search

alibaba / FederatedScope

An easy-to-use federated learning platform
https://www.federatedscope.io
Apache License 2.0
1.3k stars 210 forks source link

Trainable parameters are not properly filtered #639

Closed rayrayraykk closed 1 year ago

rayrayraykk commented 1 year ago

This is an issue to log bug fixes related to #634.

Known bugs:

  1. trainable parameters to be not properly filtered in trainer.py
  2. get_trainable_para_names are not real trainable param in context.py
  3. when using global evaluation the track_running_stats won't be updated in the server which will result in bad results
  4. bad results in gaussian_noise attack
  5. weird loss and bad results in async (caused by model delta: init_model - update, where init_model should filter running_stats.)