def test_condition_from_cryptominisat(self): parent = CategoricalHyperparameter('blkrest', ['0', '1'], default='1') child = UniformIntegerHyperparameter('blkrestlen', 2000, 10000, log=True) condition = EqualsCondition(child, parent, '1') self.assertFalse(condition.evaluate(dict(blkrest='0'))) self.assertTrue(condition.evaluate(dict(blkrest='1')))
def test_equals_condition(self): hp1 = CategoricalHyperparameter("parent", [0, 1]) hp2 = UniformIntegerHyperparameter("child", 0, 10) cond = EqualsCondition(hp2, hp1, 0) cond_ = EqualsCondition(hp2, hp1, 0) # Test invalid conditions: self.assertRaisesRegexp( ValueError, "Argument 'parent' is not an " "instance of HPOlibConfigSpace.hyperparameter." "Hyperparameter.", EqualsCondition, hp2, "parent", 0) self.assertRaisesRegexp( ValueError, "Argument 'child' is not an " "instance of HPOlibConfigSpace.hyperparameter." "Hyperparameter.", EqualsCondition, "child", hp1, 0) self.assertRaisesRegexp( ValueError, "The child and parent hyperparameter " "must be different hyperparameters.", EqualsCondition, hp1, hp1, 0) self.assertEqual(cond, cond_) cond_reverse = EqualsCondition(hp1, hp2, 0) self.assertNotEqual(cond, cond_reverse) self.assertNotEqual(cond, dict()) self.assertEqual("child | parent == 0", str(cond))
def get_hyperparameter_search_space(dataset_properties=None): loss = CategoricalHyperparameter( "loss", ["hinge", "log", "modified_huber", "squared_hinge", "perceptron"], default="hinge") penalty = CategoricalHyperparameter("penalty", ["l1", "l2", "elasticnet"], default="l2") alpha = UniformFloatHyperparameter("alpha", 10**-7, 10**-1, log=True, default=0.0001) l1_ratio = UniformFloatHyperparameter("l1_ratio", 0, 1, default=0.15) fit_intercept = UnParametrizedHyperparameter("fit_intercept", "True") n_iter = UniformIntegerHyperparameter("n_iter", 5, 1000, default=20) epsilon = UniformFloatHyperparameter("epsilon", 1e-5, 1e-1, default=1e-4, log=True) learning_rate = CategoricalHyperparameter( "learning_rate", ["optimal", "invscaling", "constant"], default="optimal") eta0 = UniformFloatHyperparameter("eta0", 10**-7, 0.1, default=0.01) power_t = UniformFloatHyperparameter("power_t", 1e-5, 1, default=0.5) # This does not allow for other resampling methods! class_weight = CategoricalHyperparameter("class_weight", ["None", "auto"], default="None") cs = ConfigurationSpace() cs.add_hyperparameter(loss) cs.add_hyperparameter(penalty) cs.add_hyperparameter(alpha) cs.add_hyperparameter(l1_ratio) cs.add_hyperparameter(fit_intercept) cs.add_hyperparameter(n_iter) cs.add_hyperparameter(epsilon) cs.add_hyperparameter(learning_rate) cs.add_hyperparameter(eta0) cs.add_hyperparameter(power_t) cs.add_hyperparameter(class_weight) # TODO add passive/aggressive here, although not properly documented? elasticnet = EqualsCondition(l1_ratio, penalty, "elasticnet") epsilon_condition = EqualsCondition(epsilon, loss, "modified_huber") # eta0 seems to be always active according to the source code; when # learning_rate is set to optimial, eta0 is the starting value: # https://github.com/scikit-learn/scikit-learn/blob/0.15.X/sklearn/linear_model/sgd_fast.pyx #eta0_and_inv = EqualsCondition(eta0, learning_rate, "invscaling") #eta0_and_constant = EqualsCondition(eta0, learning_rate, "constant") #eta0_condition = OrConjunction(eta0_and_inv, eta0_and_constant) power_t_condition = EqualsCondition(power_t, learning_rate, "invscaling") cs.add_condition(elasticnet) cs.add_condition(epsilon_condition) cs.add_condition(power_t_condition) return cs
