In this study, particle swarm optimization (PSO) is applied for the hyperparameter optimization problem of Convolutional Neural Network (CNN) model. The design objective is to improve the prediction accuracy. Several PSO algorithms are compared in the classification problem of a CIFAR-10 dataset. The design variables of PSO are defined as continuous values. Hyperparameters of CNN model may take discrete or categorical variables, so the design variables defined as continuous values are converted into discrete or categorical variables if necessary. The results showed that SG-PSO exhibited the lowest loss value of all methods and that, in all algorithms, optimization process led to early convergence in about 15 generations.