def test_val_grad(): ####################################################################### # Not all methods computes the full Jacobian, but all # compute the gradients # check that the gradient returned by all methods are the same criterion = HeldOutMSE(idx_train, idx_val) algo = Forward() val_fwd, grad_fwd = criterion.get_val_grad(model, X, y, np.array( [log_alpha1, log_alpha2]), algo.get_beta_jac_v, tol=tol) criterion = HeldOutMSE(idx_train, idx_val) algo = ImplicitForward(tol_jac=1e-16, n_iter_jac=5000) val_imp_fwd, grad_imp_fwd = criterion.get_val_grad( model, X, y, np.array([log_alpha1, log_alpha2]), algo.get_beta_jac_v, tol=tol) criterion = HeldOutMSE(idx_train, idx_val) algo = ImplicitForward(tol_jac=1e-16, n_iter_jac=5000) val_imp_fwd_custom, grad_imp_fwd_custom = criterion.get_val_grad( model, X, y, np.array([log_alpha1, log_alpha2]), algo.get_beta_jac_v, tol=tol) criterion = HeldOutMSE(idx_train, idx_val) algo = Implicit() val_imp, grad_imp = criterion.get_val_grad(model, X, y, np.array( [log_alpha1, log_alpha2]), algo.get_beta_jac_v, tol=tol) np.testing.assert_allclose(val_fwd, val_imp_fwd) np.testing.assert_allclose(grad_fwd, grad_imp_fwd) np.testing.assert_allclose(val_imp_fwd, val_imp) np.testing.assert_allclose(val_imp_fwd, val_imp_fwd_custom) # for the implcit the conjugate grad does not converge # hence the rtol=1e-2 np.testing.assert_allclose(grad_imp_fwd, grad_imp, atol=1e-3) np.testing.assert_allclose(grad_imp_fwd, grad_imp_fwd_custom)
def test_grad_search(Optimizer, model, crit): """check that the paths are the same in the line search""" n_outer = 2 criterion = HeldOutMSE(idx_train, idx_val) monitor1 = Monitor() algo = Forward() optimizer = Optimizer(n_outer=n_outer, tol=1e-16) grad_search(algo, criterion, model, optimizer, X, y, alpha0, monitor1) criterion = HeldOutMSE(idx_train, idx_val) monitor2 = Monitor() algo = Implicit() optimizer = Optimizer(n_outer=n_outer, tol=1e-16) grad_search(algo, criterion, model, optimizer, X, y, alpha0, monitor2) criterion = HeldOutMSE(idx_train, idx_val) monitor3 = Monitor() algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=5000) optimizer = Optimizer(n_outer=n_outer, tol=1e-16) grad_search(algo, criterion, model, optimizer, X, y, alpha0, monitor3) np.testing.assert_allclose(np.array(monitor1.alphas), np.array(monitor3.alphas)) np.testing.assert_allclose(np.array(monitor1.grads), np.array(monitor3.grads), rtol=1e-5) np.testing.assert_allclose(np.array(monitor1.objs), np.array(monitor3.objs)) assert not np.allclose(np.array(monitor1.times), np.array(monitor3.times))
def test_val_grad(model): criterion = HeldOutLogistic(idx_val, idx_val) algo = Forward() val_fwd, grad_fwd = criterion.get_val_grad(model, X, y, log_alpha, algo.get_beta_jac_v, tol=tol) criterion = HeldOutLogistic(idx_val, idx_val) algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=5000) val_imp_fwd, grad_imp_fwd = criterion.get_val_grad(model, X, y, log_alpha, algo.get_beta_jac_v, tol=tol) criterion = HeldOutLogistic(idx_val, idx_val) algo = Implicit() val_imp, grad_imp = criterion.get_val_grad(model, X, y, log_alpha, algo.get_beta_jac_v, tol=tol) assert np.allclose(val_fwd, val_imp_fwd, atol=1e-4) assert np.allclose(grad_fwd, grad_imp_fwd, atol=1e-4) assert np.allclose(val_imp_fwd, val_imp, atol=1e-4) # for the implcit the conjugate grad does not converge # hence the rtol=1e-2 assert np.allclose(grad_imp_fwd, grad_imp, rtol=1e-2)