def test_or_conjunction(self): self.assertRaises(TypeError, AndConjunction, "String1", "String2") hp1 = CategoricalHyperparameter("input1", [0, 1]) hp2 = CategoricalHyperparameter("input2", [0, 1]) hp3 = CategoricalHyperparameter("input3", [0, 1]) hp4 = Constant("Or", "True") cond1 = EqualsCondition(hp4, hp1, 1) self.assertRaises(ValueError, OrConjunction, cond1) cond2 = EqualsCondition(hp4, hp2, 1) cond3 = EqualsCondition(hp4, hp3, 1) andconj1 = OrConjunction(cond1, cond2) andconj1_ = OrConjunction(cond1, cond2) self.assertEqual(andconj1, andconj1_) andconj2 = OrConjunction(cond2, cond3) self.assertNotEqual(andconj1, andconj2) andconj3 = OrConjunction(cond1, cond2, cond3) self.assertEqual( "(Or | input1 == 1 || Or | input2 == 1 || Or | " "input3 == 1)", str(andconj3))
def test_and_conjunction(self): self.assertRaises(TypeError, AndConjunction, "String1", "String2") hp1 = CategoricalHyperparameter("input1", [0, 1]) hp2 = CategoricalHyperparameter("input2", [0, 1]) hp3 = CategoricalHyperparameter("input3", [0, 1]) hp4 = Constant("And", "True") cond1 = EqualsCondition(hp4, hp1, 1) # Only one condition in an AndConjunction! self.assertRaises(ValueError, AndConjunction, cond1) cond2 = EqualsCondition(hp4, hp2, 1) cond3 = EqualsCondition(hp4, hp3, 1) andconj1 = AndConjunction(cond1, cond2) andconj1_ = AndConjunction(cond1, cond2) self.assertEqual(andconj1, andconj1_) andconj2 = AndConjunction(cond2, cond3) self.assertNotEqual(andconj1, andconj2) andconj3 = AndConjunction(cond1, cond2, cond3) self.assertEqual( "(And | input1 == 1 && And | input2 == 1 && And | " "input3 == 1)", str(andconj3)) # Test __eq__ self.assertNotEqual(andconj1, andconj3) self.assertNotEqual(andconj1, "String")
def test_condition_with_cycles(self): cs = ConfigurationSpace() hp1 = CategoricalHyperparameter("parent", [0, 1]) cs.add_hyperparameter(hp1) hp2 = UniformIntegerHyperparameter("child", 0, 10) cs.add_hyperparameter(hp2) cond1 = EqualsCondition(hp2, hp1, 0) cs.add_condition(cond1) cond2 = EqualsCondition(hp1, hp2, 0) self.assertRaisesRegexp( ValueError, "Hyperparameter configuration " "contains a cycle \[\['child', 'parent'\]\]", cs.add_condition, cond2)
def get_hyperparameter_search_space(dataset_properties=None): n_components = UniformIntegerHyperparameter("n_components", 10, 2000, default=100) kernel = CategoricalHyperparameter( 'kernel', ['poly', 'rbf', 'sigmoid', 'cosine'], 'rbf') degree = UniformIntegerHyperparameter('degree', 2, 5, 3) gamma = UniformFloatHyperparameter("gamma", 3.0517578125e-05, 8, log=True, default=1.0) coef0 = UniformFloatHyperparameter("coef0", -1, 1, default=0) cs = ConfigurationSpace() cs.add_hyperparameter(n_components) cs.add_hyperparameter(kernel) cs.add_hyperparameter(degree) cs.add_hyperparameter(gamma) cs.add_hyperparameter(coef0) degree_depends_on_poly = EqualsCondition(degree, kernel, "poly") coef0_condition = InCondition(coef0, kernel, ["poly", "sigmoid"]) gamma_condition = InCondition(gamma, kernel, ["poly", "rbf"]) cs.add_condition(degree_depends_on_poly) cs.add_condition(coef0_condition) cs.add_condition(gamma_condition) return cs
