Analysis of Hyperparameters for Workout Movements Classification Using the Convolutional Neural Network Algorithm

After the Covid-19 pandemic hit Indonesia. Many people are starting to be aware of their own health. However, they do not understand about correct and safe workout movements. Therefore, we want to classify workout movements with a machine learning model using the Convolutional Neural Network algorithm and compare multiple hyperparameter combinations to achieve optimal outcomes, allowing developers to deploy it on websites and applications. By evaluating all hyperparameters, this research aims to optimize model performance in the domain of workout movement classification. In this study there are 22 classes containing workout movements that focus on weightlifting exercises. The dataset was obtained from a downloaded YouTube video that the author subsequently divided into multiple frames. The dataset was splitted into 90% training and 10% validation. To achieve optimal results, the authors examine the impact of batch size, learning rate, and optimizer on the accuracy parameter of the model. The dataset will be trained using InceptionV3 CNN model architecture. The result shows adaptive optimizers like Adam, RMSprop, and AdamW outperformed basic optimizers like SGD and AdaGrad, suggesting that dynamically adjusting learning rates based on gradient information is more effective. A learning rate of 0.001 paired with a high batch size yielded superior results compared to other hyperparameter combinations when training the InceptionV3 model on this dataset. We hope application developers can utilize the study’s findings to improve the performance of fitness apps that involve workout movement classification.