Amazon Web Services recently launched an open-source library that enables developers to implement deep learning models on image, text, or tabular data with just a few lines of code.
AutoGluon aims to be an easy-to-use and easily extensible AutoML toolkit for both machine learning beginners and experts. It enables prototyping of deep learning models with just a few lines; automatic hyperparameter tuning, model selection, and data processing; and automatic utilization of SOTA deep learning models. The AutoGluon framework can also help researchers customize and improve their existing custom models and data pipelines.
Traditionally, developers train deep learning models by manually defining neural networks and specifying hyperparameters during training. Researchers aiming to achieve SOTA performance must invest significant time deciding on the most efficient parameter updates to minimize errors, choosing the number of layers and how they should be connected, determining how best to classify and format the data, and so on.
AutoGluon aims to eliminate this tedious process and provide developers with a truly hands-free experience. After importing the AutoGluon package, developers only need to specify the task of interest, load the corresponding dataset, and finally let AutoGluon quickly and automatically train multiple models under thousands of different hyperparameter configurations, and then return the best model.
One of the applications of AutoGluon is object detection in images, achieved by object recognition and localization within a bounding box. Nick Erickson, author of AutoGluon Tabular, and his colleagues used AutoGluon to train an object detector on a small toy dataset generated using the motorcycle category of the VOC dataset. The task is to locate the motorcycle in a given picture. With one call to fit() and using predict() to test the model, AutoGluon was able to generate fairly accurate visualizations (shown below).
The AutoGluon project is on GitHub.