Neural networks for tabular data, with automatic preprocessing and a scikit-learn-style API.
NeuroTabular uses PyTorch to train neural models for binary classification, multiclass classification, and single-output regression on pandas DataFrames. It handles numerical and categorical columns, missing values, and reusable training-only preprocessing inside the estimator.
Train neural networks on pandas DataFrames without building a custom PyTorch preprocessing pipeline or training loop. NeuroTabular handles mixed features, missing values, validation, and early stopping behind a scikit-learn-style API.
A young library under active development and evaluation. Core classification and regression functionality is implemented and extensively tested, and the public API follows a documented stability policy. Hardware coverage and performance across different workloads still need broader validation.
For neural learning on mixed pandas data when you want a familiar estimator interface and integration with scikit-learn workflows.
Consider other tools if you need multiple neural architectures, multi-output learning, advanced AutoML, or more extensively benchmarked predictive performance.
python -m pip install neurotabularPython 3.11–3.14 is the recommended range for standard installation. See the support scope for tested environments and other runtime combinations.
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
from neurotabular import NeuroTabularClassifier
X, y = load_breast_cancer(return_X_y=True, as_frame=True)
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.2, random_state=42, stratify=y
)
model = NeuroTabularClassifier(random_state=42, max_epochs=10)
model.fit(X_train, y_train)
print("Test accuracy:", model.score(X_test, y_test))The same classifier handles multiclass targets. For regression, use
NeuroTabularRegressor. For numerical and categorical features, missing
values, validation sets, and all three tasks, see the
complete mixed-data example.
- Automatic preprocessing of numerical and categorical pandas columns, including missing values.
- Binary and multiclass classification, plus single-output regression.
- scikit-learn-style estimators, cloning, pipelines, and cross-validation.
- Validation, early stopping, batching, and checkpoint restoration.
sample_weightfor all tasks and optionalclass_weight="balanced"for classification.- CPU execution and optional CUDA selection; pickle/joblib model persistence.
- Input must be a pandas DataFrame; multi-output and multilabel tasks are not supported.
- No claim of universal accuracy or speed advantages over established tabular models.
- Physical NVIDIA GPU validation is incomplete; some hardware environments remain outside the verified scope.
- Very high-cardinality categorical features can increase memory use substantially.
- Concurrent
fit()calls within a process are not guaranteed thread-safe.
See the benchmark evidence for measured results and dataset-specific failure cases.
- Usage and examples
- API reference
- Support scope
- API stability and maintenance policies
- Benchmarks
- Performance measurements
- Release notes
NeuroTabular is created by Luca Lullo and released under the MIT License. Dependency notices are in THIRD_PARTY_NOTICES.md. See LEGAL.md for the project's legal/naming note.
Created by Luca Lullo.