pythios 0.1.4

The oracle for your data — low-code ML for everyone.

Pythios

Low-code machine learning for tabular data.

Pythios loads data, prepares features, compares models, tunes the best candidate, and evaluates results through one simple workflow.

Installation

python -m pip install pythios

Pythios supports Python 3.9 and newer. On macOS, XGBoost and LightGBM may require brew install libomp.

Optional features:

python -m pip install "pythios[full]"

Quick start

from pythios import Data, Models, Evaluate

data = Data.from_csv("churn.csv", target="Churn")
model = Models(data)
model.auto(metric="f1")

evaluation = Evaluate(model)
evaluation.metrics()
evaluation.confusion_matrix()

Models.auto() automatically splits data, encodes categorical features, compares applicable models, selects the best model, tunes it with Optuna, and retrains it. Skip tuning for a faster run:

model.auto(metric="f1", tune=False)

Metrics

Classification metrics are accuracy, f1, recall, precision, and auc_roc. Regression metrics are r2, rmse, and mae.

model.auto(metric="recall")
model.auto(metric="auc_roc")

Defaults are accuracy for classification and r2 for regression. You can also use model.set_metric("f1").compare().

For example, a complete regression workflow using the built-in Diabetes dataset is:

from pythios import Data, Evaluate, Models

data = Data.from_sklearn("diabetes")
model = Models(data).auto(metric="r2", tune=False)
evaluation = Evaluate(model)
evaluation.metrics()
evaluation.residuals()

An R² score of 0.455 means the model explains approximately 45.5% of the variation in the test target. This is a moderate baseline: R² of 1.0 is perfect, 0.0 is equivalent to predicting the mean, and a negative value is worse than that baseline. Always consider R² together with MAE, RMSE, and cross-validation:

evaluation.cross_validate()

Loading data

from pythios import Data

data = Data.from_csv("data.csv", target="target")
excel = Data.from_excel("data.xlsx", target="target")
remote = Data.from_url("https://example.com/data.csv", target="target")
iris = Data.from_sklearn("iris")

Built-in datasets are iris, wine, breast_cancer, digits, diabetes, california_housing, and linnerud.

data.describe()
print(data.missing_report())
data.split(test=0.2, seed=42)

String targets such as Yes/No are automatically detected as binary classification.

Preprocessing

Models.auto() handles ordinary tabular preprocessing automatically. For manual control:

from pythios import Preprocess, Models

data.split()
processed = Preprocess(data).auto().apply()
model = Models(processed).auto(metric="f1")

Configuration methods are chainable:

processed = (
    Preprocess(data)
    .impute(method="median")
    .scale(method="standard")
    .encode(method="onehot")
    .apply()
)

The target is never transformed, and preprocessing is fitted on training data only.

Models

Available names include random_forest, logistic_regression, linear_regression, svm, knn, decision_tree, gradient_boosting, naive_bayes, ridge, lasso, xgboost, and lightgbm.

data.split()
model = Models(data)
model.train("random_forest", n_estimators=200)
results = model.compare()
model.tune("xgboost", trials=30)

Evaluation

predictions = model.predict()
probabilities = model.predict_proba()

evaluation = Evaluate(model)
evaluation.metrics()
evaluation.overfit_check()
evaluation.confusion_matrix()
evaluation.roc_curve()
evaluation.feature_importance()
evaluation.cross_validate()
evaluation.report(save="report.txt")

Visualization

from pythios import Visualize

visuals = Visualize(data)
visuals.distribution()
visuals.correlation()
visuals.missing()
visuals.pairplot()
visuals.summary()

Most visualization methods accept save="plot.png".

Unsupervised learning

from pythios import Unsupervised

u = Unsupervised(data)
u.cluster(method="kmeans", k=4)
u.reduce(method="pca", n=2)
u.plot_clusters()
print(u.silhouette_score())

Supported clustering methods include KMeans, DBSCAN, agglomerative clustering, and Gaussian mixtures. Reduction methods include PCA, t-SNE, and UMAP.

Embeddings

from pythios import Embeddings

embeddings = Embeddings().text()
vectors = embeddings.encode(["first document", "second document"])
print(embeddings.similarity(vectors[0], vectors[1]))

Faiss index search is available with the full extra.

Tools

from pythios import Tools

Tools.profile(data)
Tools.suggest(data)
Tools.validate(data)
Tools.detect_outliers(data)
Tools.export(data, "cleaned.csv")

Other utilities include balancing, memory reporting, reproducible seeds, dataset comparison, and timing. LLM helpers use a locally running Ollama server.

Saving models

model.save("model.pkl")
restored = Models(data).load("model.pkl")
predictions = restored.predict()

Development

python -m venv .venv
source .venv/bin/activate
python -m pip install -e .
python -m pip install pytest
pytest -q

License

Pythios is released under the MIT License.