Reproduction of baseline

[Fodenlu]

Hi! Thanks for your excellent work.

I noticed that you stated in Table 8 of your paper that the results of EfficientNet's 4-bit quantization LSQ (baseline) were reproduced by yourself, but when I reproduced it with this code (with osc_freeze.threshold=None and osc_damp.weight_final=None), I found that the accuracy of the baseline was 73.2, which does not match the results in your paper (72.3).

I can reproduce the results of LSQ+Freeze and LSQ+Dampen in the paper, but the results of baseline are much higher than those listed in the paper, what code does your baseline use? Did I get it wrong?
Thanks! Looking forward to your reply!

[mfournarakis]

Hi Fondelu,

Looking at our EffecientNet results, I believe we may have a typo in our Table. I think the LSQ baselime result could be 73.2%. I should mention that our training pipeline for this model does not 100% represent the one it was trained on. If you want to achieve even higher accuracy, I would recommend using the training pipeline, used by https://github.com/rwightman.

Thank you for bringing this to our attention.