Questions about Equalization

quantization
1

There’re kinds of Configs within quantization, for example we have backend_config that specify which platform we are using, and QConfigMapping, but I still cannot fully understand that what the EqualiztionConfig suppose to do.

In many sample codes, this just a None input, and pytorch would generate the default one, can anyone share some informations about this?

2

Hi @cokespace2 ,

Could you link me to the EqualizationConfig you are referring to? I am not too familiar with this part of the codebase.

Thanks,
Jesse

3

Sure, for example the EqualizationConfig would be using here: pytorch/quantize_fx.py at main · pytorch/pytorch · GitHub

4

I think it’s a way to implement the weight equalization described in this paper: https://openaccess.thecvf.com/content_ICCV_2019/papers/Nagel_Data-Free_Quantization_Through_Weight_Equalization_and_Bias_Correction_ICCV_2019_paper.pdf

I don’t think you need to worry about this flag unless you’re looking to do weight equalization.

5

Thanks lot!
Then I should leave it be, thanks again.

1 Like