Questions on the README - quantization

[mkvenkit]

Hello, thanks for this very interesting project! Just wanted clarification on these two statements from your readme:

1. arm_sign_nb_45: ...This version runs but there is an issue with the way the weights were quantized and stored in the weights header.
2. quantization.ipynb: ...Note that the implementation of quantization in this code is incorrect

I was able to build your example on an nRF52840. So my question is whether the above two issues are still unresolved.

[IngyN]

Hi, I was able to get the project to build and run with the libraries and on the chip but the network didn't run properly, it ran with no errors but it didn't recognize the pictures like it should have based on the inference (I was using a pretrained version and loading the weights directly). I didn't have time to debug and figure out exactly why but I believe it might be how I set up / "quantized" the network weights in the quantization file.