Automatic pruning of convolutional networks based on Continuous search and adding skip connections

In recent years, convolutional networks have achieved impressive results in the field of machine vision. Their high accuracy has led to their use in security and military fields, such as crime prediction and prevention. However, due to the huge calculations required by these networks, they cannot be used in embedded systems with limited computing resources. One solution to this problem is network pruning. In this method, a portion of the network is pruned while maintaining the initial structure to reduce calculations and enable use in embedded systems. Traditional pruning methods typically require manual determination of the criteria and amount of pruning, and they tend to perform well on specific networks. Automatic methods, on the other hand, do not have these disadvantages. However, they must address issues such as relaxing the search space, eliminating overfit during the training process, and addressing differences resulting from relaxing the search space.This article introduces a method based on gradient decent for network pruning. The method automatically searches for the optimal pruned network and can be easily applied to networks with skip connections. To compensate for lost accuracy, dynamic skip connections are used without increasing computation. Tests on valid datasets demonstrate that this method achieves a better balance between accuracy and network pruning than previous methods.