AutoGluon

1.5.0 · active · verified Sun Apr 12

AutoGluon is an open-source AutoML library developed by AWS AI, designed for fast and accurate machine learning with minimal code. It automates model selection, hyperparameter tuning, and ensemble creation across various data types, including tabular, time series, and multimodal data. The current version is 1.5.0, with a rapid release cadence, often introducing significant performance improvements, new models, and expanded functionalities.

Warnings

breaking Models trained with an older version of AutoGluon are not guaranteed to be compatible with newer versions. Users must re-train models after upgrading AutoGluon to prevent issues.
Fix: Re-train all AutoGluon models using the new AutoGluon version.
breaking Python 3.8 support was dropped in AutoGluon v1.2.0. Additionally, Python 3.13 support is currently experimental in v1.5.0, with potential limitations on Windows.
Fix: Ensure your Python environment is between 3.10 and 3.12 (inclusive) for stable usage. If using Python 3.13, be aware of potential experimental features or platform-specific issues.
breaking Several `TabularPredictor` methods were deprecated in v1.0.0 (e.g., `persist_models`, `get_model_names`), began raising errors in v1.2.0, and were completely removed in v1.3.0.
Fix: Update method calls to their new names: `persist_models` -> `persist`, `unpersist_models` -> `unpersist`, `get_model_names` -> `model_names`, `get_model_best` -> `model_best`, `get_pred_from_proba` -> `predict_from_proba`, etc. Refer to the AutoGluon v1.3.0 release notes for a complete list.
deprecated The default behavior of `TabularPredictor.delete_models()` will change from `dry_run=True` to `dry_run=False` in a future release. A `FutureWarning` is logged since v1.3.0.
Fix: Explicitly specify the `dry_run` argument (e.g., `predictor.delete_models(dry_run=True)`) to maintain current behavior or `predictor.delete_models(dry_run=False)` to enable actual deletion.
gotcha AutoGluon (especially `TabularPredictor`) can be memory-intensive, particularly with large datasets or complex ensembles. Out-of-memory errors are common.
Fix: Use machines with ample RAM. For tabular data, GPUs are often not critical, but more memory is. Consider reducing `presets` quality, `time_limit`, or subsampling data for initial runs. Set `max_memory` parameter in `predictor.persist()` carefully.
gotcha On macOS, LightGBM and XGBoost (used by AutoGluon) can experience segmentation faults or instability if `libomp` is installed via `brew install libomp`.
Fix: If experiencing issues, uninstall `libomp` via brew and follow AutoGluon's specific instructions for installing a compatible version of `libomp`.

Install

pip install autogluon Basic Installation
pip install autogluon[all] Full Installation (includes common optional dependencies)
pip install autogluon.tabular[tabarena] Tabular with State-of-the-Art Models (e.g., TabPFN, TabDPT)

Imports

TabularPredictor, TabularDataset

from autogluon.tabular import TabularPredictor, TabularDataset

TimeSeriesPredictor, TimeSeriesDataFrame

from autogluon.timeseries import TimeSeriesPredictor, TimeSeriesDataFrame

MultiModalPredictor

from autogluon.multimodal import MultiModalPredictor

Quickstart

This quickstart demonstrates how to use AutoGluon's `TabularPredictor` to train a high-accuracy classification model and make predictions on new data with just a few lines of code. It automatically handles feature engineering, model selection, and hyperparameter tuning.

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

# Load example data (using a public dataset URL)
data_url = 'https://raw.githubusercontent.com/mli/ag-docs/main/knot_theory/'
train_data = TabularDataset(f'{data_url}train.csv')
test_data = TabularDataset(f'{data_url}test.csv')

label = 'signature'
predictor = TabularPredictor(label=label, path='AutogluonModels').fit(train_data, presets='high_quality')
predictions = predictor.predict(test_data)

print("Top 5 predictions:")
print(predictions.head())
print(f"Predictor leaderboards:\n{predictor.leaderboard(test_data, silent=True)}")

view raw JSON →