nn

nn

def backprop(self, thetas):
    theta = self.shape_thetas(thetas)
    a1 = add_ones(X)

    z2 = a1.dot(theta[0])
    a2 = add_ones(g(z2))

    z3 = a2.dot(theta[1])
    a3 = add_ones(g(z3))

    z4 = a3.dot(theta[2])
    a4 = g(z4)

    d4 = -(Y-a4) * a4 * (1-a4)
    d3 = (theta[2].dot(d4.T).T * (a3 * (1-a3)))[:, 1:]
    d2 = (theta[1].dot(d3.T).T * (a2 * (1-a2)))[:, 1:]

    T3 = a3.T.dot(d4) / self.m
    T2 = a2.T.dot(d3) / self.m
    T1 = a1.T.dot(d2) / self.m

    grad = np.hstack((T1.flat, T2.flat, T3.flat))
    return grad

def J(self, thetas):
    Y = self.Y
    theta = self.shape_thetas(thetas)
    h0 = self.h(self.X, thetas)
    h0 = np.maximum(np.minimum(h0, 1 - 1e-15), 1e-15)

    return 1/self.m * .5 * np.sum(np.power(h0 - Y, 2))