Regarding LotFL performance

[kgbssj]

Hi all,

is anyone successful in reproducing the results (e.g. near 90% for CIFAR10 personalized model, 20 participants, 400 clients etc.) in LotteryFL paper? From my trials, I'm not sure why, but it just converges around 70~75% with the same parameters.

Some questions I have are,

• which data is used for local validations?
• are the test sets the sum of the n-classes' entire test data (e.g. 1000 for class 1 and 1000 for class 6)?

Need explanation for the performance gap. Don't see any major bugs.