Parallel selection of hyperparameters to design a multi-branch convolutional neural network method for pedestrian recognition

Abstract

In the present invention, the method for selecting hyperparameters in parallel to design a multi-branch convolutional neural network for identifying pedestrians includes the steps of: 1) initializing the architecture of the multi-branch convolutional neural network; 2) self-adaptive, automatic screening of construction units according to specific tasks, and adding them iteratively and in parallel Each branch constitutes a branched convolutional neural network with separable accuracy and computational load; 3) verification and evaluation, retaining all models with the highest accuracy within 5% drop, and selecting the model with the least computational load as the selected structure; The multi-branch convolutional neural network is split, and a single branch or a combination of two branches is used as a benchmark model, stored in the terminal device, and offline reasoning and identification of pedestrians. In the present invention, each branch of the multi-branch convolutional neural network, according to the specific pedestrian recognition data set, selects more than one parameter in each iteration cycle to generate building blocks, and these building blocks are added to each branch in parallel for training, and then selected A hyperparameter preservation model with excellent performance is obtained.

Description

technical field [0001] The invention belongs to the field of deep learning research, in particular to a method for selecting hyperparameters in parallel to design a multi-branch convolutional neural network for identifying pedestrians. Background technique [0002] The quality of machine learning and deep learning models largely comes down to "tuning parameters", that is, the optimization of hyperparameters. Convolutional neural network, one of the main algorithms of deep learning, handles tasks such as target recognition, detection, instance segmentation, scene understanding, and reinforcement learning. It mainly needs to complete the design of the network structure and the adjustment and optimization of hyperparameters. Depending on manual parameter tuning, it needs to rely on experience and intuition such as hyperparameters, training errors, generalization errors, and computing resources, and it takes a lot of time and effort. Therefore, more and more hyperparameter tuni...

AI Technical Summary

Problems solved by technology

First, a distribution is selected for each hyperparameter set, for example, the Bernoulli distribution, and then a random combination of hyperparameters is generated from the hyperparameter set according to this distribution, which is sent to the network for training, and the best combination of accuracy is selected, but random Search is suitable for rough selection and census, and it is easy to miss key data

Images