Case study of Hyperparameter Optimization framework Optuna on a Multi-column Convolutional Neural Network

To observe the condition of the flower growth during the blooming period and estimate the harvest forecast of the Canola crops, the ‘Flower Counter’ application has been developed by the researchers ofP2IRC at the University of Saskatchewan. The model has been developed using a Deep Learning based Multi-column Convolutional Neural Network (MCNN) algorithm and the TensorFlow framework, in order to count the Canola flowers from the images based on the learning from a given set of training images. To ensure better accuracy score with respect to flower prediction, proper training of the model is essential involving appropriate values of hyperparameters. Among numerous possible values of these hyperparameters, selecting the suitable ones is certainly a time-consuming and tedious task for humans. Ongoing research for developing Automated Hyperparameter Optimization (HPO) frameworks has attracted researchers and practitioners to develop and utilize such frameworks to give directions towards \nfinding better hyperparameters according to their applications.\nThe primary goal of this research work is to apply the Automated HPO Optuna on the Flower Counterapplication with the purpose of directing the researchers towards among the best observed hyperparameter configurations for good overall performance in terms of prediction accuracy and resource utilization. This work would help the researchers and plant scientists gain knowledge about the practicality of Optuna while treating it as a black-box and apply it for this application as well as other similar applications.\nIn order to achieve this goal, three essential hyperparameters, batch size, learning rate and number of epochs, have been chosen for assessing their individual and combined impacts. Since the training of the model depends on the datasets collected during diverse weather conditions, there could be factors that could impact Optuna’s functionality and performance. The analysis of the results of the current work and comparison of the accuracy scores with the previous work have yielded almost equal scores while testing the model’s performance on different test populations. Moreover, for the tuned version of the model, the current work has shown the potential for achieving that result with substantially lower resource utilization. The findings have provided useful concepts about making the better usage of Optuna; the search space can be restricted ormore complicated objective functions can be implemented to ensure better stability of the models obtained when chosen parameters are used in training

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.