New model HW layer acceleration

[reshkor]

Hello,
I have added some extra layers on a CIFAR-10. I managed to make a successful simulation of the model and execute software inference on PYNQ board. However if I try to run HW acceleration the results for all classes are 1. I guess the weights are not properly loaded on FPGA or some configuration in main_python.cpp is wrong (in OFFLOAD sections as software inference runs just fine).
Could you suggest how to fix this issue?

[giuliogamba]

Hi,
without having a look at the code it's very difficult to debug it. If you fork the repo and push your code in the forked repository, maybe I can have a look at it.

[reshkor]

Thank you for your fast reply.
I have created a fork and pushed to https://github.com/reshkor/BNN-PYNQ (the time on PYNQ is wrong, so don't mid that commit was done 12 days ago).
The new network is called cifar10_v5 and that one actually has only 3 convolution layers.

[Changyiyu]

@reshkor Hi, I referenced the link of https://github.com/reshkor/BNN-PYNQ and I don’t see the settings for these parameters in "cifar10-gen-binary-weights.py".

#topology of convolutional layers (only for config.h defines)
ifm = [32, 30, 14, 12, 5, 3]
ofm = [30, 28, 12, 10, 3, 1]
ifm_ch = [ 3, 64, 64, 128, 128, 256]
ofm_ch = [64, 64, 128, 128, 256, 256]
filterDim = [ 3, 3, 3, 3, 3, 3]

WeightsPrecisions_fractional =    [0 , 0 , 0 , 0 , 0 , 0 , 0, 0,  0]
ActivationPrecisions_fractional = [0 , 0 , 0 , 0 , 0 , 0 , 0, 0,  0]
InputPrecisions_fractional =      [7 , 0 , 0 , 0 , 0 , 0 , 0, 0,  0]
WeightsPrecisions_integer =       [1 , 1 , 1 , 1 , 1 , 1 , 1, 1,  1]
ActivationPrecisions_integer =    [1 , 1 , 1 , 1 , 1 , 1 , 1, 1,  1]
InputPrecisions_integer =         [1 , 1 , 1 , 1 , 1 , 1 , 1, 1,  1]

Can these be omitted? Because I'm worried about changing the size of the input image.