def test_grad_search(model, crit): """check that the paths are the same in the line search""" if crit == 'cv': n_outer = 2 criterion = HeldOutMSE(idx_train, idx_val) else: n_outer = 2 criterion = SmoothedSURE(sigma_star) # TODO MM@QBE if else scheme surprising criterion = HeldOutMSE(idx_train, idx_val) monitor1 = Monitor() algo = Forward() grad_search(algo, criterion, model, X, y, log_alpha, monitor1, n_outer=n_outer, tol=1e-16) criterion = HeldOutMSE(idx_train, idx_val) monitor2 = Monitor() algo = Implicit() grad_search(algo, criterion, model, X, y, log_alpha, monitor2, n_outer=n_outer, tol=1e-16) criterion = HeldOutMSE(idx_train, idx_val) monitor3 = Monitor() algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=5000) grad_search(algo, criterion, model, X, y, log_alpha, monitor3, n_outer=n_outer, tol=1e-16) np.testing.assert_allclose(np.array(monitor1.log_alphas), np.array(monitor3.log_alphas)) np.testing.assert_allclose(np.array(monitor1.grads), np.array(monitor3.grads), atol=1e-8) np.testing.assert_allclose(np.array(monitor1.objs), np.array(monitor3.objs)) assert not np.allclose(np.array(monitor1.times), np.array(monitor3.times))
def test_monitor(): model = Lasso(estimator=estimator) criterion = HeldOutMSE(idx_train, idx_val) algo = ImplicitForward() monitor = Monitor(callback=callback) optimizer = LineSearch(n_outer=10, tol=tol) grad_search(algo, criterion, model, optimizer, X, y, alpha0, monitor) np.testing.assert_allclose(np.array(monitor.objs), np.array(objs))
def test_grad_search_custom(model, model_custom, crit): """check that the paths are the same in the line search""" n_outer = 5 criterion = HeldOutLogistic(idx_val, idx_val) monitor = Monitor() algo = ImplicitForward(tol_jac=tol, n_iter_jac=5000) grad_search(algo, criterion, model, X, y, log_alpha, monitor, n_outer=n_outer, tol=tol) criterion = HeldOutLogistic(idx_val, idx_val) monitor_custom = Monitor() algo = ImplicitForward(tol_jac=tol, n_iter_jac=5000) grad_search(algo, criterion, model_custom, X, y, log_alpha, monitor_custom, n_outer=n_outer, tol=tol) np.testing.assert_allclose(np.array(monitor.log_alphas), np.array(monitor_custom.log_alphas), atol=1e-3) np.testing.assert_allclose(np.array(monitor.grads), np.array(monitor_custom.grads), atol=1e-4) np.testing.assert_allclose(np.array(monitor.objs), np.array(monitor_custom.objs), atol=1e-5) assert not np.allclose(np.array(monitor.times), np.array(monitor_custom.times))
def test_val_grad_custom(model, model_custom): criterion = HeldOutLogistic(idx_train, idx_val) algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=5000) val, grad = criterion.get_val_grad(model, X, y, log_alpha, algo.get_beta_jac_v, tol=tol) criterion = HeldOutLogistic(idx_train, idx_val) algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=5000) val_custom, grad_custom = criterion.get_val_grad(model_custom, X, y, log_alpha, algo.get_beta_jac_v, tol=tol) assert np.allclose(val, val_custom) assert np.allclose(grad, grad_custom)
def test_grad_search(model, crit): """check that the paths are the same in the line search""" n_outer = 2 criterion = HeldOutLogistic(idx_val, idx_val) monitor1 = Monitor() algo = Forward() grad_search(algo, criterion, model, X, y, log_alpha, monitor1, n_outer=n_outer, tol=tol) criterion = HeldOutLogistic(idx_val, idx_val) monitor2 = Monitor() algo = Implicit() grad_search(algo, criterion, model, X, y, log_alpha, monitor2, n_outer=n_outer, tol=tol) criterion = HeldOutLogistic(idx_val, idx_val) monitor3 = Monitor() algo = ImplicitForward(tol_jac=tol, n_iter_jac=5000) grad_search(algo, criterion, model, X, y, log_alpha, monitor3, n_outer=n_outer, tol=tol) assert np.allclose(np.array(monitor1.log_alphas), np.array(monitor3.log_alphas)) assert np.allclose(np.array(monitor1.grads), np.array(monitor3.grads), atol=1e-4) assert np.allclose(np.array(monitor1.objs), np.array(monitor3.objs)) assert not np.allclose(np.array(monitor1.times), np.array(monitor3.times))
