Feuermagier/Beyond_Deep_Ensembles

Missing/wrong hparams in cifar.yaml

emirkonuk opened this issue · 1 comments

Hello,

Thanks for the great work. When trying to replicate your experiments I noticed the "update_interval" parameter is missing for CIFAR10's swag optimizer.

There are some other minor typos like e.g. unnecessary "mean_samples" key or "swag_config" should be replaced by "swag" etc. in the same cifar.yaml file. Maybe you could share a more recent version of the cifar.yaml? If not, can you tell me what the "update_interval" parameter should be?

Best,
Emir

Hi Emir,

sorry for the confusion around the wrong configuration file. We ran most CIFAR experiments with an older version of the codebase...

I will test and update the file on the weekend, but for now the update_interval parameter should be 390:
update_interval is the interval (in batches) between successive mean samples taken by SWAG. Since we have 50000 training datapoints, a batch size of 128, SWAG runs for 50 epochs, and we want 50 mean samples, we get an update interval of (50 * 50000) / (128 * 50) = 390.625 which should be rounded down to 390.

Best,
Florian