HyperFast : Instant Classification for Tabular Data

[Update Nov 26, 2025] Interested in more accurate tabular data classification and support for regression? See our new model iLTM: Integrated Large Tabular Model (arXiv).

HyperFast is a hypernetwork designed for fast classification of tabular data, capable of scaling to large datasets. Utilizing a meta-trained hypernetwork, HyperFast generates a dataset-specific target network in a single forward pass, eliminating the need for time-consuming model training.

Installation

HyperFast can be installed from PyPI using

pip install hyperfast

This package contains example code to run HyperFast. Ensure that you are using Python 3.9 or later to run this project. When using HyperFast for the first time, the model weights will be downloaded automatically. The model weights are also available here.

Usage

Here's a quick example on how to use HyperFast's scikit-learn-like interface:

import torch
import numpy as np
from hyperfast import HyperFastClassifier
from sklearn.metrics import accuracy_score

# Load your dataset
X_train, y_train = np.load("data/hapmap1_X_train.npy"), np.load( "data/hapmap1_y_train.npy")
X_test, y_test = np.load("data/hapmap1_X_test.npy"), np.load("data/hapmap1_y_test.npy")

# Set the device
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')

# Initialize HyperFast
model = HyperFastClassifier(device=device)

# Generate a target network and make predictions
model.fit(X_train, y_train)
predictions = model.predict(X_test)

accuracy = accuracy_score(y_test, predictions)
print(f"Accuracy: {accuracy * 100:.2f}%")

Warning:

• Set cat_features as the list of indices of the categorical features in the dataset (if any), now in __init__.
• The current default hyperparameters provide medium speed-accuracy performance.
• 🚀 For the fastest inference (but less accurate) set n_ensemble=1 and optimization=None.
• 📊 If you are dealing with an imbalanced dataset, consider setting stratify_sampling=True with n_ensemble > 1.
• 🌐 If you are dealing with a very high-dimensional dataset (e.g., >3000 features), consider setting feature_bagging=True with n_ensemble > 1.
• 👌 For slower but most accurate predictions, optimize the inference parameters of HyperFast for each dataset. In this case, we recommend the following search space:

from scipy.stats import loguniform, uniform

param_grid = {
    'n_ensemble': [4, 8, 16, 32],
    'batch_size': [1024, 2048],
    'nn_bias': [True, False],
    'nn_bias_scale': loguniform(0.01, 1.0),  # Only when nn_bias=True.
    'stratify_sampling': [True, False],
    'optimization': ['ensemble_optimize', None],  # Sampling weights: 9:1.
    # Remaining parameters apply only when optimization is enabled.
    'optimize_steps': [1, 4, 8, 16, 32, 64, 128, None],
    'finetuning_lr': loguniform(1e-5, 3e-3),
    'finetuning_dropout': uniform(0.0, 0.5),
    'finetuning_weight_decay': loguniform(1e-6, 0.1),
    'finetuning_parameter_scope': ['all', 'network', 'head'],  # Sampling weights: 8:1:1.
    'finetuning_batch_size': [128, 256, 1024],
    'label_smoothing': [0.0, 0.05, 0.1],
}

Pass model.fit(X_train, y_train, eval_set=(X_val, y_val)) to select each ensemble member's best validation checkpoint, including its initial weights. Binary tasks use ROC-AUC and multiclass tasks use log loss. Validation rows are excluded from preprocessing fitting, weight generation, gradient updates, and nearest-neighbor references. Without an explicit validation set, fine-tuning reserves a stratified 10% holdout before preprocessing.

Controlling fine-tuning

finetuning_parameter_scope="all" retains full predictor adaptation. Use "network" to freeze the random projection and PCA, or "head" to also freeze all hidden MLP layers. Enabled nearest-neighbor biases remain trainable. These options do not update the pretrained hypernetwork or checkpoint.

License

This project is under the CC BY-NC 4.0 license. See LICENSE for details.

Cite us

If you use HyperFast in your research, please cite our paper:

@inproceedings{bonet2024hyperfast,
  title={HyperFast: Instant Classification for Tabular Data},
  author={Bonet, David and Montserrat, Daniel Mas and Gir{\'o}-i-Nieto, Xavier and Ioannidis, Alexander},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2024}
}