Learning to Learn Parameterized Classification Networks for Scalable Input Images

Convolutional Neural Networks (CNNs) do not have a predictable recognition\nbehavior with respect to the input resolution change. This prevents the\nfeasibility of deployment on different input image resolutions for a specific\nmodel. To achieve efficient and flexible image classification at runtime, we\nemploy meta learners to generate convolutional weights of main networks for\nvarious input scales and maintain privatized Batch Normalization layers per\nscale. For improved training performance, we further utilize knowledge\ndistillation on the fly over model predictions based on different input\nresolutions. The learned meta network could dynamically parameterize main\nnetworks to act on input images of arbitrary size with consistently better\naccuracy compared to individually trained models. Extensive experiments on the\nImageNet demonstrate that our method achieves an improved accuracy-efficiency\ntrade-off during the adaptive inference process. By switching executable input\nresolutions, our method could satisfy the requirement of fast adaption in\ndifferent resource-constrained environments. Code and models are available at\nhttps://github.com/d-li14/SAN.\n