def get_hyperparameter_search_space(dataset_properties=None): n_neighbors = UniformIntegerHyperparameter(name="n_neighbors", lower=1, upper=100, default=1) weights = CategoricalHyperparameter(name="weights", choices=["uniform", "distance"], default="uniform") metric = UnParametrizedHyperparameter(name="metric", value="minkowski") algorithm = Constant(name='algorithm', value="auto") p = CategoricalHyperparameter(name="p", choices=[1, 2, 5], default=2) leaf_size = Constant(name="leaf_size", value=30) # Unparametrized # TODO: If we further parametrize 'metric' we need more metric params metric = UnParametrizedHyperparameter(name="metric", value="minkowski") cs = ConfigurationSpace() cs.add_hyperparameter(n_neighbors) cs.add_hyperparameter(weights) cs.add_hyperparameter(metric) cs.add_hyperparameter(algorithm) cs.add_hyperparameter(p) cs.add_hyperparameter(leaf_size) # Conditions metric_p = EqualsCondition(parent=metric, child=p, value="minkowski") cs.add_condition(metric_p) return cs
def get_hyperparameter_search_space(dataset_properties=None): cs = ConfigurationSpace() loss = cs.add_hyperparameter(CategoricalHyperparameter("loss", ["squared_loss", "huber", "epsilon_insensitive", "squared_epsilon_insensitive"], default="squared_loss")) penalty = cs.add_hyperparameter(CategoricalHyperparameter( "penalty", ["l1", "l2", "elasticnet"], default="l2")) alpha = cs.add_hyperparameter(UniformFloatHyperparameter( "alpha", 10e-7, 1e-1, log=True, default=0.01)) l1_ratio = cs.add_hyperparameter(UniformFloatHyperparameter( "l1_ratio", 1e-9, 1., log=True, default=0.15)) fit_intercept = cs.add_hyperparameter(UnParametrizedHyperparameter( "fit_intercept", "True")) n_iter = cs.add_hyperparameter(UniformIntegerHyperparameter( "n_iter", 5, 1000, log=True, default=20)) epsilon = cs.add_hyperparameter(UniformFloatHyperparameter( "epsilon", 1e-5, 1e-1, default=1e-4, log=True)) learning_rate = cs.add_hyperparameter(CategoricalHyperparameter( "learning_rate", ["optimal", "invscaling", "constant"], default="optimal")) eta0 = cs.add_hyperparameter(UniformFloatHyperparameter( "eta0", 10 ** -7, 0.1, default=0.01)) power_t = cs.add_hyperparameter(UniformFloatHyperparameter( "power_t", 1e-5, 1, default=0.5)) average = cs.add_hyperparameter(CategoricalHyperparameter( "average", ["False", "True"], default="False")) # TODO add passive/aggressive here, although not properly documented? elasticnet = EqualsCondition(l1_ratio, penalty, "elasticnet") epsilon_condition = InCondition(epsilon, loss, ["huber", "epsilon_insensitive", "squared_epsilon_insensitive"]) # eta0 seems to be always active according to the source code; when # learning_rate is set to optimial, eta0 is the starting value: # https://github.com/scikit-learn/scikit-learn/blob/0.15.X/sklearn/linear_model/sgd_fast.pyx # eta0_and_inv = EqualsCondition(eta0, learning_rate, "invscaling") #eta0_and_constant = EqualsCondition(eta0, learning_rate, "constant") #eta0_condition = OrConjunction(eta0_and_inv, eta0_and_constant) power_t_condition = EqualsCondition(power_t, learning_rate, "invscaling") cs.add_condition(elasticnet) cs.add_condition(epsilon_condition) cs.add_condition(power_t_condition) return cs
def test_hyperparameters_with_valid_condition(self): cs = ConfigurationSpace() hp1 = CategoricalHyperparameter("parent", [0, 1]) cs.add_hyperparameter(hp1) hp2 = UniformIntegerHyperparameter("child", 0, 10) cs.add_hyperparameter(hp2) cond = EqualsCondition(hp2, hp1, 0) cs.add_condition(cond) self.assertEqual(len(cs._hyperparameters), 2)
def test_all_components_have_the_same_child(self): hp1 = CategoricalHyperparameter("input1", [0, 1]) hp2 = CategoricalHyperparameter("input2", [0, 1]) hp3 = CategoricalHyperparameter("input3", [0, 1]) hp4 = CategoricalHyperparameter("input4", [0, 1]) hp5 = CategoricalHyperparameter("input5", [0, 1]) hp6 = Constant("AND", "True") cond1 = EqualsCondition(hp1, hp2, 1) cond2 = EqualsCondition(hp1, hp3, 1) cond3 = EqualsCondition(hp1, hp4, 1) cond4 = EqualsCondition(hp6, hp4, 1) cond5 = EqualsCondition(hp6, hp5, 1) AndConjunction(cond1, cond2, cond3) AndConjunction(cond4, cond5) self.assertRaisesRegexp( ValueError, "All Conjunctions and Conditions must have " "the same child.", AndConjunction, cond1, cond4)
