Although state-of-the-art machine learning models are getting bigger and bigger, ML engineers are also facing more and more requirements for training and inference efficiency, including as they begin to deploy ML models at the edge. Recently, Google OSS’ed the code for EfficientNetV2, a family of smaller and faster image classification models that outperform previous models in training speed and parameter efficiency. Unlike EfficientNetV1, EfficientNetV2 models use training aware neural architecture search – an AutoML methodology – to jointly optimize the model size and training speed.