def test_compare_gradientless_and_gradient_learning(): seq = [1, 1, 1, 1, 0, 1, 1, 0, 1, 1, 1, 0, 1, 1, 1] seqh = [1, 0, 1, 0, 0, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1] factors = [] hidden_factors = [] evidence = {} o_parameters = np.array([['a', 'b'], ['c', 'd']]) h_parameters = np.array([['e', 'f'], ['g', 'h']]) for i, (s, sh) in enumerate(zip(seq, seqh)): obs = DiscreteFactor(['o_{}'.format(i), 'h_{}'.format(i)], parameters=o_parameters) evidence['o_{}'.format(i)] = s evidence['h_{}'.format(i)] = sh factors.append(obs) if i < len(seq): hid = DiscreteFactor(['h_{}'.format(i), 'h_{}'.format(i + 1)], parameters=h_parameters) factors.append(hid) hidden_factors.append(hid) model = Model(factors) update_order = DistributeCollectProtocol(model) learn = LearnMrfParameters(model, update_order=update_order) learn.fit(evidence) print learn._optimizer_result #print learn.iterations ans1 = learn._optimizer_result[:2] print update_order = DistributeCollectProtocol(model) learn = LearnMrfParameters(model, update_order=update_order) fit_without_gradient(learn, evidence) print learn._optimizer_result #print learn.iterations ans2 = learn._optimizer_result[:2] assert_array_almost_equal(ans1[1], ans2[1], decimal=4) assert_array_almost_equal(ans1[0], ans2[0], decimal=4)
def test_learn_with_gradient_binary(self): tc1 = 70 tc2 = 14 obs = [DiscreteFactor([(i, 2)], parameters=np.array(['m0', 'm1'])) for i in xrange(tc1 + tc2)] model = Model(obs) evidence = dict((i, 0 if i < tc1 else 1) for i in xrange(tc1 + tc2)) print 'evidence', evidence print model.edges learner = LearnMrfParameters(model, prior=1.0) def nlog_posterior(x): c1 = tc1 c2 = tc2 x = x.reshape((2, 1)) N = np.zeros((2, 1)) Lamb = np.array([[1.0, 0], [0, 1.0]]) lp = -0.5 * np.dot(np.dot((x - N).T, Lamb), (x - N)) lp += -0.5 * x.shape[0] * np.log(2.0 * np.pi) lp += +0.5 * np.log(np.linalg.det((Lamb))) lp += np.dot(x.T, np.array([c1, c2])) - (c1 + c2) * np.log(sum(np.exp(x))) return -lp[0, 0] x0 = np.zeros(2) print 'zeros', x0 expected_ans = scipy.optimize.fmin_l_bfgs_b(nlog_posterior, x0, approx_grad=True, pgtol=10.0**-10) actual_ans = learner.fit(evidence).result() print actual_ans print expected_ans self.assertAlmostEqual(actual_ans[0], expected_ans[1]) assert_array_almost_equal(actual_ans[1], expected_ans[0], decimal=5)