data quantization

[ElonGUI]

Hi,
This is a comprehensive project.Have you ever quantified these models to make it run faster？(ex. bls2017)
I want to quantify weights and bias to 8-bit or 16-bit. The weights of GDN may be quantized to 32-bit fixed-point format. But it would negatively affect the quality of the compressed image. It is easy to compare the accuracy after quantization for classification networks, but not for compressed networks.
Do you think this is feasible？Thanks for you advice.