Efficient-Net-B0 AIMET quantization

EfficientNet-B0 from timm repository, post quantization Sim is leading to very less accuracy.

If I try to update timm version (4.5.1) vs what AIMET requires (0.3.1) the AIMET quantization fails at onnx importer.

@akhobare Can you please suggest how to proceed?

Hi @abinila_siva

EfficientNet-B0 is a somewhat difficult to quantize model from our experience. Could you try applying the AdaRound technique in AIMET. Check the accuracy after that.
And then try the QAT (Quantization-aware training) feature in AIMET after that.

Thanks
Abhi

Hi @akhobare,

I have do some experiments on EfficientNet-B0, and find that the quantization of the SE layer is the main reason that harm the entire model accuracy. We then switch to EfficientNet-Lite0, which remove the SE layer and replace SiLU with ReLU, and found that it can be retrained to close to original accuracy. Is your observation echos to the above statement? Thank you for your attention!

Best regards,
Edan