-
Notifications
You must be signed in to change notification settings - Fork 3
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Checkpoint Ablation for Flamby #71
Conversation
… heart disease implementations. Work towards ablation study.
@@ -50,7 +54,7 @@ def __init__( | |||
self.alpha_learning_rate = alpha_learning_rate | |||
self.client_number = client_number | |||
|
|||
def get_dataloader(self, config: Config) -> Tuple[DataLoader, DataLoader]: |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Somehow a typo crept into this APFL client
@@ -23,9 +23,9 @@ def __init__( | |||
metric_meter_type: MetricMeterType = MetricMeterType.AVERAGE, | |||
checkpointer: Optional[TorchCheckpointer] = None, | |||
temperature: Optional[float] = 0.5, | |||
perfcl_loss_weights: Optional[Tuple[float, float]] = (0.0, 0.0), | |||
cos_sim_loss_weight: Optional[float] = 0.0, | |||
contrastive_loss_weight: Optional[float] = 0.0, |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
If these are not None, then some of these losses (at least perfcl) are still computed, but just "zeroed" out. This is problematic if you want to use asymmetric latent spaces in a FENDA architecture.
Looks good to me. Only a few comments on some small typos |
Thanks for pointing out the helper message. Will fix! |
Adding in the ability to perform "non-federated checkpointing" for the FLamby experiments. This facilitates the checkpoint ablation study that was added to the ICLR paper.
Fixed two small bugs as well. The first was the dataloader method name for APFL (had a typo) and the second was defaulting the auxiliary FENDA losses to zeros rather than
None
. This prevents using non-symmetric latent spaces in FENDA, because you need symmetric latent spaces for some the auxiliary losses and they are computed even if the coefficients are zero.