Exporting FINN-ONNX from pre-trained network in PyTorch

[MJokar1643]

Hello;
Thank you for your wonderful tool.

Here it is said that to export FINN-ONNX, the network can be trained in PyTorch or Brevitas. As far as I looked in the tutorials, I did not see that FINN-ONNX was generated from a pre-trained network in PyTorch. Is there any example or documentation for this?

[Ba1tu3han]

Hello,

I'm looking for the same example. Did you find any document?

[MJokar1643]

Hello,

No, unfortunately. I couldn't find an example. But I defined my own model and was able to produce FINN-ONNX in both training modes with PyTorch and training with Brevitas.

[Ba1tu3han]

Hello,

It is written here how to export basic model to ONNX format after training.

Weight-activation-bias model
export_onnx_qcdq(quant_weight_act_bias_lenet, torch.randn(1, 3, 32, 32), export_path='4b_weight_act_bias_lenet.onnx')

FINN expects FINN-ONNX and not all ONNX graphs are supported.

I could not find how to export to FINN-ONNX. Could you help me about in this point?

Best,

Batuhan