test.py 文件源码

def adagrad():
    theta = ones((features.num_features)) * -.1 
    theta[0] = 1.0


    master_stepsize = 1e-1 #for example
    fudge_factor = 1e-6 #for numerical stability
    historical_grad = 0

    for iteration in xrange(10):
        print "iteration", iteration
        for i, (x, y) in enumerate(train):
            print i
            theta_g = zeros_like(theta)
            t.grad_features(x, y, i, theta, theta_g, features, threshold)
            historical_grad += theta_g * theta_g
            adjusted_grad = theta_g / (fudge_factor + np.sqrt(historical_grad))

            theta -= master_stepsize * adjusted_grad

        print f(theta)

    return theta
#cProfile.runctx("adagrad()", globals(), locals(), '.prof')
#s = pstats.Stats('.prof')
#s.strip_dirs().sort_stats('time').print_stats(30)