These are necessary to get decent performance out of the local mode.
Tree's in-memory expand wasn't taking the sampler or tree id, so it couldn't correctly do feature sampling. This led to overly-correlated trees and poor performance.
The local validate method was still generating (empty) predictions for instances that had been completely filtered out of the validation set (eg by the OutOfTimeSampler). This made performance look worse than it actually was.
tixxit
commented
8 years ago
These are necessary to get decent performance out of the local mode.
expandwasn't taking the sampler or tree id, so it couldn't correctly do feature sampling. This led to overly-correlated trees and poor performance.validatemethod was still generating (empty) predictions for instances that had been completely filtered out of the validation set (eg by theOutOfTimeSampler). This made performance look worse than it actually was.