SGD fails to converge when order of the data is sequential about samantha HOT 3 CLOSED

@qian2015 When the production samantha server rebuilds its model, is the order of the data randomized or sequential? If it's sequential, we should think about randomizing it for better performance.

from samantha.

Sgd generally have better performance after permutating input, but I think you should be able to tune parameters to converge. Indexer might be a better place to do permutation, e.g. implementing a PermutatedIndexer just like GroupedIndexer. Then in production, training model can be configured to use this type of indexer.

from samantha.

thanks for pointing out. this is a serious issue. production has the same problem. i'm working on it now.

also, you are right that dao might be the place to do permutation. currently, the relationship of indexer, dao and retriever is a little messy. it needs cleaning up and refactoring.

from samantha.