DyNet: Dynamic Convolution for Accelerating Convolution Neural Networks

Kane Zhang, Jian Zhang, Qiang Wang, Zhao Zhong

TL;DR: We propose a dynamic convolution method to significantly accelerate inference time of CNNs while maintaining the accuracy.

Abstract: Convolution operator is the core of convolutional neural networks (CNNs) and occupies the most computation cost. To make CNNs more efficient, many methods have been proposed to either design lightweight networks or compress models. Although some efficient network structures have been proposed, such as MobileNet or ShuffleNet, we find that there still exists redundant information between convolution kernels. To address this issue, we propose a novel dynamic convolution method named \textbf{DyNet} in this paper, which can adaptively generate convolution kernels based on image contents. To demonstrate the effectiveness, we apply DyNet on multiple state-of-the-art CNNs. The experiment results show that DyNet can reduce the computation cost remarkably, while maintaining the performance nearly unchanged. Specifically, for ShuffleNetV2 (1.0), MobileNetV2 (1.0), ResNet18 and ResNet50, DyNet reduces 40.0%, 56.7%, 68.2% and 72.4% FLOPs respectively while the Top-1 accuracy on ImageNet only changes by +1.0%, -0.27%, -0.6% and -0.08%. Meanwhile, DyNet further accelerates the inference speed of MobileNetV2 (1.0), ResNet18 and ResNet50 by 1.87x,1.32x and 1.48x on CPU platform respectively. To verify the scalability, we also apply DyNet on segmentation task, the results show that DyNet can reduces 69.3% FLOPs while maintaining the Mean IoU on segmentation task.

Keywords: CNNs, dynamic convolution kernel