def test_add_second_condition_wo_conjunction(self): hp1 = CategoricalHyperparameter("input1", [0, 1]) hp2 = CategoricalHyperparameter("input2", [0, 1]) hp3 = Constant("And", "True") cond1 = EqualsCondition(hp3, hp1, 1) cond2 = EqualsCondition(hp3, hp2, 1) cs = ConfigurationSpace() cs.add_hyperparameter(hp1) cs.add_hyperparameter(hp2) cs.add_hyperparameter(hp3) cs.add_condition(cond1) self.assertRaisesRegexp( ValueError, "Adding a second condition \(different\) for a " "hyperparameter is ambigouos and " "therefore forbidden. Add a conjunction " "instead!", cs.add_condition, cond2)
def test_get_conditions(self): cs = ConfigurationSpace() hp1 = CategoricalHyperparameter("parent", [0, 1]) cs.add_hyperparameter(hp1) hp2 = UniformIntegerHyperparameter("child", 0, 10) cs.add_hyperparameter(hp2) self.assertEqual([], cs.get_conditions()) cond1 = EqualsCondition(hp2, hp1, 0) cs.add_condition(cond1) self.assertEqual([cond1], cs.get_conditions())
def test_add_conjunction(self): hp1 = CategoricalHyperparameter("input1", [0, 1]) hp2 = CategoricalHyperparameter("input2", [0, 1]) hp3 = CategoricalHyperparameter("input3", [0, 1]) hp4 = Constant("And", "True") cond1 = EqualsCondition(hp4, hp1, 1) cond2 = EqualsCondition(hp4, hp2, 1) cond3 = EqualsCondition(hp4, hp3, 1) andconj1 = AndConjunction(cond1, cond2, cond3) cs = ConfigurationSpace() cs.add_hyperparameter(hp1) cs.add_hyperparameter(hp2) cs.add_hyperparameter(hp3) cs.add_hyperparameter(hp4) cs.add_condition(andconj1) self.assertNotIn(hp4, cs.get_all_uncoditional_hyperparameters())
def get_hyperparameter_search_space(dataset_properties=None): cs = ConfigurationSpace() shrinkage = cs.add_hyperparameter(CategoricalHyperparameter( "shrinkage", ["None", "auto", "manual"], default="None")) shrinkage_factor = cs.add_hyperparameter(UniformFloatHyperparameter( "shrinkage_factor", 0., 1., 0.5)) n_components = cs.add_hyperparameter(UniformIntegerHyperparameter( 'n_components', 1, 250, default=10)) tol = cs.add_hyperparameter(UniformFloatHyperparameter( "tol", 1e-5, 1e-1, default=1e-4, log=True)) cs.add_condition(EqualsCondition(shrinkage_factor, shrinkage, "manual")) return cs
def test_get_hyperparameters(self): cs = ConfigurationSpace() self.assertEqual(0, len(cs.get_hyperparameters())) hp1 = CategoricalHyperparameter("parent", [0, 1]) cs.add_hyperparameter(hp1) self.assertEqual([hp1], cs.get_hyperparameters()) hp2 = UniformIntegerHyperparameter("child", 0, 10) cs.add_hyperparameter(hp2) cond1 = EqualsCondition(hp2, hp1, 1) cs.add_condition(cond1) self.assertEqual([hp1, hp2], cs.get_hyperparameters()) # TODO: I need more tests for the topological sort! self.assertEqual([hp1, hp2], cs.get_hyperparameters())
def test_check_configuration2(self): # Test that hyperparameters which are not active must not be set and # that evaluating forbidden clauses does not choke on missing # hyperparameters cs = ConfigurationSpace() classifier = CategoricalHyperparameter( "classifier", ["k_nearest_neighbors", "extra_trees"]) metric = CategoricalHyperparameter("metric", ["minkowski", "other"]) p = CategoricalHyperparameter("k_nearest_neighbors:p", [1, 2]) metric_depends_on_classifier = EqualsCondition(metric, classifier, "k_nearest_neighbors") p_depends_on_metric = EqualsCondition(p, metric, "minkowski") cs.add_hyperparameter(metric) cs.add_hyperparameter(p) cs.add_hyperparameter(classifier) cs.add_condition(metric_depends_on_classifier) cs.add_condition(p_depends_on_metric) forbidden = ForbiddenEqualsClause(metric, "other") cs.add_forbidden_clause(forbidden) configuration = Configuration(cs, dict(classifier="extra_trees"))
