| The architecture of a neural network is closely related to its performance,and a highperformance Convolutional Neural Network(CNN)requires experienced professionals to consume a lot of time and effort to design.Therefore,Neural Architecture Search(NAS)is dedicated to automatically finding the optimal network architecture and reducing the cost of CNN design.However,in the automatic blind inspection project of product quality(referred to as product blind inspection),the number of categories in the data set is often increased according to the demand,causing NAS to re-execute and retrain the found models,which greatly reduces production efficiency.In order to solve the problem of retraining the CNN model due to the increase of data set categories in NAS,this paper proposes a differentiable incremental branch search space and an adaptive NAS algorithm for analog increase for the image classification data set of product blind inspection,proposing the differentiable incremental branching in neural architecture search(DIB-NAS)to solve the retrain problem.The main work of this paper is as follows:(1)A differentiable incremental branch search space is designed,which enables the CNN model to have dynamic growth and search capabilities.The search space is mainly composed of differentiable incremental branching cells(DIB).In order to adapt to the dynamic increase of data set categories,DIB uses the feature fusion mechanism to access other modules to realize the growth function,and searches for the optimal candidate operation through the gradient optimization strategy.The differentiable incremental branch search space uses a breadth-first growth strategy to make the CNN model incrementally grow branches to extract different features at the same scale.Compared with increasing the depth of the neural network,increasing the width during the growth of the neural network is beneficial to extracting multi-scale features to enhance network performance.Therefore,the priority is to increase the width of the growth branch of the neural network,and then grow the depth when the performance is insufficient,and then access the branch.(2)Based on the differentiable incremental branch search space,an adaptive neural architecture search algorithm for dynamically increasing categories is designed.The algorithm combines differentiable incremental branch search space,uses real-time pruning,weight sharing and unbalanced learning to adapt to category increase,and plans branch growth timing,real-time pruning timing,and category-oriented growth process.Real-time pruning can improve search efficiency.Existing DIBs are pruned every time a branch grows,preventing too many DIBs searched in the model from causing lower and lower search efficiency.Therefore,when the search is completed,the network weights have been optimized enough and do not need to be retrained.Weight sharing and unbalanced learning allow the neural network to converge quickly when facing increasing and decreasing categories.Experiments have shown that DIB-NAS can solve retraining problems and achieve high metrics when categories increase.The experiment uses a product blind inspection data set of 3to 6 categories with a sample size of 13,500.The searched CNN does not require retraining and exceeds the classification performance of DARTS(the m AP of DIB-NAS is about 97.4%,and the EMR is about 89.9%),and surpassed manual CNNs such as Alex Net,Res Net-18,Res Net-50.DIB-NAS search takes about 20 GPU-hours,which is only about 44% of the total timeconsuming of DARTS. |
