Chameleon is an efficient continuous adaptation framework based on NVIDIA TAO. To bridge the gap between one-time domain adaptation and continuous learning, we propose Chameleon, which updates models on new data (labeled or unlabeled) via existing domain adaptation techniques and select the suitable model to recovery accuracy through adaptive model selection methods. In implementation, we provide different adaptation strategies and optimization techniques for different visual tasks (object detection, segmentation, tracking and SLAM). In the end, we summary existing common optimization techniques (GPU sharing, ...).
- GPU Sharing between training and inference: Ekya: Continuous Learning of Video Analytics Models on Edge Compute Servers. Published in NSDI'22.