The SGM uses uniform sampling with replacement.
OK, thank you very much
Was the SGM not sampled with a Poisson?Does Opacus have an RDP implementing Poisson Gaussian sampling?
there are some code or examples to achieve the Dynamic Privacy Parameters?
New Opacus (any version > 1.0) supports dynamic privacy parameters. This is handled by scheduler
: (See “Dynamic Privacy Parameters” in https://github.com/pytorch/opacus/releases/tag/v1.0.0 ).
sorry,I don’t know the meaning of the “distrubuted” . and I have a question that RDP calculations remain unchanged whether of Poisson sampling or not?
Hi @jeff20210616,
Yes, you are right. Currently we do not have an example in our tutorials about dynamic noise scheduling. I added to our list of items to add, and maybe in the future we will add such tutorial!
I am listing here the API reference for dynamic privacy parameter
You can also look at the code. Scheduling noise is achieved similarly to how learning rate in traditional machine learning training is schedule.
Hope you find this information useful
thank you very much!!!