def test_monitor(): model = Lasso(estimator=estimator) criterion = HeldOutMSE(idx_train, idx_val) algo = ImplicitForward() monitor = Monitor(callback=callback) grad_search(algo, criterion, model, X, y, np.log(alpha_max / 10), monitor, n_outer=10, tol=tol) np.testing.assert_allclose(np.array(monitor.objs), np.array(objs))
def test_grad_search(model, crit): """check that the paths are the same in the line search""" if crit == 'MSE': n_outer = 2 criterion = HeldOutMSE(idx_train, idx_val) else: n_outer = 2 criterion = FiniteDiffMonteCarloSure(sigma_star) # TODO MM@QBE if else scheme surprising criterion = HeldOutMSE(idx_train, idx_val) monitor1 = Monitor() algo = Forward() optimizer = LineSearch(n_outer=n_outer, tol=1e-16) grad_search(algo, criterion, model, optimizer, X, y, alpha0, monitor1) criterion = HeldOutMSE(idx_train, idx_val) monitor2 = Monitor() algo = Implicit() optimizer = LineSearch(n_outer=n_outer, tol=1e-16) grad_search(algo, criterion, model, optimizer, X, y, alpha0, monitor2) criterion = HeldOutMSE(idx_train, idx_val) monitor3 = Monitor() algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=5000) optimizer = LineSearch(n_outer=n_outer, tol=1e-16) grad_search(algo, criterion, model, optimizer, X, y, alpha0, monitor3) np.testing.assert_allclose(np.array(monitor1.alphas), np.array(monitor3.alphas)) np.testing.assert_allclose(np.array(monitor1.grads), np.array(monitor3.grads), rtol=1e-5) np.testing.assert_allclose(np.array(monitor1.objs), np.array(monitor3.objs)) assert not np.allclose(np.array(monitor1.times), np.array(monitor3.times))
############################################################################## # Grad-search # ----------- print('sparse-ho started') t0 = time.time() estimator = LogisticRegression(penalty='l1', fit_intercept=False, solver='saga', tol=tol) model = SparseLogreg(max_iter=max_iter, estimator=estimator) criterion = HeldOutLogistic(idx_train, idx_val) monitor_grad = Monitor() algo = ImplicitForward(tol_jac=tol, n_iter_jac=1000) grad_search(algo, criterion, model, X, y, np.log(0.1 * alpha_max), monitor_grad, n_outer=10, tol=tol) objs_grad = np.array(monitor_grad.objs) t_grad_search = time.time() - t0 print('sparse-ho finished') print("Time to compute CV for sparse-ho: %.2f" % t_grad_search)
def parallel_function(dataset_name, method, tol=1e-5, n_outer=50, tolerance_decrease='constant'): # load data X, y = fetch_libsvm(dataset_name) y -= np.mean(y) # compute alpha_max alpha_max = np.abs(X.T @ y).max() / len(y) if model_name == "logreg": alpha_max /= 2 alpha_min = alpha_max * dict_palphamin[dataset_name] if model_name == "enet": estimator = linear_model.ElasticNet(fit_intercept=False, max_iter=10_000, warm_start=True, tol=tol) model = ElasticNet(estimator=estimator) elif model_name == "logreg": model = SparseLogreg(estimator=estimator) # TODO improve this try: n_outer = dict_n_outers[dataset_name, method] except Exception: n_outer = 20 size_loop = 2 for _ in range(size_loop): if model_name == "lasso" or model_name == "enet": sub_criterion = HeldOutMSE(None, None) elif model_name == "logreg": criterion = HeldOutLogistic(None, None) kf = KFold(n_splits=5, shuffle=True, random_state=42) criterion = CrossVal(sub_criterion, cv=kf) algo = ImplicitForward(tol_jac=1e-3) monitor = Monitor() t_max = dict_t_max[dataset_name] if method == 'grid_search': num1D = dict_point_grid_search[dataset_name] alpha1D = np.geomspace(alpha_max, alpha_min, num=num1D) alphas = [np.array(i) for