SCA.py

import numpy as np
from cnn_model import build_cnn_model

# -------------------------
# Fonction objectif pour SCA
# -------------------------
def objective_function(x, X_train, y_train):
    filter1 = int(x[0])
    filter2 = int(x[1])
    filter3 = int(x[2])
    learning_rate = float(x[3])
    dropout = float(x[4])

    model = build_cnn_model(
        input_shape=(X_train.shape[1], X_train.shape[2]),
        num_classes=y_train.shape[1],
        filter1=filter1,
        filter2=filter2,
        filter3=filter3,
        learning_rate=learning_rate,
        dropout=dropout
    )

    history = model.fit(
        X_train, y_train,
        validation_split=0.2,
        epochs=3,  # rapide pour optimisation
        batch_size=32,
        verbose=0
    )

    val_acc = history.history['val_accuracy'][-1]
    return 1 - val_acc  # minimiser 1 - précision

# -------------------------
# Implémentation simple de SCA
# -------------------------
def devsca(objf, lb, ub, dim, N, T, X_train, y_train):
    X = np.random.uniform(0, 1, (N, dim)) * (np.array(ub) - np.array(lb)) + np.array(lb)
    fitness = np.array([objf(ind, X_train, y_train) for ind in X])
    best_idx = np.argmin(fitness)
    best_X = X[best_idx].copy()
    best_fit = fitness[best_idx]
    convergence_curve = []

    for t in range(T):
        r1 = np.linspace(2, 0, T)[t]  # diminue avec le temps
        for i in range(N):
            for j in range(dim):
                r2 = 2 * np.pi * np.random.rand()
                r3 = 2 * np.random.rand()
                r4 = np.random.rand()
                if r4 < 0.5:
                    X[i, j] = X[i, j] + r1 * np.sin(r2) * abs(r3 * best_X[j] - X[i, j])
                else:
                    X[i, j] = X[i, j] + r1 * np.cos(r2) * abs(r3 * best_X[j] - X[i, j])

            X[i] = np.clip(X[i], lb, ub)

            fit = objf(X[i], X_train, y_train)
            if fit < best_fit:
                best_fit = fit
                best_X = X[i].copy()

        convergence_curve.append(best_fit)
        print(f"Iteration {t+1}/{T} | Best Accuracy: {1 - best_fit:.4f}")

    return best_X, best_fit, convergence_curve