salearn 1.0.0

From SQL to serving — in one import.

sklearn-compatible and fast (NumPy + optional Numba), with a built-in data stack: databases · loading · EDA · cleaning · features · time-series · NLP · AutoML · explainability · model registry.

Quickstart · Modules · Benchmarks · Roadmap · Contributing

Why salearn?

Formerly developed as sslearn. The package was renamed to salearn for its 1.0 launch — same API, bigger scope.

Install

pip install salearn
pip install "salearn[speed]"   # numba — JIT for hot loops
pip install "salearn[db]"      # sqlalchemy + duckdb + pandas + pyarrow
pip install "salearn[eda]"     # pandas + matplotlib
pip install "salearn[all]"     # everything

Requires Python 3.9+, numpy, scipy. Everything else is optional.

Quickstart

1. Drop-in sklearn replacement

# from sklearn.ensemble import RandomForestClassifier
from salearn.ensemble import RandomForestClassifier
from salearn.preprocessing import StandardScaler
from salearn.pipeline import make_pipeline

pipe = make_pipeline(StandardScaler(), RandomForestClassifier(n_estimators=200, n_jobs=-1))
pipe.fit(X_train, y_train)
print(pipe.score(X_test, y_test))

2. Databases — train straight from SQL

from salearn.db import Database

db = Database.sqlite_file("shop.db")
db.to_sql("sales", df=df, if_exists="replace")
X, y, feats = db.load_X_y("SELECT * FROM sales WHERE amount > 0", target="churn")
model = db.train_sql("SELECT * FROM sales", target="churn")

3. Load → profile → clean in three lines

from salearn import io, eda
from salearn.clean import DataCleaner

df = io.load_csv("data.csv")                       # or load_parquet / load_url(...)
eda.profile(df, html_path="report.html")           # shape, missing, outliers, duplicates
df = DataCleaner(outlier_clip="iqr").fit_transform(df)

4. Features → AutoML → explain → ship

from salearn.features import auto_featurize
from salearn.automl import find_best
from salearn.explain import model_report
from salearn.persist import save_model

X, y, names = auto_featurize(df, target="churn", date_cols=["signup_date"])
best, board = find_best(X, y, time_budget=15)
print(board[0], model_report(best, X, y, names))
save_model(best, "churn.ssl")

5. Time-series (NumPy only) and multilingual NLP

from salearn.timeseries import ARForecaster, temporal_split
train, test = temporal_split(sales, test_size=30)
forecast = ARForecaster(lags=14).fit(train).predict(30)

from salearn.text import TextClassifier   # English + Arabic out of the box
clf = TextClassifier().fit(["great movie", "فيلم رائع", "boring", "ممل"], [1, 1, 0, 0])
clf.predict(["absolutely great", "رائع جدا"])

Module map

salearn/
├── linear_model tree ensemble neighbors naive_bayes svm cluster   # classic ML (sklearn API)
├── decomposition metrics model_selection pipeline preprocessing   # ...faster (BLAS / Numba / threads)
├── impute feature_selection mixture neural_network extras datasets calibration
├── db.py          # Database (sqlite/duckdb/postgres/mysql), QueryBuilder, train_sql
├── io.py          # csv/parquet/json/excel/url, chunked reading, to_X_y
├── eda.py         # profile / missing / outliers / correlation / HTML report
├── clean.py       # DataCleaner transformer
├── features.py    # DateFeaturizer, TargetEncoder, FrequencyEncoder, auto_featurize
├── timeseries.py  # Naive / Seasonal / MA / ETS / AR + walk-forward CV
├── text.py        # clean_text, Count/TfidfVectorizer, TextClassifier
├── automl.py      # find_best + leaderboard
├── explain.py     # permutation importance, partial dependence, model_report
├── persist.py     # save_model / load_model / ModelRegistry
└── viz.py         # confusion / residuals / importance (matplotlib optional)

flowchart LR
    DB[("SQL databases")] --> IO["salearn.io"]
    IO --> EDA["salearn.eda"]
    EDA --> CLEAN["clean + features"]
    CLEAN --> ML["classic ML"]
    CLEAN --> AUTO["salearn.automl"]
    ML --> EXP["salearn.explain"]
    AUTO --> EXP
    EXP --> REG[("model registry")]

Benchmarks

python bench/bench.py
python -m pytest -q        # 12 tests, ~20s
python examples_fullstack.py

Vectorized NumPy paths (BLAS pairwise distances, Cholesky Ridge, lstsq), optional Numba JIT for KMeans/KNN/SGD/Lasso-CD/Pegasos, and ThreadPool-parallel forests.

Roadmap

• PyPI release + docs site
• Deep-learning bridge (torch/onnx export)
• Streaming / incremental estimators
• More connectors (BigQuery, Snowflake via SQLAlchemy)
• Multilingual docs (Arabic, French)

Vote with a ⭐ or open an issue — it shapes priorities.

Contributing

See CONTRIBUTING.md. Bug reports and PRs welcome.

License

MIT — see LICENSE. Changelog in CHANGELOG.md.