i in product(alpha1D, alpha1D)] grid_search(algo, criterion, model, X, y, alpha_min, alpha_max, monitor, max_evals=100, tol=tol, alphas=alphas) elif method == 'random' or method == 'bayesian': hyperopt_wrapper(algo, criterion, model, X, y, alpha_min, alpha_max, monitor, max_evals=30, tol=tol, method=method, size_space=2, t_max=t_max) elif method.startswith("implicit_forward"): # do gradient descent to find the optimal lambda alpha0 = np.array([alpha_max / 100, alpha_max / 100]) n_outer = 30 if method == 'implicit_forward': optimizer = GradientDescent(n_outer=n_outer, p_grad_norm=1, verbose=True, tol=tol, t_max=t_max) else: optimizer = GradientDescent(n_outer=n_outer, p_grad_norm=1, verbose=True, tol=tol, t_max=t_max, tol_decrease="geom") grad_search(algo, criterion, model, optimizer, X, y, alpha0, monitor) else: raise NotImplementedError monitor.times = np.array(monitor.times) monitor.objs = np.array(monitor.objs) monitor.objs_test = 0 # TODO monitor.alphas = np.array(monitor.alphas) return (dataset_name, method, tol, n_outer, tolerance_decrease, monitor.times, monitor.objs, monitor.objs_test, monitor.alphas, alpha_max, model_name)
n_jobs=2).fit(X, y) # Measure mse on test mse_cv = mean_squared_error(y_test, model_cv.predict(X_test)) print("Vanilla LassoCV: Mean-squared error on test data %f" % mse_cv) ############################################################################## ############################################################################## # Weighted Lasso with sparse-ho. # We use the vanilla lassoCV coefficients as a starting point log_alpha0 = np.log(model_cv.alpha_) * np.ones(n_features) # Weighted Lasso: Sparse-ho: 1 param per feature estimator = Lasso(fit_intercept=False, max_iter=10, warm_start=True) model = WeightedLasso(estimator=estimator) criterion = HeldOutMSE(idx_train, idx_val) algo = ImplicitForward() monitor = Monitor() grad_search(algo, criterion, model, X, y, log_alpha0, monitor, n_outer=20, tol=1e-6) ############################################################################## ############################################################################## # MSE on validation set mse_sho_val = mean_squared_error(y[idx_val], estimator.predict(X[idx_val, :]))
def parallel_function(dataset_name, div_alpha, method, ind_rep, random_state=10): maxit = dict_maxits[(dataset_name, div_alpha)][ind_rep] print("Dataset %s, algo %s, maxit %i" % (dataset_name, method, maxit)) X, y = fetch_libsvm(dataset_name) n_samples = len(y) kf = KFold(n_splits=5, random_state=random_state, shuffle=True) for i in range(2): alpha_max = np.max(np.abs(X.T.dot(y))) / n_samples log_alpha = np.log(alpha_max / div_alpha) monitor = Monitor() if method == "celer": clf = Lasso_celer( alpha=np.exp(log_alpha), fit_intercept=False, # TODO maybe change this tol tol=1e-8, max_iter=maxit) model = Lasso(estimator=clf, max_iter=maxit) criterion = HeldOutMSE(None, None) cross_val = CrossVal(cv=kf, criterion=criterion) algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=maxit, use_stop_crit=False) algo.max_iter = maxit val, grad = cross_val.get_val_grad(model, X, y, log_alpha, algo.get_beta_jac_v, tol=tol, monitor=monitor, max_iter=maxit) elif method == "ground_truth": for file in os.listdir("results/"): if file.startswith("hypergradient_%s_%i_%s" % (dataset_name, div_alpha, method)): return else: clf = Lasso_celer(alpha=np.exp(log_alpha), fit_intercept=False, warm_start=True, tol=1e-13, max_iter=10000) criterion = HeldOutMSE(None, None) cross_val = CrossVal(cv=kf, criterion=criterion) algo = Implicit(criterion) model = Lasso(estimator=clf, max_iter=10000) val, grad = cross_val.get_val_grad(model, X, y, log_alpha, algo.get_beta_jac_v, tol=1e-13, monitor=monitor) else: model = Lasso(max_iter=maxit) criterion = HeldOutMSE(None, None) cross_val = CrossVal(cv=kf, criterion=criterion) if method == "forward": algo = Forward(use_stop_crit=False) elif method == "implicit_forward": algo = ImplicitForward(use_stop_crit=False, tol_jac=1e-8, n_iter_jac=maxit, max_iter=1000) elif method == "implicit": algo = Implicit(use_stop_crit=False, max_iter=1000) elif method == "backward": algo = Backward() else: 1 / 0 algo.max_iter = maxit algo.use_stop_crit = False val, grad = cross_val.get_val_grad(model, X, y, log_alpha, algo.get_beta_jac_v, tol=tol, monitor=monitor, max_iter=maxit) results = (dataset_name, div_alpha, method, maxit, val, grad, monitor.times[0]) df = pandas.DataFrame(results).transpose() df.columns = [ 'dataset', 'div_alpha', 'method', 'maxit', 'val', 'grad', 'time' ] str_results = "results/hypergradient_%s_%i_%s_%i.pkl" % ( dataset_name, div_alpha, method, maxit) df.to_pickle(str_results)
def parallel_function(dataset_name, method): X, y = fetch_libsvm(dataset_name) X, y = fetch_libsvm(dataset_name) if dataset_name == "real-sim": X = X[:, :2000] X = csr_matrix(X) # very important for SVM my_bool = norm(X, axis=1) != 0 X = X[my_bool, :] y = y[my_bool] logC = dict_logC[dataset_name] for max_iter in dict_max_iter[dataset_name]: print("Dataset %s, max iter %i" % (method, max_iter)) for i in range(2): # TODO change this sss1 = StratifiedShuffleSplit(n_splits=2, test_size=0.3333, random_state=0) idx_train, idx_val = sss1.split(X, y) idx_train = idx_train[0] idx_val = idx_val[0] monitor = Monitor() criterion = HeldOutSmoothedHinge(idx_train, idx_val) model = SVM(estimator=None, max_iter=10_000) if method == "ground_truth": for file in os.listdir("results_svm/"): if file.startswith("hypergradient_svm_%s_%s" % (dataset_name, method)): return clf = LinearSVC(C=np.exp(logC), tol=1e-32, max_iter=10_000, loss='hinge', permute=False) algo = Implicit(criterion) model.estimator = clf val, grad = criterion.get_val_grad(model, X, y, logC, algo.compute_beta_grad, tol=1e-14, monitor=monitor) else: if method == "sota": clf = LinearSVC(C=np.exp(logC), loss='hinge', max_iter=max_iter, tol=1e-32, permute=False) model.estimator = clf algo = ImplicitForward(tol_jac=1e-32, n_iter_jac=max_iter, use_stop_crit=False) elif method == "forward": algo = Forward(use_stop_crit=False) elif method == "implicit_forward": algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=max_iter, use_stop_crit=False) else: raise NotImplementedError algo.max_iter = max_iter algo.use_stop_crit = False val, grad = criterion.get_val_grad(model, X, y, logC, algo.compute_beta_grad, tol=tol, monitor=monitor, max_iter=max_iter) results = (dataset_name, method, max_iter, val, grad, monitor.times[0]) df = pandas.DataFrame(results).transpose() df.columns = ['dataset', 'method', 'maxit', 'val', 'grad', 'time'] str_results = "results_svm/hypergradient_svm_%s_%s_%i.pkl" % ( dataset_name, method, max_iter) df.to_pickle(str_results)
tol=tol) objs = np.array(monitor_grid_sk.objs) t_sk = time.time() - t0 print('scikit-learn finished') ############################################################################## # Grad-search with sparse-ho # -------------------------- print('sparse-ho started') t0 = time.time() model = Lasso(estimator=estimator) criterion = HeldOutMSE(idx_train, idx_val) algo = ImplicitForward(criterion) monitor_grad = Monitor() grad_search(algo, criterion, model, X, y, np.log(alpha_max / 10), monitor_grad, n_outer=10, tol=tol) t_grad_search = time.time() - t0 print('sparse-ho finished')
