Not getting the accuracy for Radioml_with_cnns on FPGA

[zsrabbani]

I've been using the RadioML CNN notebook as my reference to implement a model on an FPGA. I followed the standard structure to generate the bitstream and I've managed to get it running on the FPGA (AMD Zynq™ UltraScale+™ MPSoC ZCU104). The challenge I'm facing now is that the accuracy on the FPGA is only about 14%, even though the model itself achieves around 70% accuracy. So I'm looking for help in figuring out where I might be going wrong.

[zsrabbani]

The quantization function that I used for training the model is based on brevitas-radioml-challenge-21.
It seems that the quantization in the tutorial is wrong.
I used the below function, but the accuracy of the model on FPGA is still 13%.

def quantize(x):
# First: HardTanh clips to [-1, 1]
x = np.clip(x, -1.0, 1.0)

 # Second: The quantization in Brevitas with CONST scaling
 # For symmetric quantization with 8 bits and range [-2, 2]:
 # The scale factor is: (2^(n_bits-1) - 1) / max_val = 127 / 2 = 63.5
 scale = 63.5

 # Quantize
 q = np.round(x * scale)

 # IMPORTANT: Since inputs are in [-1, 1] and scale is 63.5,
 # the output range is [-63.5, 63.5], so we clip to [-64, 63]
 # This is "narrow range" quantization for int8
 q = np.clip(q, -64, 63).astype(np.int8)

 return q