Convolutional neural networks (CNNs) have been used to achieve computer vision applications for the past few years. These networks can be trained and applied in many fields, including image classification, object detection, semantic segmentation.
Inference speed on mobile devices based ARM architecture or CPU devices based x86 architecture has been challenging to get with the increase of model feature extraction capability and model parameters. Even many good mobile networks have been proposed to resolve this issue, but the speed of these proposed networks is not good enough on the Intel CPU due to various limitations of MKLDNN.
Baidu researchers have developed a lightweight CPU network based on the MKLDNN acceleration strategy, named PP-LCNet. This new system improves the performance of models in many tasks making it perfect for future artificial intelligence (AI) systems. The research group’s rethink the lightweight model elements for a network designed on Intel-CPU. During the research, the research team brings up following three questions to resolve:
- ‘How to promote the network to learn stronger feature presentations without increasing latency.’
- ‘What are the elements to improve the accuracy of lightweight models on CPU.’
- ‘How to effectively combine different strategies for designing lightweight models on CPU.’
The main contribution of this research is summarizing a series of methods to improve the accuracy without an increase in inference time and how they can be combined for a greater balance between speed. With this information, researchers can come up with several general rules for designing lightweight CNNs. They also provide new ideas on how other people can build their networks if they’re trying to get better models faster and find architectures that work best for them.
The proposed network, PPLCNet, has a better speed balance and shows strong performance with better results on computer vision tasks. This research network also reduces the search space of NAS and offers quicker access to NAS with lightweight models. The research experiments in this research were implemented based on PaddlePaddle.