Current models perform ensembling through averaging layer. Independent training and network exposure, with averaging only during testing.