Shulou(Shulou.com)06/01 Report--

This article mainly shows you "what is the meaning of EfficientNet", the content is easy to understand, clear, hope to help you solve your doubts, the following let the editor lead you to study and learn "what does EfficientNet mean" this article?

Brief introduction

EfficientNet is derived from Google Brain's paper EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. As can be seen from the title, the main innovation of this paper is Model Scaling. This paper proposes three ways of compound scaling, hybrid scaling and network scaling: depth, width and resolution, which are combined to scale according to certain rules, so as to improve the effect of the network. When the network becomes larger, the effect of EfficientNet is improved obviously, and the upper limit of accuracy is further raised, which has become the strongest network at present. EfficientNet-B7 achieved the most advanced top-1 accuracy of 84.4% and top-5 accuracy of 97.1% on ImageNet, which is 8.4 times smaller and 6.1 times faster than the previous best convolution network (GPipe, Top-1: 84.3%, Top-5: 97.0%).

The main innovation of EfficientNet is not structure. Unlike ResNet and SENet who invented shortcut or attention mechanism, the base structure of EfficientNet is found by structure search, and then zooms out using compound scaling rules to get a series of excellent networks: B0~B7. The following two pictures show the variation of Top-1 Accuracy with the number of parameters and flops of ImageNet, respectively. You can see that the saturation value of EfficientNet is high and the arrival speed is fast.

Principle

Better accuracy can be obtained by increasing the network parameters (there is enough data, but under the condition of fitting). For example, ResNet can deepen the baseline model from ResNet-18 to ResNet-200,GPipe and quadruple the top-1 accuracy of 84.3% on the ImageNet data set. There are three ways to increase network parameters: depth, width, and resolution.

Depth refers to the number of layers of the network, width refers to the number of channel convoluted in the network (for example, wide resnet gains precision benefits by increasing the number of channel), and resolution refers to the input size through the network (for example, from 112x112 to 224x224)

Intuitively, these three zooming methods are not independent. For high-resolution images, a deeper network should be used because a larger receptive field is needed, and the width of the network should be increased to obtain finer-grained features.

Previously, the increase of network parameters is one of the three ways to zoom in separately, and there is no simultaneous adjustment, and there is no research on the adjustment mode. EfficientNet uses the compound scaling method to scale the depth, width and resolution of the network uniformly.

As shown in the following figure, (a) is the baseline network, (b), (c), (d) is the way to make the network larger by adding width,depth and resolution, and (e) is the compound scaling way.

The base network in EfficientNet adopts a similar method to MNAS (the only difference is that the target is changed from hardware delay to FLOPS). After using compound scaling, the effect is very significant, and the number of parameters and the amount of calculation are improved many times.

In addition, in order to verify the benefits of individual compound scaling, the base network does not use the search results of MNAS network structure, but directly uses MobileNets and ResNet as base for scaling. The results show that when the FLOPS is close, Top-1 Acc has an improvement of 0.6% to 1.4%.

Finally, the latency of EfficientNet is 5-6 times higher than that of ResNet/GPipe.

The above is all the contents of this article "what does EfficientNet mean?" Thank you for reading! I believe we all have a certain understanding, hope to share the content to help you, if you want to learn more knowledge, welcome to follow the industry information channel!

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.