Installing AutoGluon

Install specific AutoGluon modules and dependencies

AutoGluon is modularized into sub-modules specialized for tabular, multimodal, or time series data. You can reduce the number of dependencies required by solely installing a specific sub-module via: pip install <submodule>, where <submodule> may be one of the following options:

autogluon.tabular - functionality for tabular data (TabularPredictor)
- The default installation of autogluon.tabular standalone is a skeleton installation.
- Install via pip install autogluon.tabular[all] to get the same installation of tabular as via pip install autogluon
- Available optional dependencies: lightgbm,catboost,xgboost,fastai,ray. These are included in all.
- Optional dependencies not included in all: tabpfn,vowpalwabbit,imodels,skex,skl2onnx.
- To run autogluon.tabular with only the optional LightGBM and CatBoost models for example, you can do: pip install autogluon.tabular[lightgbm,catboost]
- Optional dependency: skex. This will speedup KNN models by 25x in training and inference on CPU. Use pip install autogluon.tabular[all,skex] to enable. Note: Not compatible with ARM processors.
- Experimental optional dependency: tabpfn. This will enable the usage of the TabPFN model, and allow you to fit TabPFN in TabularPredictor. Use pip install autogluon.tabular[all,tabpfn] to enable.
- Optional dependency: vowpalwabbit. This will install the VowpalWabbit package and allow you to fit VowpalWabbit in TabularPredictor.
- Experimental optional dependency: imodels. This will install the imodels package and allow you to fit interpretable models in TabularPredictor.
- Optional dependency: skl2onnx. This will enable ONNX model compilation via predictor.compile() on supported models.
autogluon.multimodal - functionality for image, text, and multimodal problems. Focus on deep learning models.
- To try object detection functionality using MultiModalPredictor, please install additional dependencies via mim install "mmcv==2.1.0", pip install "mmdet==3.2.0" and pip install pycocotools. Note that Windows users should also install pycocotools by: pip install pycocotools-windows, but it only supports python 3.6/3.7/3.8.
autogluon.timeseries - only functionality for time series data (TimeSeriesPredictor).
autogluon.common - helper functionality. Not useful standalone.
autogluon.core - only core functionality (Searcher/Scheduler) useful for hyperparameter tuning of arbitrary code/models.
autogluon.features - only functionality for feature generation / feature preprocessing pipelines (primarily related to Tabular data).
autogluon.eda - (Deprecated) only functionality for exploratory data analysis.

To install a submodule from source, follow the instructions for installing the entire package from source but replace the line cd autogluon && ./full_install.sh with cd autogluon && pip install -e {SUBMODULE_NAME}/{OPTIONAL_DEPENDENCIES}

For example, to install autogluon.tabular[lightgbm,catboost] from source, the command would be: cd autogluon && pip install -e tabular/[lightgbm,catboost]

To install all AutoGluon optional dependencies:

pip install autogluon && pip install autogluon.tabular[tabpfn,vowpalwabbit,imodels,skex,skl2onnx]

AutoGluon in Amazon SageMaker Studio

Amazon SageMaker Distribution is the docker environment for data science used as the default image of JupyterLab notebook instances and Code Editor in Amazon SageMaker Studio. The AutoGluon library comes pre-installed in all releases of Amazon SageMaker Distribution. SageMaker Studio users can access AutoGluon’s automation capabilities without needing to install anything additional.

To find the AutoGluon and PyTorch versions available in a SageMaker Distribution image, refer to the RELEASE.md file for your image version in the SageMaker Distribution GitHub repository.

Install from source for a specific pull-request

To build AutoGluon from source for the purposes of testing a pull-request, you can clone and install the exact branch by following these instructions. This process is useful if you are a code reviewer or want to test if a PR fixes a bug you have reported.

In this example, we are using this PR. It is from the user innixma and the PR branch is called accel_preprocess_bool. This information is provided in the PR page directly under the title of the PR (where it says into autogluon:master from Innixma:accel_preprocess_bool).

# Edit these two variables to change which PR / branch is being installed
GITHUB_USER=innixma
BRANCH=accel_preprocess_bool

pip install -U pip
git clone --depth 1 --single-branch --branch ${BRANCH} --recurse-submodules https://github.com/${GITHUB_USER}/autogluon.git
cd autogluon && ./full_install.sh

Note that the above example is only valid while the branch still exists. A user could delete the branch after the PR is merged, so this advice is primarily focused on unmerged PRs.

Install nightly builds

AutoGluon offers nightly builds that can be installed using the --pre argument. Nightly builds have the latest features but have not been as rigorously tested as stable releases.

pip install -U uv
python -m uv pip install --pre autogluon

M1 and M2 Apple Silicon

Apple Silicon is now supported via the conda installation instructions outlined above. conda-forge will install the GPU version if a user’s machine supports it.

Kaggle

AutoGluon is actively used by the Kaggle community. You can find hundreds of Kaggle notebooks using AutoGluon here.

For Kaggle competitions that allow internet access in notebooks, you can install AutoGluon via the following line at the start of the notebook:

!pip install -U autogluon > /dev/null

For competitions without internet access, you can obtain AutoGluon by using one of the Kaggle community’s packaged AutoGluon artifacts in the form of a Kaggle dataset.

If you encounter issues after installing AutoGluon, try restarting the notebook runtime to ensure a clean memory state.

Installing AutoGluon¶