Working with Custom model

[ppccww0201]

Hi.

I tried the vgg10-radioml example. I synthesized the vgg10-radioml model with the command ./run-docker.sh build_custom $FINN_EXAMPLES/build/vgg10-radioml using FINN.

Ultimately, I want to export a custom model, constructed based on Brevitas (or any other framework), as ONNX and synthesize it with FINN, rather than using the provided vgg10-radioml ONNX. I have few questions and I don't necessarily need answers to all of them. I hope to receive advice that helps me achieve my ultimate goal.

1. After exporting my custom Brevitas model to ONNX and replacing the ONNX file in $FINN_EXAMPLES/build/vgg10-radioml, I encountered numerous errors. Is the $FINN_EXAMPLES/build/vgg10-radioml/build.py specifically written for the vgg10-radioml model? Can I not synthesize my custom model using this build.py code?

2. If there is code available for the vgg10-radioml model before exporting it to ONNX, I think I can build my custom model based on that and synthesize it with FINN smoothly. The vgg10-radioml model code likely incorporates an understanding of FINN and adheres to many FINN constraints, making it easier to transplant my custom model. Could I obtain the code for the vgg10-radioml model? (Even if it's not based on Brevitas)

Thank you.

[fpjentzsch]

Hi,
I think the FINN Github discussions might be a better place to discuss this: https://github.com/Xilinx/finn/discussions

The model should be based on this notebook and you can refer to this related thread for how to "pre-process" the exported model for FINN, although the export in the "FINN-ONNX" format has since been deprecated in favor of the QONNX export.

In general, you will most likely need to customize the build steps to fit your new model, as you can tell by the different "custom steps" that are used for the different finn-examples. For example, the VGG10 requires additional transformations to deal with the 1D convolution, while a ResNet will require additional care to streamline residual connections.