Reproduce “PP-LCNet: A Lightweight CPU Convolutional Neural Network” on OpenResearch

About this paper

Cheng Cui, Tingquan Gao, Shengyu Wei, Yuning Du, Ruoyu Guo, Shuilong Dong, Bin Lu, Ying Zhou and 5 more

Baidu Inc. developed PP-LCNet, a lightweight convolutional neural network specifically optimized for efficient inference on Intel CPU devices utilizing MKLDNN. This model achieved improved accuracy and reduced inference latency across image classification, object detection, and semantic segmentation tasks compared to other lightweight architectures.

Key insights

• Replacing ReLU with H-Swish activation, strategically integrating Squeeze-and-Excitation (SE) modules at later layers, and employing larger 5x5 convolutional kernels in the network's tail significantly improve accuracy with minimal latency impact on Intel CPUs.
• A high-dimensional 1x1 convolutional layer positioned after Global Average Pooling enhances the model's fitting ability and feature retention without substantially increasing inference time.
• Optimizing architectural elements for specific hardware acceleration libraries, like MKLDNN on Intel CPUs, is crucial for achieving superior accuracy-latency trade-offs compared to general lightweight designs.

Summary by alphaXiv.