Zilinghan Li comments

Results 14 comments of


                                            Zilinghan Li

credit dataset loading and more readiness metrics included

Hi @kaveenh, the code looks great. However, I think `train.csv` and `test.csv` are too large to put on Github. Could you please: 1. Revert the commit and do not make...

Inquiry on Integrating Homomorphic Encryption into the APPFL Framework

Hi, thanks for your question! 1. To use homomorphic encryption, I recommend using the following repo https://github.com/OpenMined/TenSEAL, which is built on top of Microsoft SEAF for tensor. You can use...

Inquiry on Integrating Homomorphic Encryption into the APPFL Framework

I think a great first step to write two functions `homo_encrpytion` and `homo_decryption` with the function interfaces to encrypt the model state dictionary `def homo_encrpytion(model: torch.nn.Module, **kwargs) -> Dict[str, Any]:`...

Update client optimizer

I will merge this branch into `zilinghan/refactor` branch first, and then for a merge together to the main branch

Clip norm does not clip gradients to given value

@minseok-ryu Hi Minseok, could you please leave some suggestions here, thanks

Clip norm does not clip gradients to given value

I think the main concern here is that Shourya just wants to check whether the norm of the model weight difference is equal to clip_value (i.e., 1 in this case),...

Memory-augmented asynchronous algorithms

> Is it better to move the initialization inside the client (needs to happen for all client types)? Also, it looks like the branch no longer passes all the checks?...

Memory-augmented asynchronous algorithms

@chariako I fixed it accordingly.

Memory-augmented asynchronous algorithms

Hi @chariako, I will look into this. However, the traceback for MIFT seems to be incomplete.

Checkpointing Evaluation Results and Enable Resume of Evaluation

Thanks for the prompt response! A quick follow-up: how about the evaluation via a local inference server? can it be resumed?