python离散建模之感知器学习算法

def perce(X, y, w_init, rho, max_iter=1000):
    
    (N, nfeatures) = X.shape

    # Augment the feature vectors by adding a 1 to each one. (see lecture notes)
    X = np.hstack((X, np.ones((N, 1))))
    nfeatures += 1

    w = w_init  # initialise weights
    iter = 0
    mis_class = N  # start by assuming all samples are misclassified

    while mis_class > 0 and iter < max_iter:
        iter += 1
        mis_class = 0
        gradient = np.zeros(nfeatures)  # initaliase the gradients to 0

        # loop over every training sample.
        for i in range(N):
            # each misclassified point will cause the gradient to change
            if np.inner(X[i, :], w) * y[i] <= 0:
                mis_class += 1
                gradient += -y[i] * X[i, :]
        # update the weight vector ready for the next iteration
        # Note, also that the learning rate decays over time (rho/iter)
        w -= rho / iter * gradient

    return w, iter, mis_class

def perce_fast(X, y, w_init, rho, max_iter=10000):
  
    (N, nfeatures) = X.shape
    X = np.hstack((X, np.ones((N, 1))))
    nfeatures += 1
    w = w_init
    iter = 0
    mis_class = N
    yy = np.tile(y, (nfeatures, 1)).T
    while mis_class > 0 and iter < max_iter:
        iter += 1
        # Compute set of misclassified points
        mc = (np.dot(X, w.transpose()) * y) <= 0
        mis_class = np.sum(mc)
        # Update weights. Note, the learning rate decays over time (rho/iter)
        w -= rho / iter * (np.sum(-yy[mc, :] * X[mc, :], axis=0))
    return w, iter, np.sum(mc)