Cloud Qualcomm hint more about AIMET's stochastic method?

#1

Thanks Qualcomm AIMET team.

we want to inquiry about the stochastic method.
What exactly is the AIMET’s stochastic method? Is that a different from AdaRound?
If there were any sample code will be helpful.

https://quic.github.io/aimet-pages/AimetDocs/api_docs/torch_quantsim.html

  • rounding_mode ( str ) – Rounding mode. Supported options are ‘nearest’ or ‘stochastic’

Best Regards.

#2

Hi @herokuvps - stochastic rounding is like the name suggests a rounding method that randomly chooses rounding per tensor value. This is meant to be used during QAT.

To be fair - we have not seen compelling results with this scheme and I will suggest to avoid it.

AdaRound has some similarities to stochastic rounding - the big difference being that AdaRound is a standalone post-training technique that learns how to round. So in a way AdaRound tries to chase the optimal stochastic rounding. We have seen excellent results with AdaRound.

1 Like
#3

Thanks a lot for your reply. well noted.