Caketool

A Python MLOps toolkit for common machine learning and data science workflows. Provides EDA, feature engineering, model training, explainability, experiment tracking, model monitoring, calibration, and metrics — all designed for production-ready ML pipelines.

Installation

# Core (pandas, numpy, sklearn, xgboost)
pip install caketool

# With MLflow support
pip install "caketool[onprem]"

# With Google Cloud (Vertex AI, BigQuery)
pip install "caketool[gcp]"

# Everything
pip install "caketool[all]"

Quick Start

EDA

from caketool import eda

# Dataset overview
eda.profile(df)
eda.plot_correlations(df)

# Univariate
eda.plot_numeric_distribution(df["age"])
eda.plot_categorical_frequency(df["category"])

# Bivariate
eda.plot_scatter(df, x="income", y="spend", color_by="segment")
eda.plot_distribution_by_group(df, cat_col="segment", num_col="income", mode="box")

Feature Generation

from caketool.feature import generate_features_by_window

result = generate_features_by_window(
    df,
    client_id_col="user_id",
    report_date_col="event_date",
    fs_event_timestamp="snapshot_date",
    numeric_cols=("amount", "balance"),
    string_cols=("category",),
    boolean_cols=("is_active",),
    lookback_days=(0, 7, 30),   # 0 = lifetime
    backend="pandas",           # "pandas" | "polars" | "spark" | "bigframes"
)

Model Training

from caketool.model import BoostTree

model = BoostTree()
model.fit(X_train, y_train, eval_set=[(X_val, y_val)])
proba = model.predict_proba(X_test)[:, 1]
importance = model.get_feature_importance()

Out-of-Fold Cross-Validation

from caketool.model import BoostTree, EnsembleBoostTree

models, oof_preds, oof_labels = BoostTree.fit_oof(X_train, y_train, n_splits=5)
ensemble = EnsembleBoostTree(models)
proba = ensemble.predict_proba(X_test)[:, 1]

Explainability

from caketool.explainability import PermutationExplainer

explainer = PermutationExplainer(model=model)
explainer.fit(X_test)

# Global feature importance
importance = explainer.get_feature_importance()

# Local explanation for a single sample
local = explainer.get_local_explanation(row_index=0)

# Visualize
explainer.show_summary()
explainer.show_waterfall(row_index=0)
explainer.show_dependence(feature="income")

Score Calibration

from caketool.calibration import calibrate_score_to_normal

calibrated = calibrate_score_to_normal(raw_scores, standard=False)

Metrics

from caketool.metric import gini, psi

print(gini(y_true, y_pred))          # Gini coefficient
print(psi(expected, actual))          # Population Stability Index

Risk Report

from caketool.report import decribe_risk_score

report = decribe_risk_score(score_df, pred_col="score", label_col="label")

Drift Detection

from caketool.monitor import AdversarialModel

model = AdversarialModel()
model.fit(reference_df, current_df)
model.show()    # prints ROC AUC and top important features

Experiment Tracking

from caketool.experiment import create_tracker

# MLflow
with create_tracker("mlflow", experiment_name="my-exp", run_name="run-001") as tracker:
    tracker.log_params({"lr": 0.01, "depth": 6})
    tracker.log_metrics({"gini": 0.72})
    tracker.log_pickle(model, "model")

# Vertex AI
with create_tracker("vertex_ai", experiment_name="my-exp", run_name="run-001",
                    project="my-gcp-project", location="us-central1",
                    bucket_name="my-bucket") as tracker:
    tracker.log_params({"lr": 0.01})

API Overview

Development

conda create -n caketool python=3.10
conda activate caketool
pip-compile pyproject.toml --all-extras
pip install -e ".[dev,all]"
pre-commit install

Linting

Pre-commit hooks run ruff automatically on commit. To run manually:

ruff check src/ tests/ --fix  # Lint and auto-fix
ruff format src/ tests/        # Format code
pre-commit run --all-files     # Run all hooks

Tests

pytest tests/ -v --tb=short

Docs

pip install -e ".[docs]"
pdoc src/caketool   # Preview at http://localhost:8080

Docs are published automatically to GitHub Pages when a version tag is pushed.

Publishing

Version is automatically derived from git tags via setuptools-scm.

# Test on TestPyPI (RC/beta/alpha tags)
git tag v1.8.0-rc1
git push origin v1.8.0-rc1

# Publish to PyPI (stable tags)
git tag v1.8.0
git push origin v1.8.0

GitHub Actions builds and publishes automatically on tag push.

Local Development

python -m pip install -e .
python -c "from caketool import __version__; print(__version__)"