We study network pruning which aims to remove redundant channels/kernels and accelerate the inference of deep networks. Existing pruning methods either train from scratch with sparsity constraints or minimize the reconstruction error between the feature maps of the pre-trained models and the compressed ones. Both strategies suffer from some limitations: the former kind is computationally expensive and difficult to converge, while the latter kind optimizes the reconstruction error but ignores the discriminative power of channels. In this paper, we propose a discrimination-aware channel pruning (DCP) method to choose the channels that actually contribute to the discriminative power. Based on DCP, we further propose several techniques to improve the optimization efficiency. Note that the parameters of a channel (3D tensor) may contain redundant kernels (each with a 2D matrix). To solve this issue, we propose a discrimination-aware kernel pruning (DKP) method to select the kernels with promising discriminative power. Experiments on image classification and face recognition demonstrate the effectiveness of our methods. For example, on ILSVRC-12, the resultant ResNet-50 with 30% reduction of channels even outperforms the baseline model by 0.36% on Top-1 accuracy. The pruned MobileNetV1 and MobileNetV2 achieve 1.93x and 1.42x inference acceleration on a mobile device, respectively, with negligible performance degradation.

Automatic reading for water meter is one of the practical demands in smart city applications. Due to the high cost, it is not feasible to replace the old mechanical water meter with a new embedded electronic device. Recently, image recognition based meter reading methods have become research hotspots. However, illumination, occlusion, energy and computational consuming in IoT environment bring challenges to these methods. In this paper, we design and implement a smart water meter reading system to handle this issue. Specifically, we first propose a novel light-weight spliced convolution network to recognize the meter number, which simplifies standard 3×3 convolutions by splicing a certain number of 1×1 and 3×3 size kernel. We then prove the superiority of our network by theoretical analysis. Second, we have implemented the prototype which can handle huge real-time data base on the distributed cloud platform. Base on this system, our system can provide industrial service. Finally, we conduct real-world dataset to verify the performance of the system. The experimental results demonstrate that our proposed light-weight spliced convolution network can reduce nearly 10× computational consuming, 7× model space, and save 3× running time comparing with standard convolution network.

Tingfeng Cao, Chengyu Wang, Bingyan Liu, Ziheng Wu, Jinhui Zhu, Jun Huang. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track. 2023.