lernkurve.png

def fig_lernkurve(X_train, y_train) -> None:
    """Plot MLP training loss curve and validation accuracy curve."""
    mlp_pipe = Pipeline([
        ("pre", build_preprocessor()),
        ("mlp", MLPClassifier(
            hidden_layer_sizes=(64, 32),
            activation="relu",
            alpha=0.01,
            early_stopping=True,
            validation_fraction=0.15,
            max_iter=300,
            random_state=SEED,
        )),
    ])
    mlp_pipe.fit(X_train, y_train)
    mlp = mlp_pipe.named_steps["mlp"]

    loss_curve = mlp.loss_curve_
    val_scores = mlp.validation_scores_
    epochs = range(1, len(loss_curve) + 1)

    fig, ax1 = plt.subplots(figsize=(7, 4))
    ax1.plot(epochs, loss_curve, color=PRIMARY, lw=1.8, label="Trainingsverlust")
    ax1.set_xlabel("Epoche")
    ax1.set_ylabel("Verlust (log loss)", color=PRIMARY)
    ax1.tick_params(axis="y", labelcolor=PRIMARY)

    ax2 = ax1.twinx()
    ax2.plot(epochs, val_scores, color=EVENT, lw=1.8, linestyle="--",
             label="Validierungsgüte (Genauigkeit)")
    ax2.set_ylabel("Validierungsgüte", color=EVENT)
    ax2.tick_params(axis="y", labelcolor=EVENT)
    ax2.grid(False)

    # Mark early-stopping point
    best_ep = int(np.argmax(val_scores)) + 1
    ax2.axvline(best_ep, color=SECONDARY, lw=1, ls=":")
    ax2.text(best_ep + 0.5, min(val_scores) + 0.01,
             f"Early stop\n(Epoche {best_ep})", color=SECONDARY, fontsize=9)

    ax1.set_title("MLP-Lernkurve: Verlust und Validierungsgüte über Epochen")
    # Combined legend
    h1, l1 = ax1.get_legend_handles_labels()
    h2, l2 = ax2.get_legend_handles_labels()
    ax1.legend(h1 + h2, l1 + l2, loc="center right")

    save(fig, ASSETS / "lernkurve.png")