A Pruning Method Based on Feature Rank and Channel Importance

Abstract

The invention relates to a pruning method based on feature rank and channel importance, comprising the following steps: preparing a data set, selecting an original network model, and dividing it into a training set and a test set; training a converged unpruned network through the data set The original network model of the data set is used to input the trained model with some pictures in the data set; the unpruned network model added to the channel attention mechanism through the data set training is used to input the trained model with the training set, and the model parameters are fixed to obtain the network layer. Channel attention mechanism for weight vectors of different convolution kernels in different layers; pruning; get the global pruning results of different layers, and then calculate the number of remaining channels after each layer of pruning, and change the number of channels in each layer of the original network to The number of pruned channels is used to obtain the pruned network model, and then the data set is used to retrain to restore the accuracy. After continuous gradient backpropagation optimization, the optimal lightweight model is finally obtained.

Description

technical field [0001] The invention relates to the field of lightweight models in image processing, and is especially suitable for platforms with less computing resources. Background technique [0002] At present, because the neural network has the advantage of automatically learning and extracting appropriate visual features, it can avoid the trouble of manual design and extraction of features, and has high classification accuracy, low training difficulty, and strong robustness and fault tolerance to noise in features. It can fully fit the complex nonlinear relationship required in classification tasks, so using it for classification tasks is a current hot research issue. The classification problem of neural network refers to the use of relevant algorithms to design and train the model to classify the pictures in the task for a specific classification task. Its application fields are extremely wide, including face recognition in daily life, license plate recognition and g...

AI Technical Summary

Problems solved by technology

However, most of the current pruning methods have certain limitations, ignoring that the importance judgment conditions of parameters at different depths in the network should be different.

Images