dict_monitor = {} all_algo_name = ['grid_search'] # , 'implicit_forward', "implicit_forward_approx", 'bayesian'] # , 'random_search'] # all_algo_name = ['random_search'] for algo_name in all_algo_name: model = ElasticNet(estimator=estimator) sub_criterion = HeldOutMSE(None, None) alpha0 = np.array([alpha_max / 10, alpha_max / 10]) monitor = Monitor() kf = KFold(n_splits=5, shuffle=True, random_state=42) criterion = CrossVal(sub_criterion, cv=kf) algo = ImplicitForward(tol_jac=1e-3) # optimizer = LineSearch(n_outer=10, tol=tol) if algo_name.startswith('implicit_forward'): if algo_name == "implicit_forward_approx": optimizer = GradientDescent(n_outer=30, p_grad_norm=1., verbose=True, tol=tol, tol_decrease="geom") else: optimizer = GradientDescent(n_outer=30, p_grad_norm=1., verbose=True, tol=tol) grad_search(algo, criterion, model, optimizer, X, y, alpha0, monitor) elif algo_name == 'grid_search':
t_grid_search += time.time() print("Finished grid-search") ############################################################################## # Grad-search with sparse-ho # -------------------------- estimator = linear_model.ElasticNet(fit_intercept=False, max_iter=max_iter, warm_start=True) print("Started grad-search") t_grad_search = -time.time() monitor = Monitor() n_outer = 10 model = ElasticNet(max_iter=max_iter, estimator=estimator) criterion = HeldOutMSE(idx_train, idx_val) algo = ImplicitForward(tol_jac=1e-7, n_iter_jac=1000, max_iter=max_iter) grad_search(algo, criterion, model, X, y, verbose=True, log_alpha0=np.array( [np.log(alpha_max * 0.3), np.log(alpha_max / 10)]), tol=tol, n_outer=n_outer, monitor=monitor) t_grad_search += time.time() alphas_grad = np.exp(np.array(monitor.log_alphas)) alphas_grad /= alpha_max
from sparse_ho.algo.forward import get_beta_jac_iterdiff from sparse_ho.algo.implicit_forward import get_beta_jac_fast_iterdiff from sparse_ho.algo.implicit import get_beta_jac_t_v_implicit from sparse_ho.criterion import ( HeldOutMSE, FiniteDiffMonteCarloSure, HeldOutLogistic) from sparse_ho.tests.common import ( X, X_s, y, sigma_star, idx_train, idx_val, dict_log_alpha, models, custom_models, dict_cvxpy_func, dict_vals_cvxpy, dict_grads_cvxpy, dict_list_log_alphas, get_v, list_model_crit, list_model_names) # list of algorithms to be tested list_algos = [ Forward(), ImplicitForward(tol_jac=1e-16, n_iter_jac=5000), Implicit() # Backward() # XXX to fix ] tol = 1e-15 X_r = X_s.tocsr() X_c = X_s @pytest.mark.parametrize('key', list(models.keys())) def test_beta_jac(key): """Tests that algorithms computing the Jacobian return the same Jacobian""" if key == "svm" or key == "svr" or key == "ssvr": X_s = X_r else:
for maxit in maxits: for method in methods: print("Dataset %s, maxit %i" % (method, maxit)) for i in range(2): alpha_max = np.max(np.abs(X.T.dot(y))) / n_samples log_alpha = np.log(alpha_max * p_alpha_max) monitor = Monitor() if method == "celer": clf = Lasso_celer(alpha=np.exp(log_alpha), fit_intercept=False, tol=1e-12, max_iter=maxit) model = Lasso(estimator=clf, max_iter=maxit) criterion = HeldOutMSE(idx_train, idx_val) algo = ImplicitForward(tol_jac=1e-32, n_iter_jac=maxit, use_stop_crit=False) algo.max_iter = maxit val, grad = criterion.get_val_grad(model, X, y, log_alpha, algo.get_beta_jac_v, tol=1e-12, monitor=monitor, max_iter=maxit) else: model = Lasso(max_iter=maxit) criterion = HeldOutMSE(idx_train, idx_val) if method == "forward": algo = Forward()
