samuel-maddock / federated-boosted-dp-trees Goto Github PK

I want to do some experiments with Non i.i.d data, but I couldn't find the interface to set the number of clients and change the data on each client.

Hello, I encountered some problems when dealing with multi-classification problems. I have noticed that SoftmaxCrossEntropyLoss and one-hot coding are used for multi-classification, but the dimension of gradient g will also increase. For example, the shape of g will become (num_samples,num_classes) when dealing with four-category classification, and I don't know how to deal with g when calculating the split scores later (because g was one-dimensional before)

Hello Samuel,

I have a question in /models/gbdt/components/split_candidate_manager.py

At line 112, why flip the sign if min_val > 0? I thought the sketch should be operated on the interval [min_val, max_val]?

Thank you!