AutoGluon Tabular

1.5.0 · active · verified Sun Apr 12

AutoGluon Tabular provides a fast and accurate AutoML library specifically designed for tabular data, allowing users to train and deploy high-accuracy machine learning models with just a few lines of code. Developed by AWS AI, it offers automated stack ensembling, deep learning integration, and handles feature engineering and hyperparameter tuning automatically. The library maintains an active release cadence with major updates every few months and intermediate patch releases.

Warnings

breaking Models trained with an older version of AutoGluon are NOT compatible with newer versions. Users must re-train models after upgrading the library.
Fix: Re-train all AutoGluon models after upgrading the `autogluon-tabular` library.
breaking Several `TabularPredictor` methods were renamed in v1.3.0. For example, `persist_models` became `persist`, and `get_model_names` became `model_names`.
Fix: Update method calls to their new names (e.g., `predictor.persist()` instead of `predictor.persist_models()`). Refer to the v1.3.0 release notes or documentation for a full list of renames.
gotcha Python version support changes: AutoGluon dropped support for Python 3.8 in v1.2.0, while adding support for 3.12. Current versions (1.5.0) support Python 3.10-3.13, with experimental support for 3.13 on Windows.
Fix: Ensure your Python environment uses a supported version (3.10-3.13 for v1.5.0). Check release notes for specific version compatibility.
gotcha If using 'TABPFNV2' model, it is strongly recommended to switch to 'REALTABPFN-V2' due to breaking changes in the underlying TabPFN library. 'REALTABPFN-V2.5' also exists but has a non-commercial license and requires HuggingFace authentication.
Fix: Update your model hyperparameters to use 'REALTABPFN-V2' if you were using 'TABPFNV2'. For 'REALTABPFN-V2.5', ensure you understand the licensing and perform the required HuggingFace authentication. Note that 'REALTABPFN-V2.5' is not included in default presets.

Install

pip install autogluon.tabular[all] Full installation with common ML dependencies
pip install autogluon.tabular Minimal (skeleton) installation

Imports

TabularPredictor

from autogluon.tabular import TabularPredictor

TabularDataset

from autogluon.tabular import TabularDataset

Quickstart

This quickstart demonstrates how to train a `TabularPredictor` on a dataset and then make predictions. It uses a dummy DataFrame for immediate execution, with commented-out lines showing how to load data from AutoGluon's public S3 bucket for a more realistic scenario. The `presets='best'` option instructs AutoGluon to use its best-performing configuration.

import pandas as pd
from autogluon.tabular import TabularPredictor, TabularDataset

# Create dummy dataframes if not using S3 URLs for demonstration
train_data = pd.DataFrame({
    'feature_1': [1, 2, 3, 4, 5],
    'feature_2': ['A', 'B', 'A', 'C', 'B'],
    'target_column': [0, 1, 0, 1, 0]
})
test_data = pd.DataFrame({
    'feature_1': [6, 7],
    'feature_2': ['C', 'A']
})

# Or load directly from AutoGluon's S3 bucket (uncomment for real usage)
# data_root = 'https://autogluon.s3.amazonaws.com/datasets/Inc/'
# train_data = TabularDataset(data_root + 'train.csv')
# test_data = TabularDataset(data_root + 'test.csv')

# Initialize and train the predictor
predictor = TabularPredictor(label='target_column', path='./AutogluonModels').fit(train_data, presets='best')

# Make predictions
predictions = predictor.predict(test_data)
print("Predictions:\n", predictions)

# Evaluate the model (requires a label column in test_data, not present in dummy test_data)
# Assuming test_data_with_labels exists:
# test_data_with_labels = pd.DataFrame({
#     'feature_1': [6, 7],
#     'feature_2': ['C', 'A'],
#     'target_column': [1, 0]
# })
# leaderboards = predictor.leaderboard(test_data_with_labels, silent=True)
# print("Leaderboard:\n", leaderboards)

view raw JSON →