Automated Super-Network Generation for Scalable Neural Architecture Search

Weight-sharing Neural Architecture Search (NAS) solutions often discover neural network architectures that outperform their human-crafted counterparts. Weight-sharing allows the creation and training of super-networks that contain many smaller and more efficient child models, a.k.a., sub-networks. For an average deep learning practitioner, generating and training one of these super-networks for an arbitrary neural network architecture design space can be a daunting experience. In this paper, we present BootstrapNAS, a software framework that addresses this challenge by automating the generation and training of super-networks. Developers can use this solution to convert a pre-trained model into a super-network. BootstrapNAS then trains the super-network using a weight-sharing NAS technique available in the framework or provided by the user. Finally, a search component discovers high-performing sub-networks that are returned to the end-user. We demonstrate BootstrapNAS by automatically generating super-networks from popular pre-trained models (MobileNetV2, MobileNetV3, EfficientNet, ResNet50 and HyperSeg), available from Torchvision and other repositories. BootstrapNAS can achieve up to 9.87{\texttimes} improvement in throughput in comparison to the pre-trained Torchvision ResNet-50 (FP32) on Intel Xeon platform.