Brief reading of dynamic networks and conditional computing papers and code collection

Preface

In this paper, the conditional computation and dynamic computation for computer vision are summarized cnn, This paper focuses on how to reduce the computing cost of the existing network structure . Compared with static network , The dynamic network disables part of the network according to the input image during inference . This can save computation and speed up reasoning , for example , Processing simple images with fewer operations . Be careful , This list focuses on ways to reduce the computational cost of existing models ( for example ResNet Model ), It doesn't list all the ways to use dynamic computing in a custom schema .

background

The method has three important distinguishing factors :

• The architecture of the method , for example , Skip layers or pixels , And these run or skip decisions are separate policy networks 、 The result of submodules or other mechanisms in the network .

• Methods of training strategies , For example, use reinforcement learning , Gradient estimator , Such as Gumbel-Softmax Or custom methods .

• The implementation of this method , And whether the method can be efficiently implemented on the existing platform ( That is, whether the method accelerates the reasoning , Or just reduce the amount of theoretical calculation )

indicators : Compared to the loss of accuracy , Most methods reduce the amount of computation ( That is to say, it is measured by floating-point operation ,FLOPS) To demonstrate performance . Methods usually show baseline models of varying complexity ( for example , By reducing the number of channels ) A chart comparing the method used by the largest model with different cost savings .

Related papers and codes

https://github.com/thomasverelst/awesome-dynamic-conditional-networks-cv