CNN top-5 results are completely wrong for SqueezeNet and ZynqNet inference!

[pmousoul]

Hello,

I tried both DietCHai and CHaiDNN to classify images such as car, cat, espresso and the results don't even seem relevant; so something is going wrong.

I used the "python XportDNN.pyc" tool for quantization since the DynamicFixed method does not accept negative quantization parameters as it is the case with Ristretto.

I attach the .caffemodel files, the .prototxt files, as well as the FPGA executable output if anyone is interested to give a try - maybe I did something wrong..

Cheers,
Panos
model-results.zip