Fast Neural Network Adaptation via Parameter Remapping and Architecture Search

Deep neural networks achieve remarkable performance in many computer vision\ntasks. Most state-of-the-art (SOTA) semantic segmentation and object detection\napproaches reuse neural network architectures designed for image classification\nas the backbone, commonly pre-trained on ImageNet. However, performance gains\ncan be achieved by designing network architectures specifically for detection\nand segmentation, as shown by recent neural architecture search (NAS) research\nfor detection and segmentation. One major challenge though, is that ImageNet\npre-training of the search space representation (a.k.a. super network) or the\nsearched networks incurs huge computational cost. In this paper, we propose a\nFast Neural Network Adaptation (FNA) method, which can adapt both the\narchitecture and parameters of a seed network (e.g. a high performing manually\ndesigned backbone) to become a network with different depth, width, or kernels\nvia a Parameter Remapping technique, making it possible to utilize NAS for\ndetection/segmentation tasks a lot more efficiently. In our experiments, we\nconduct FNA on MobileNetV2 to obtain new networks for both segmentation and\ndetection that clearly out-perform existing networks designed both manually and\nby NAS. The total computation cost of FNA is significantly less than SOTA\nsegmentation/detection NAS approaches: 1737$\\times$ less than DPC, 6.8$\\times$\nless than Auto-DeepLab and 7.4$\\times$ less than DetNAS. The code is available\nat https://github.com/JaminFong/FNA.\n