Federated Learning Using Pytorch

Healthcare Clara Train AIAA
#1

Hi,

I appreciate the newly included feature called Federated Learning in Clara train SDK 2.0.
I wonder if I can write a pytorch model and prepare some decentralized data to try federated learning by using this SDK.
In https://docs.nvidia.com/clara/tlt-mi/clara-train-sdk-v2.0/nvmidl/byom.html#bring-your-own-model,
I would like to ask if I can use FL by making my model inheriting the class <ai4med.components.models.model.Model>.
If yes, can I use pytorch module to define the model? or I should only use TensorFlow module as the example in https://docs.nvidia.com/clara/tlt-mi/clara-train-sdk-v2.0/nvmidl/byom.html#bring-your-own-model.

gratefully thanks.

Terence.

#2

Hi
Thanks for your complements and interest in Clara train.

Bring your own model and components only supports TF for now. You should write your own model, run it as single standalone (non federated learning). then it should be straight forward to run it as federated learning.

Please let us know if you face any issues.

#3

Thanks for the information about TensorFlow!

Will there be support for PyTorch in the roadmap of Clara?

#4

Thank you for your interest in Clara Federated Learning.

PyTorch support is on our roadmap in the future, I cannot at this point commit to a timeline for that. Will keep the forums posted on latest and greatest updates for Clara SDKs.

#5

Hi. I’m not from the CLARA SDK team.

But I have a simulator that supports single/multi GPU for simulation: FL_PyTorch is publicly available on GitHub. – Konstantin Burlachenko – CS Ph.D. student in CEMSE Division at KAUST

But the thing is that currently, we did not include HE, SMPC, and DP in it.

  1. If privacy is not an issue for you
  2. If you’re interested in prototyping, not in deployment
  3. If you are ready (With help) to augment the simulator with things that you need (in case you wish for something specific).

You are welcome.

Our simulator is more orientated for research.