def callback(params): print("Log likelihood {}".format(-objective(params))) plt.cla() # Show posterior marginals. plot_xs = np.reshape(np.linspace(-7, 7, 300), (300,1)) pred_mean, pred_cov = predict(params, X, y, plot_xs) marg_std = np.sqrt(np.diag(pred_cov)) ax.plot(plot_xs, pred_mean, 'b') ax.fill(np.concatenate([plot_xs, plot_xs[::-1]]), np.concatenate([pred_mean - 1.96 * marg_std, (pred_mean + 1.96 * marg_std)[::-1]]), alpha=.15, fc='Blue', ec='None') # Show samples from posterior. rs = npr.RandomState(0) sampled_funcs = rs.multivariate_normal(pred_mean, pred_cov, size=10) ax.plot(plot_xs, sampled_funcs.T) ax.plot(X, y, 'kx') ax.set_ylim([-1.5, 1.5]) ax.set_xticks([]) ax.set_yticks([]) plt.draw() plt.pause(1.0/60.0)
def rmsprop(grad, x, callback=None, num_iters=100, step_size=0.1, gamma=0.9, eps = 10**-8): """Root mean squared prop: See Adagrad paper for details.""" avg_sq_grad = np.ones(len(x)) for i in range(num_iters): g = grad(x, i) if callback: callback(x, i, g) avg_sq_grad = avg_sq_grad * gamma + g**2 * (1 - gamma) x -= step_size * g/(np.sqrt(avg_sq_grad) + eps) return x
def adam(grad, x, callback=None, num_iters=100, step_size=0.001, b1=0.9, b2=0.999, eps=10**-8): """Adam as described in http://arxiv.org/pdf/1412.6980.pdf. It's basically RMSprop with momentum and some correction terms.""" m = np.zeros(len(x)) v = np.zeros(len(x)) for i in range(num_iters): g = grad(x, i) if callback: callback(x, i, g) m = (1 - b1) * g + b1 * m # First moment estimate. v = (1 - b2) * (g**2) + b2 * v # Second moment estimate. mhat = m / (1 - b1**(i + 1)) # Bias correction. vhat = v / (1 - b2**(i + 1)) x -= step_size*mhat/(np.sqrt(vhat) + eps) return x
def quick_grad_check(fun, arg0, extra_args=(), kwargs={}, verbose=True, eps=EPS, rtol=RTOL, atol=ATOL, rs=None): """Checks the gradient of a function (w.r.t. to its first arg) in a random direction""" if verbose: print("Checking gradient of {0} at {1}".format(fun, arg0)) if rs is None: rs = np.random.RandomState() random_dir = rs.standard_normal(np.shape(arg0)) random_dir = random_dir / np.sqrt(np.sum(random_dir * random_dir)) unary_fun = lambda x : fun(arg0 + x * random_dir, *extra_args, **kwargs) numeric_grad = unary_nd(unary_fun, 0.0, eps=eps) analytic_grad = np.sum(grad(fun)(arg0, *extra_args, **kwargs) * random_dir) assert np.allclose(numeric_grad, analytic_grad, rtol=rtol, atol=atol), \ "Check failed! nd={0}, ad={1}".format(numeric_grad, analytic_grad) if verbose: print("Gradient projection OK (numeric grad: {0}, analytic grad: {1})".format( numeric_grad, analytic_grad))