Function test_donut

ann_class/xor_donut.py:104–152

()

def test_donut():
    # donut example
    N = 1000
    R_inner = 5
    R_outer = 10

    # distance from origin is radius + random normal
    # angle theta is uniformly distributed between (0, 2pi)
    R1 = np.random.randn(N//2) + R_inner
    theta = 2*np.pi*np.random.random(N//2)
    X_inner = np.concatenate([[R1 * np.cos(theta)], [R1 * np.sin(theta)]]).T

    R2 = np.random.randn(N//2) + R_outer
    theta = 2*np.pi*np.random.random(N//2)
    X_outer = np.concatenate([[R2 * np.cos(theta)], [R2 * np.sin(theta)]]).T

    X = np.concatenate([ X_inner, X_outer ])
    Y = np.array([0]*(N//2) + [1]*(N//2))

    n_hidden = 8
    W1 = np.random.randn(2, n_hidden)
    b1 = np.random.randn(n_hidden)
    W2 = np.random.randn(n_hidden)
    b2 = np.random.randn(1)
    LL = [] # keep track of log-likelihoods
    learning_rate = 0.00005
    regularization = 0.2
    last_error_rate = None
    for i in range(3000):
        pY, Z = forward(X, W1, b1, W2, b2)
        ll = get_log_likelihood(Y, pY)
        prediction = predict(X, W1, b1, W2, b2)
        er = np.abs(prediction - Y).mean()
        LL.append(ll)

        # get gradients
        gW2 = derivative_w2(Z, Y, pY)
        gb2 = derivative_b2(Y, pY)
        gW1 = derivative_w1(X, Z, Y, pY, W2)
        gb1 = derivative_b1(Z, Y, pY, W2)

        W2 += learning_rate * (gW2 - regularization * W2)
        b2 += learning_rate * (gb2 - regularization * b2)
        W1 += learning_rate * (gW1 - regularization * W1)
        b1 += learning_rate * (gb1 - regularization * b1)
        if i % 300 == 0:
            print("i:", i, "ll:", ll, "classification rate:", 1 - er)
    plt.plot(LL)
    plt.show()


if __name__ == '__main__':