highFIS

highFIS is a PyTorch-based library for high-dimensional Takagi–Sugeno–Kang (TSK) fuzzy systems. It brings differentiable fuzzy inference, numerical stability, and scikit-learn compatible estimators to both classification and regression.

🚀 Overview

Differentiable TSK fuzzy systems built for high-dimensional data.
Thirteen model families from the literature behind one API, each with a classifier and a regressor.
Available as scikit-learn compatible estimators and as plain PyTorch model classes.
Designed for numerical stability with log-space and inverse-log defuzzifiers.

📦 Installation

Install from PyPI:

pip install highfis

highFIS requires Python 3.11 or newer and depends on PyTorch, NumPy, scikit-learn, and tqdm. The diagnostic plots need matplotlib, an optional dependency:

pip install highfis[plot]

🧠 Quick Start

from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import MinMaxScaler

from highfis import HTSKClassifier

X, y = make_classification(n_samples=800, n_features=10, n_informative=8, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=42)

# Membership functions are sensitive to feature scale.
scaler = MinMaxScaler()
X_train = scaler.fit_transform(X_train)
X_test = scaler.transform(X_test)

clf = HTSKClassifier(
    n_mfs=3,
    mf_init="kmeans",
    epochs=100,
    learning_rate=0.01,
    random_state=42,
)
clf.fit(X_train, y_train)
print(f"Test accuracy: {clf.score(X_test, y_test):.2%}")

highFIS integrates with sklearn.pipeline.Pipeline, GridSearchCV, and cross_val_score.

🧩 Model families

highFIS implements thirteen TSK model families, each following a published high-dimensional inference strategy.

TSK — generic TSK; by default the vanilla system with product antecedent aggregation and sum-based normalization, with selectable membership function, T-norm, and defuzzifier.
HTSK — geometric mean aggregation with log-space softmax normalization.
LogTSK — inverse-log normalization of log-domain rule weights.
HDFIS — high-dimensional inference with product T-norm (HDFISProd) and minimum T-norm (HDFISMin) variants.
DombiTSK — Dombi T-norm aggregation with a fixed shape parameter.
ADMTSK — adaptive Dombi T-norm, with the shape parameter derived from the input dimensionality, and composite Gaussian membership functions.
AYATSK — adaptive Yager T-norm with composite exponential membership functions.
ADATSK — adaptive softmin antecedent aggregation.
ADPTSK — adaptive double-parameter softmin antecedent aggregation.
FSRE-ADATSK — ADATSK extended with embedded feature selection and rule extraction.
DGTSK — double groups of gates for simultaneous feature selection and rule extraction.
DGALETSK — adaptive Ln-Exp softmin with simultaneous feature selection and rule extraction.
MHTSK — multihead TSK built from sparse subantecedents over random feature subsets.

Each family exposes classifier and regressor variants.

🔧 Core components

highFIS exposes a compact, model-family-driven API with both concrete model classes and sklearn-compatible estimator wrappers.

Estimators: *Classifier and *Regressor variants for each model family, importable directly from highfis (for example HTSKClassifier, FSREADATSKRegressor, HDFISProdClassifier)
PyTorch models: the underlying *ClassifierModel and *RegressorModel classes in highfis.models
Building blocks: membership functions (highfis.memberships), defuzzifiers (highfis.defuzzifiers), T-norms (highfis.t_norms), and layers (highfis.layers)
Interpretation: estimator.rules_as_text() writes the rule base as IF ... THEN ... sentences; the rule table, consequent coefficients, and gate values are available as estimator methods
Diagnostic plots: estimator.plot(kind=...) for the learned membership functions, the training history, the rule activations, and the predictions
Utilities: evaluation metrics (highfis.metrics), estimator checkpoints (highfis.persistence), and a membership-function initialization cache

For the full class list and API reference, see the documentation:

🛠️ Training options

highFIS uses gradient-based optimization and supports:

an optimizer selected per model family (SGD, Adam, or AdamW), following the source paper
mini-batch training with learning-rate schedulers and weight decay on the consequent parameters
early stopping on a validation set passed to fit
uniform regularization (ur_weight) for balanced rule activation
a choice of membership function, T-norm, and defuzzifier in the generic TSKClassifier and TSKRegressor; the other families keep their published combination
custom T-norms, rule bases, and defuzzifiers through the PyTorch model classes

📚 Documentation

The published documentation is available at:

https://dcruzf.github.io/highFIS

Start with the model families overview, the user guides, and the cookbook. The diagnostic plots guide shows how to inspect a fitted model.

Model reference pages:

🧪 Testing & quality

Format, lint, and type check:

hatch check --fix

Run the test suite with coverage:

hatch test -c -a

Run security scan:

hatch run security

📝 Citation

If you use highFIS in your research, please cite it using the metadata in CITATION.cff or the Zenodo DOI.

🤝 Contributing

Contributions are welcome! Please open issues or pull requests, and refer to our development guide in the documentation: contributing.

When reporting a bug, include the output of highfis.show_versions().

📄 License

Distributed under the GPLv3.

Metadata

Release files for highfis 0.33.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for highfis 0.33.0
File	Size	Uploaded
highfis-0.33.0.tar.gz	239.8 kB	Oct 10, 2026	Details

Built distribution (wheel)

Table of built distributions (wheels) for highfis 0.33.0
File	Interpreter	ABI	Platform	Reset
highfis-0.33.0-py3-none-any.whl 203.9 kB Oct 10, 2026	Python 3	none	any	Details

Total release size: 443.7 kB

Release files / highfis-0.33.0.tar.gz

Download URL	highfis-0.33.0.tar.gz
Size	239.8 kB
Tags	Source
SHA-256 checksum How to use checksums	`dadb9fa09117fb363d5a8b5f44cc5fafd9b5fd2320a2d1ffe7e161eb95ea0197`
BLAKE2b-256 checksum How to use checksums	`d1cdd9fd4cc62b03e61f3aced612e699d3456ab3e24a74e4f140713c30b97fe4`
Upload date	Oct 10, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	Yes
Uploaded via	`uv/0.13.0 {"installer":{"name":"uv","version":"0.13.0","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"24.04","id":"noble","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}`

Release files / highfis-0.33.0-py3-none-any.whl

Download URL	highfis-0.33.0-py3-none-any.whl
Size	203.9 kB
Tags	Python 3
SHA-256 checksum How to use checksums	`3c800702c3b1cd40f6525ef4b37f70b6186030b83028a26badf5cdc227b08da1`
BLAKE2b-256 checksum How to use checksums	`55662468d9d7439911d8725c9fec3a97d7029e2ac1a07ceb7d9d7726c39906c1`
Upload date	Oct 10, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	Yes
Uploaded via	`uv/0.13.0 {"installer":{"name":"uv","version":"0.13.0","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"24.04","id":"noble","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}`