A Channel Pruning Optimization With Layer-Wise Sensitivity in a Single-Shot Manner Under Computational Constraints

In the constrained computing environments such as mobile device or satellite on-board system, various computational factors of hardware resource can restrict processing deep learning (DL) services. Recent DL models image analysis mainly require larger memory occupation for intermediate feature map footprint than given specification and overhead (in FLOP) to meet service-level objective in sense accelerator. As one solutions, we propose a new method controlling layer-wise channel pruning single-shot manner that decide how much channels prune each layer by observing dataset once without full pretraining. To improve robustness performance degradation, also sensitivity formulate optimization problems deciding ratio under target constraints. paper, optimal conditions are theoretically derived, practical optimum searching schemes proposed using conditions. On empirical evaluation, methods show on present feasibility serving reducing occupation, providing acceleration effect throughput improvement while keeping accuracy performance.