Working on a system which allows everyone to easily train a computer vision neural network, our ML team has created an app to check which ML model works the best on a given mobile app. Before developing their own ML-based app for image recognition and classification, users need to choose a machine learning model to train. The choice is heavily influenced by the mobile device on which the app will run as more advanced models work faster on newer devices and slower on older ones. To help users find the model which will work the fastest on their mobile device, we have developed a simple but effective model testing app.
We have adapted four image recognition and classification TensorFlow Lite models and three PyTorch models to be used on an Android device. Users can choose between two image sizes to better reflect real working conditions, i.e. the size of images in a real dataset. After that the app tests each model and scores them according to their performance.
The app allows users to easily make the right choice between various TensorFlow Lite and PyTorch models in a matter of minutes, ensuring their project will be as effective as possible.