Can feature in autoscheduler apply into autotvm by stmt? if feature in autoscheduler is better than the feature in autotvm now?
Auto-scheduler’s feature is better. But we cannot directly feed auto-scheduler’s feature into autotvm’s xgb model. Because auto-scheduler’s model and autotvm’s model are different and they accept different feature formats.
Although auto-scheduler and autotvm both use xgboost, they do training and inference in different ways. Auto-scheduler implements a custom format and custom loss to train xgboost models.
We cannot merely port the feature, but we can port the whole cost model from auto-scheduler to autotvm. It should work well and it is expected to be more accurate and general.