN layer? I've tried multiple devices with the same results, so I suspect the issue is with the plugin in OpenVINO.

Thanks,

-Donnie

[idata]

Hi @donniet

You can post your question in the Computer Vision forums here where the OpenVINO community is.

Thanks so much,

Sahira

[idata]

I have posted it here: https://software.intel.com/en-us/forums/computer-vision/topic/806929