bizhani-linreg 1.0.0

Linear Regression from scratch — no ML libraries required.

linreg

Linear Regression from scratch — no ML libraries required.
Pure Python · Batch Gradient Descent · Z-score normalisation · Early stopping

---

Installation

# Editable (development) install — changes to the source are reflected immediately
pip install -e .

# Or a regular install
pip install .

linreg has a single runtime dependency: matplotlib (for plotting).
The model, scaler, metrics, and data utilities use only the Python standard library.

---

Quick start

from linreg import (
    LinearRegression,
    ZScoreScaler,
    train_test_split,
    r2_score, rmse, mae,
    plot_all,
)

# 1. Prepare data  (list of lists — no NumPy needed)
X = [[x1, x2], ...]   # shape (n_samples, n_features)
y = [y1, y2, ...]     # shape (n_samples,)

# 2. Split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

# 3. Z-score normalise
x_scaler  = ZScoreScaler()
X_train_s = x_scaler.fit_transform(X_train)   # fit + transform on train
X_test_s  = x_scaler.transform(X_test)        # transform only on test

y_scaler  = ZScoreScaler()
y_train_s = y_scaler.fit_transform_1d(y_train)
y_test_s  = y_scaler.transform_1d(y_test)

# 4. Train
model = LinearRegression(lr=0.01, max_iter=10_000, tol=1e-6)
model.fit(X_train_s, y_train_s)

# 5. Predict  (invert target scaling to recover original units)
y_pred_s = model.predict(X_test_s)
y_pred   = y_scaler.inverse_transform_1d(y_pred_s)

# 6. Evaluate
print(f"R²   = {r2_score(y_test, y_pred):.4f}")
print(f"RMSE = {rmse(y_test, y_pred):.4f}")
print(f"MAE  = {mae(y_test, y_pred):.4f}")

# 7. Plot
fig = plot_all(model.loss_history, model.n_iter_, y_test, y_pred)
fig.savefig("results.png", dpi=150)

Run the bundled demo:

python main.py

---

API reference

LinearRegression

LinearRegression(lr=0.01, max_iter=10_000, tol=1e-6, verbose=True)

Method	Description
fit(X, y)	Train the model; returns self for chaining
predict(X)	Return predicted values
score(X, y)	Return R² on the given data
summary()	Print fitted parameters and training stats

Key attributes after fit()

Attribute	Description
weights	Fitted weight vector [w₁, …, wₙ]
bias	Fitted intercept b
loss_history	MSE at every iteration
n_iter_	Actual iterations run (early stopping may fire early)

---

ZScoreScaler

ZScoreScaler()

Method	Description
fit_transform(X)	Fit on X, return scaled matrix
transform(X)	Apply existing fit to new data
fit_transform_1d(y)	Fit on target vector, return scaled list
transform_1d(y)	Apply existing fit to new target list
inverse_transform_1d(y_scaled)	Recover original units

---

Metrics

from linreg import mse, rmse, mae, r2_score

All functions accept two plain Python lists: (y_true, y_pred).

---

Data utilities

from linreg import train_test_split, generate_dataset

Function	Description
train_test_split(X, y, test_size, seed)	Reproducible random split
generate_dataset(n_samples, noise, seed)	Synthetic y = 3x₁ + 5x₂ + 10 + ε

---

Plotting

from linreg import plot_loss, plot_predictions, plot_residuals, plot_all

All functions return a matplotlib.Figure — they never call plt.show(), so you control display, saving, or embedding.

Function	Description
plot_loss(loss_history, n_iter)	MSE vs iteration with early-stop marker
plot_predictions(y_true, y_pred)	Predicted vs actual scatter
plot_residuals(y_true, y_pred)	Residuals vs predicted scatter
plot_all(loss_history, n_iter, y_true, y_pred)	All three side-by-side

---

Algorithm

Model:       ŷ = b + w₁x₁ + w₂x₂ + … + wₙxₙ

Loss (MSE):  L = (1/n) Σ (ŷᵢ − yᵢ)²

Gradients:   ∂L/∂wⱼ = (1/n) Σ (ŷᵢ − yᵢ) · xᵢⱼ
             ∂L/∂b  = (1/n) Σ (ŷᵢ − yᵢ)

Update:      wⱼ ← wⱼ − lr · ∂L/∂wⱼ
             b  ← b  − lr · ∂L/∂b

Stop when:   |MSE(t) − MSE(t−1)| < tol

---

Project layout

linreg/
├── __init__.py     public API surface
├── _math.py        dot, list_mean, list_std, transpose
├── _scaler.py      ZScoreScaler
├── _model.py       LinearRegression
├── _metrics.py     mse, rmse, mae, r2_score
├── _data.py        train_test_split, generate_dataset
└── plot.py         plot_loss, plot_predictions, plot_residuals, plot_all
main.py             runnable demo
setup.py            packaging
README.md           this file

---

License

MIT