def get_hyperparameter_search_space(dataset_properties=None): cs = ConfigurationSpace() n_components = cs.add_hyperparameter(UniformIntegerHyperparameter( "n_components", 10, 2000, default=100)) algorithm = cs.add_hyperparameter(CategoricalHyperparameter('algorithm', ['parallel', 'deflation'], 'parallel')) whiten = cs.add_hyperparameter(CategoricalHyperparameter('whiten', ['False', 'True'], 'False')) fun = cs.add_hyperparameter(CategoricalHyperparameter( 'fun', ['logcosh', 'exp', 'cube'], 'logcosh')) cs.add_condition(EqualsCondition(n_components, whiten, "True")) return cs
def test_nested_conjunctions(self): hp1 = CategoricalHyperparameter("input1", [0, 1]) hp2 = CategoricalHyperparameter("input2", [0, 1]) hp3 = CategoricalHyperparameter("input3", [0, 1]) hp4 = CategoricalHyperparameter("input4", [0, 1]) hp5 = CategoricalHyperparameter("input5", [0, 1]) hp6 = Constant("AND", "True") cond1 = EqualsCondition(hp6, hp1, 1) cond2 = EqualsCondition(hp6, hp2, 1) cond3 = EqualsCondition(hp6, hp3, 1) cond4 = EqualsCondition(hp6, hp4, 1) cond5 = EqualsCondition(hp6, hp5, 1) conj1 = AndConjunction(cond1, cond2) conj2 = OrConjunction(conj1, cond3) conj3 = AndConjunction(conj2, cond4, cond5) # TODO: this does not look nice, And should depend on a large # conjunction, there should not be many ANDs inside this string! self.assertEqual( "(((AND | input1 == 1 && AND | input2 == 1) || AND | " "input3 == 1) && AND | input4 == 1 && AND | input5 " "== 1)", str(conj3))
def get_hyperparameter_search_space(dataset_properties=None): cs = ConfigurationSpace() use_minimum_fraction = cs.add_hyperparameter( CategoricalHyperparameter("use_minimum_fraction", ["True", "False"], default="True")) minimum_fraction = cs.add_hyperparameter( UniformFloatHyperparameter("minimum_fraction", lower=.0001, upper=0.5, default=0.01, log=True)) cs.add_condition( EqualsCondition(minimum_fraction, use_minimum_fraction, 'True')) return cs
def get_hyperparameter_search_space(dataset_properties=None): C = UniformFloatHyperparameter("C", 0.03125, 32768, log=True, default=1.0) # No linear kernel here, because we have liblinear kernel = CategoricalHyperparameter(name="kernel", choices=["rbf", "poly", "sigmoid"], default="rbf") degree = UniformIntegerHyperparameter("degree", 1, 5, default=3) gamma = UniformFloatHyperparameter("gamma", 3.0517578125e-05, 8, log=True, default=0.1) # TODO this is totally ad-hoc coef0 = UniformFloatHyperparameter("coef0", -1, 1, default=0) # probability is no hyperparameter, but an argument to the SVM algo shrinking = CategoricalHyperparameter("shrinking", ["True", "False"], default="True") tol = UniformFloatHyperparameter("tol", 1e-5, 1e-1, default=1e-4, log=True) # cache size is not a hyperparameter, but an argument to the program! max_iter = UnParametrizedHyperparameter("max_iter", -1) cs = ConfigurationSpace() cs.add_hyperparameter(C) cs.add_hyperparameter(kernel) cs.add_hyperparameter(degree) cs.add_hyperparameter(gamma) cs.add_hyperparameter(coef0) cs.add_hyperparameter(shrinking) cs.add_hyperparameter(tol) cs.add_hyperparameter(max_iter) degree_depends_on_poly = EqualsCondition(degree, kernel, "poly") coef0_condition = InCondition(coef0, kernel, ["poly", "sigmoid"]) cs.add_condition(degree_depends_on_poly) cs.add_condition(coef0_condition) return cs
