Filter Pruning Method for Inference Time Acceleration Based on YOLOX in Edge Device

Abstract

Convolutional neural network (CNN) has a lot of parameters and floating point operations (FLOPs), so it is difficult to use it in edge devices with limited resources. To solve this problem, the filter pruning method of our previous study was extended and applied to the state-of-the-art object detection network, YOLOX. In addition, the inference time of the pruned network was measured on NVIDIA Jetson Xavier NX using the PASCAL VOC dataset to confirm performance improvement in the actual edge device. When the target pruning rates of parameters and FLOPs were 40% and 30%, mean average precision <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\text{mAP})(0.5)$</tex> improved by 0.07%, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{mAP}(0.5:0.95)$</tex> decreased by 0.8%, and inference time improved by 19.48%. Also, when the target pruning rates of parameters and FLOPs were 40% and 50%, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{mAP}(0.5)$</tex> decreased by 0.57%, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{mAP}(0.5:0.95)$</tex> decreased by 2.84%, but the inference time was improved by 36.21%.