v2.7.2
Use sparse_categorical_crossentropy for test as the loss object default does not work on tpustrategy + the single task trainer already handles the reduction.
Internal change