def parallel_function(dataset_name, div_alpha, method): X, y = fetch_libsvm(dataset_name) n_samples = len(y) if dataset_name == "news20" and div_alpha == 100: rng = np.random.RandomState(42) y += rng.randn(n_samples) * 0.01 for maxit in dict_maxits[(dataset_name, div_alpha)]: print("Dataset %s, maxit %i" % (method, maxit)) for i in range(2): rng = np.random.RandomState(i) idx_train = rng.choice(n_samples, n_samples // 2, replace=False) idx = np.arange(0, n_samples) idx_val = idx[np.logical_not(np.isin(idx, idx_train))] alpha_max = np.max(np.abs(X[idx_train, :].T.dot(y[idx_train]))) alpha_max /= len(idx_train) log_alpha = np.log(alpha_max / div_alpha) monitor = Monitor() if method == "celer": clf = Lasso_celer(alpha=np.exp(log_alpha), fit_intercept=False, tol=1e-12, max_iter=maxit) model = Lasso(estimator=clf, max_iter=maxit) criterion = HeldOutMSE(idx_train, idx_val) algo = ImplicitForward(tol_jac=1e-32, n_iter_jac=maxit, use_stop_crit=False) algo.max_iter = maxit val, grad = criterion.get_val_grad(model, X, y, log_alpha, algo.compute_beta_grad, tol=1e-12, monitor=monitor, max_iter=maxit) elif method == "ground_truth": for file in os.listdir("results/"): if file.startswith("hypergradient_%s_%i_%s" % (dataset_name, div_alpha, method)): return clf = Lasso_celer(alpha=np.exp(log_alpha), fit_intercept=False, warm_start=True, tol=1e-14, max_iter=10000) criterion = HeldOutMSE(idx_train, idx_val) if dataset_name == "news20": algo = ImplicitForward(tol_jac=1e-11, n_iter_jac=100000) else: algo = Implicit(criterion) model = Lasso(estimator=clf, max_iter=10000) val, grad = criterion.get_val_grad(model, X, y, log_alpha, algo.compute_beta_grad, tol=1e-14, monitor=monitor) else: model = Lasso(max_iter=maxit) criterion = HeldOutMSE(idx_train, idx_val) if method == "forward": algo = Forward(use_stop_crit=False) elif method == "implicit_forward": algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=maxit, use_stop_crit=False) elif method == "implicit": algo = Implicit(max_iter=1000) elif method == "backward": algo = Backward() else: raise NotImplementedError algo.max_iter = maxit algo.use_stop_crit = False val, grad = criterion.get_val_grad(model, X, y, log_alpha, algo.compute_beta_grad, tol=tol, monitor=monitor, max_iter=maxit) results = (dataset_name, div_alpha, method, maxit, val, grad, monitor.times[0]) df = pandas.DataFrame(results).transpose() df.columns = [ 'dataset', 'div_alpha', 'method', 'maxit', 'val', 'grad', 'time' ] str_results = "results/hypergradient_%s_%i_%s_%i.pkl" % ( dataset_name, div_alpha, method, maxit) df.to_pickle(str_results)
def test_val_grad(): ####################################################################### # Not all methods computes the full Jacobian, but all # compute the gradients # check that the gradient returned by all methods are the same for key in models.keys(): log_alpha = dict_log_alpha[key] model = models[key] criterion = HeldOutMSE(idx_train, idx_val) algo = Forward() val_fwd, grad_fwd = criterion.get_val_grad(model, X, y, log_alpha, algo.get_beta_jac_v, tol=tol) criterion = HeldOutMSE(idx_train, idx_val) algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=5000) val_imp_fwd, grad_imp_fwd = criterion.get_val_grad(model, X, y, log_alpha, algo.get_beta_jac_v, tol=tol) criterion = HeldOutMSE(idx_train, idx_val) algo = Implicit() val_imp, grad_imp = criterion.get_val_grad(model, X, y, log_alpha, algo.get_beta_jac_v, tol=tol) criterion = HeldOutMSE(idx_train, idx_val) algo = Backward() val_bwd, grad_bwd = criterion.get_val_grad(model, X, y, log_alpha, algo.get_beta_jac_v, tol=tol) assert np.allclose(val_fwd, val_imp_fwd) assert np.allclose(grad_fwd, grad_imp_fwd) # assert np.allclose(val_imp_fwd, val_imp) assert