Thread by @smolix

#AutoGluon 0.4 is officially released. It's a big update with better performance (up to 90% win rate), better speed (up to 2x faster), removed dependencies, more compatibility (full Windows support), and more models. 151 commits from 14 contributors. auto.gluon.ai 🧵 1/

To get started on #Windows follow the instructions at auto.gluon.ai/0.4.0/index.ht… - TL;DR 1) Install Anaconda; 2) Install Python 3.9, pip and wheel; 3) pip install PyTorch 1.10.1; 4) pip install autogluon. 2/

Text support got a massive upgrade. It now supports @huggingface transformer backbones and uses @PyTorchLightnin. Relative to the 0.3 release arxiv.org/abs/2111.02705, the 0.4 release has an over 70% win rate. We support multiple languages and transfer learning. 3/

You can now win benchmarks with 3 lines of code in @autogluon. Check out machinehack.com/hackathon/prod… machinehack.com/hackathon/pred… and kaggle.com/c/mercari-pric… for a first, second and third prize respectively. 4/

This release contains 151 commits from 14 contributors, 6 of which are first time contributors. github.com/awslabs/autogl… 5/

#Tabulardata in @autogluon has seen significant speedups, for training (2x) and inference (1.3x), while offering an over 60% win rate relative to the 0.3.1 release. Much better for large data. Improved quantile regression, conformal calibration, GPUs auto.gluon.ai/0.4.0/tutorial… 6/

There are even more base models for #Tabulardata, including Vowpal Wabbit @JohnCLangford and NN_Torch @PyTorch, besides XGBoost @tqchenml, CatBoost @CatBoostML, LightGBM, FastAIModel @jeremyphoward and Linear Models @sklearn_inria 7/

AutoGluon no longer depends on ConfigSpace, cython, dill, paramiko, autograd, openml, d8, and graphviz. MXNet is no longer needed by default. All of this simplifies installation, in particular on Windows 7/

To get started, check out this cheat sheet auto.gluon.ai/stable/autoglu… 8/