Custom NN Model Conversion to Brevitas Version

[Ba1tu3han]

Hello Community,

I have a NN and I want to convert each layer to Brevitas version. Is there a list of the Brevitas layers? I could not find it in the website and GitHub.

Additionally, could you provide any source that explains how to change each layer with the Brevitas one? I think the website is not enough.

Best Regards,

[Giuseppe5]

You are right, we need to work on expanding that part of our documentation.

Regarding what layers we support, they are in the folder brevitas/nn, but in general we support all the most commonly used PyTorch layers, including RNN and LSTM.
I will open an issue to add a full list of supported layers to our documentation.

Regarding how to convert floating point layers to quantized one, Brevitas offers the possibility of automatically converting a floating point model to its quantized counterpart, preserving all the weights and structure.

You can take a look at

https://github.com/Xilinx/brevitas/tree/master/src/brevitas_examples/imagenet_classification/ptq/ptq_evaluate.py

To see how we do that for imagenet models.
In particular, I believe you are interested in the quantize function. As a reminder, this function takes care only of inserting quantization nodes, without optimizing any of the quantization scales or zero point. The rest of the script shows how to perform those operations as well.
In the same folder there is a tutorial with some explaination about the different quantization options we expose.

Another option would be to manually recreate the network but you'd need to preserve hierarchy so that you can re-load the state dict from the floating point one.

Feel free to reach out in case more help is needed.

[Ba1tu3han]

Hello Giuseppe,

Thank you for your reply and action.

I will look at the documents deeply.

Have a good day!

Batuhan