np.allclose(val_bwd, val_fwd) assert np.allclose(val_bwd, val_imp_fwd) assert np.allclose(grad_fwd, grad_bwd) assert np.allclose(grad_bwd, grad_imp_fwd) # for the implcit the conjugate grad does not converge # hence the rtol=1e-2 assert np.allclose(grad_imp_fwd, grad_imp, atol=1e-3) for key in models.keys(): log_alpha = dict_log_alpha[key] model = models[key] criterion = SmoothedSURE(sigma_star) algo = Forward() val_fwd, grad_fwd = criterion.get_val_grad(model, X, y, log_alpha, algo.get_beta_jac_v, tol=tol) criterion = SmoothedSURE(sigma_star) algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=5000) val_imp_fwd, grad_imp_fwd = criterion.get_val_grad(model, X, y, log_alpha, algo.get_beta_jac_v, tol=tol) criterion = SmoothedSURE(sigma_star) algo = Implicit(criterion) val_imp, grad_imp = criterion.get_val_grad(model, X, y, log_alpha, algo.get_beta_jac_v, tol=tol) criterion = SmoothedSURE(sigma_star) algo = Backward() val_bwd, grad_bwd = criterion.get_val_grad(model, X, y, log_alpha, algo.get_beta_jac_v, tol=tol) assert np.allclose(val_fwd, val_imp_fwd) assert np.allclose(grad_fwd, grad_imp_fwd) assert np.allclose(val_imp_fwd, val_imp) assert np.allclose(val_bwd, val_fwd) assert np.allclose(val_bwd, val_imp_fwd) assert np.allclose(grad_fwd, grad_bwd) assert np.allclose(grad_bwd, grad_imp_fwd)
# n_features = 3_200 # X, y = fetch_libsvm('sensit') # X, y = fetch_libsvm('usps') X, y = fetch_libsvm('rcv1_multiclass') # X, y = fetch_libsvm('sector_scale') # X, y = fetch_libsvm('sector') # X, y = fetch_libsvm('smallNORB') # X, y = fetch_libsvm('mnist') # clean data and subsample X, y = clean_dataset(X, y, n_samples, n_features) idx_train, idx_val, idx_test = get_splits(X, y) n_samples, n_features = X.shape algo = ImplicitForward(n_iter_jac=1000) estimator = LogisticRegression( C=1, fit_intercept=False, warm_start=True, max_iter=2000, verbose=False) model = SparseLogreg(estimator=estimator) logit_multiclass = LogisticMulticlass( idx_train, idx_val, algo, idx_test=idx_test) alpha_max, n_classes = alpha_max_multiclass(X, y) tol = 1e-5 n_alphas = 10 p_alphas = np.geomspace(1, 0.001, n_alphas) p_alphas = np.tile(p_alphas, (n_classes, 1))
def test_grad_search(): n_outer = 3 criterion = HeldOutMSE(idx_train, idx_val) monitor1 = Monitor() algo = Forward() grad_search(algo, criterion, model, X, y, np.array([log_alpha1, log_alpha2]), monitor1, n_outer=n_outer, tol=1e-16) criterion = HeldOutMSE(idx_train, idx_val) monitor2 = Monitor() algo = Implicit() grad_search(algo, criterion, model, X, y, np.array([log_alpha1, log_alpha2]), monitor2, n_outer=n_outer, tol=1e-16) criterion = HeldOutMSE(idx_train, idx_val) monitor3 = Monitor() algo = ImplicitForward(tol_jac=1e-3, n_iter_jac=1000) grad_search(algo, criterion, model, X, y, np.array([log_alpha1, log_alpha2]), monitor3, n_outer=n_outer, tol=1e-16) [np.linalg.norm(grad) for grad in monitor1.grads] [np.exp(alpha) for alpha in monitor1.log_alphas] np.testing.assert_allclose(np.array(monitor1.log_alphas), np.array(monitor3.log_alphas)) np.testing.assert_allclose(np.array(monitor1.grads), np.array(monitor3.grads), rtol=1e-6) np.testing.assert_allclose(np.array(monitor1.objs), np.array(monitor3.objs), rtol=1e-6) assert not np.allclose(np.array(monitor1.times), np.array(monitor3.times)) np.testing.assert_allclose(np.array(monitor1.log_alphas), np.array(monitor2.log_alphas), atol=1e-2) np.testing.assert_allclose(np.array(monitor1.grads), np.array(monitor2.grads), atol=1e-2) np.testing.assert_allclose(np.array(monitor1.objs), np.array(monitor2.objs), atol=1e-2) assert not np.allclose(np.array(monitor1.times), np.array(monitor2.times))