Can finn accelerate 4w4a or 8w8a models ?

[YangNuoCheng]

Can finn accelerate 4w4a or 8w8a models ?

I noticed there no CNV_4w4a and CNV_8w8a data in TABLE
is the FINN didn't support 4bit accelerate?I got the accelerate idea from FINN-R.

I modified BNN_PYNQ in brevitas_examples to get CNV_2w2a 、CNV_4w4a and CNV_8w8a modules and use following code for export.

from brevitas.export import FINNManager
FINNManager.export(self.model, input_shape=(1, 3, 32, 32), export_path=self.args.network+".onnx")

I'm tring to accelerate the quantized model on PYNQ-Z2 board.
When I using cnv=get_test_model_trained("CNV",2,2)) or my export CNV_2w2a.onnx there have same bugs as '#437' (InferQuantizedStreamingFCLayer: Signed output requres actval < 0) in Lowering and Streamlining part.
But I didn't modified the construction of CNV.
How can I get the same CNV_2w2a result from TABLE？

[YangNuoCheng]

here are my exported onnx and notebook based on BNN-PYNQ
CNV.zip

[Ba1tu3han]

Hello @YangNuoCheng ,

Have you solved it? Could you explain it please?

Bests,