Inconsistent Output from CPU vs NCS

[idata]

Hi there,

I'm having difficulty getting equivalent results from the ncs and cpu, and was wondering if anyone else has encountered similar issues?

Steps to reproduce:

1. Take the AlexNet example from the ncappzoo: https://github.com/movidius/ncappzoo/tree/ncsdk2/caffe/AlexNet
2. Run make all, a deploy.prototxt and bvlc_alexnet.caffemodel weights file will appear
3. Take deploy.prototxt and remove everything after the conv1 layer - expected output size is now 96 x 55 x 55
4. Generate a new graph using mvNCCompile
5. Run the graph on the stick, and store the output of 290,440 floats ( use script similar to this: https://1drv.ms/u/s!AosBHN3IHTLWhoJ7Hk3DnpRZMOo31w )
6. Run the model on the cpu, and store the output of 290.440 floats ( use script similar to this: https://1drv.ms/u/s!AosBHN3IHTLWhoJ840l2yBCQlVo-aA )
7. Compare results in something like Matlab, Octave, or Python
8. Output looks like this: https://1drv.ms/u/s!AosBHN3IHTLWhoJ9GmzKhS7sIxurqw

Environment:

NCSDK v2.05.00.02Caffe v1.0.0Ubuntu v18.04 in VirtualBox 5.2.16

Many thanks..

[idata]

@Jay Have you tried compaing the results from the ncappzoo Alexnet example (https://github.com/movidius/ncappzoo/blob/ncsdk2/caffe/AlexNet/run.py) instead of the example at https://1drv.ms/u/s!AosBHN3IHTLWhoJ7Hk3DnpRZMOo31w? I'm curious to know what the results would be. I wonder if it could be a preprocessing issue since none of the appzoo examples use the transpose preprocessing step seen in the python scripts above img = img.transpose((2,0,1))

[idata]

Hi @Tome_at_Intel,

Thanks for that - it seems to have fixed things! The new histograms are here:

NCS Version:

https://1drv.ms/u/s!AosBHN3IHTLWhoMCk3rqG5v6z2-Crw

PC Version:

https://1drv.ms/u/s!AosBHN3IHTLWhoMBzM8rCJ1h_NztIQ

All the best!

[idata]

:)