`ExtractBiasFromConv` is not working properly with quantized biases

[balditommaso]

I have a quantized NN trained with Brevitas, I would like to apply ExtractBiasFromConv
transformation, but it cannot find the bias.

[makoeppel]

Hi @balditommaso, can you give a bit more content? Best would be a minimal example, which can be reproduced?

[maltanar]

@balditommaso I have a version of ExtractBiasFromConv on another branch which should be able to deal with a bias that still has its Quant node (i.e. not constant-folded):

https://github.com/fastmachinelearning/qonnx/blob/custom/RAS/src/qonnx/transformation/extract_conv_bias.py

Can you give this a try and see if it solves your problem?

[balditommaso]

Hi @maltanar, sure let me try it!

[balditommaso]

I am getting the following error with @maltanar implementation.

I think it might be probably due to the layer-wise quantization. Is it handle?

[balditommaso]

I wrote a personal transformation which deal with both quantized and full precision bias. I can shere it with you if you are interested.