GatherNet: The Lightest Convolutional Neural Network Architecture So far

Abstract It is generally difficult to establish convolutional neural networks with many operations on mobile devices due to their limited memory and computation resources. This paper proposes a novel Gather module that combines depthwise separable convolution and standard convolution with Ghost module to generate feature maps cheaply and efficiently and uses channel shuffle to rearrange these learned feature maps to improve the information flow between different feature channels. Based on the Gather module, we construct a novel network architecture called GatherNet which is the lightest convolutional neural network architecture so far. We also introduce a hard-swish activation function to effectively solve the data collapse when low-dimensional features are embedded in a high-dimensional space during training. Three benchmark datasets of CIFAR-10, ImageNet-1K, and VOC are used to evaluate our network, with the validation results showing that our proposed GatherNet achieves competitive classification and detection results with much fewer weight parameters than state-of-the-art lightweight network models. Particularly, our GatherNet works much better on a small set of training samples than other lightweight models and still shows the best performance with much fewer parameters and better accuracy when applying it to ocular surface disease recognition. The pre-trained GatherNet model with its code is available at GitHub: https://github.com/Rchen3233/GatherNet.