def test_sample_configuration(self): cs = ConfigurationSpace() hp1 = CategoricalHyperparameter("parent", [0, 1]) cs.add_hyperparameter(hp1) hp2 = UniformIntegerHyperparameter("child", 0, 10) cs.add_hyperparameter(hp2) cond1 = EqualsCondition(hp2, hp1, 0) cs.add_condition(cond1) # This automatically checks the configuration! Configuration(cs, dict(parent=0, child=5)) # and now for something more complicated cs = ConfigurationSpace(seed=1) hp1 = CategoricalHyperparameter("input1", [0, 1]) cs.add_hyperparameter(hp1) hp2 = CategoricalHyperparameter("input2", [0, 1]) cs.add_hyperparameter(hp2) hp3 = CategoricalHyperparameter("input3", [0, 1]) cs.add_hyperparameter(hp3) hp4 = CategoricalHyperparameter("input4", [0, 1]) cs.add_hyperparameter(hp4) hp5 = CategoricalHyperparameter("input5", [0, 1]) cs.add_hyperparameter(hp5) hp6 = Constant("AND", "True") cs.add_hyperparameter(hp6) cond1 = EqualsCondition(hp6, hp1, 1) cond2 = NotEqualsCondition(hp6, hp2, 1) cond3 = InCondition(hp6, hp3, [1]) cond4 = EqualsCondition(hp5, hp3, 1) cond5 = EqualsCondition(hp4, hp5, 1) cond6 = EqualsCondition(hp6, hp4, 1) cond7 = EqualsCondition(hp6, hp5, 1) conj1 = AndConjunction(cond1, cond2) conj2 = OrConjunction(conj1, cond3) conj3 = AndConjunction(conj2, cond6, cond7) cs.add_condition(cond4) cs.add_condition(cond5) cs.add_condition(conj3) samples = [] for i in range(5): cs.seed(1) samples.append([]) for j in range(100): sample = cs.sample_configuration() samples[-1].append(sample) if i > 0: for j in range(100): self.assertEqual(samples[-1][j], samples[-2][j])
def test_condition_without_added_hyperparameters(self): cs = ConfigurationSpace() hp1 = CategoricalHyperparameter("parent", [0, 1]) hp2 = UniformIntegerHyperparameter("child", 0, 10) cond = EqualsCondition(hp2, hp1, 0) self.assertRaisesRegexp( ValueError, "Child hyperparameter 'child' not " "in configuration space.", cs.add_condition, cond) cs.add_hyperparameter(hp1) self.assertRaisesRegexp( ValueError, "Child hyperparameter 'child' not " "in configuration space.", cs.add_condition, cond) # Test also the parent hyperparameter cs2 = ConfigurationSpace() cs2.add_hyperparameter(hp2) self.assertRaisesRegexp( ValueError, "Parent hyperparameter 'parent' " "not in configuration space.", cs2.add_condition, cond)
def get_hyperparameter_search_space(dataset_properties=None): if dataset_properties is not None and \ (dataset_properties.get("sparse") is True or dataset_properties.get("signed") is False): allow_chi2 = False else: allow_chi2 = True possible_kernels = ['poly', 'rbf', 'sigmoid', 'cosine'] if allow_chi2: possible_kernels.append("chi2") kernel = CategoricalHyperparameter('kernel', possible_kernels, 'rbf') degree = UniformIntegerHyperparameter('degree', 2, 5, 3) gamma = UniformFloatHyperparameter("gamma", 3.0517578125e-05, 8, log=True, default=0.1) coef0 = UniformFloatHyperparameter("coef0", -1, 1, default=0) n_components = UniformIntegerHyperparameter("n_components", 50, 10000, default=100, log=True) cs = ConfigurationSpace() cs.add_hyperparameter(kernel) cs.add_hyperparameter(degree) cs.add_hyperparameter(gamma) cs.add_hyperparameter(coef0) cs.add_hyperparameter(n_components) degree_depends_on_poly = EqualsCondition(degree, kernel, "poly") coef0_condition = InCondition(coef0, kernel, ["poly", "sigmoid"]) gamma_kernels = ["poly", "rbf", "sigmoid"] if allow_chi2: gamma_kernels.append("chi2") gamma_condition = InCondition(gamma, kernel, gamma_kernels) cs.add_condition(degree_depends_on_poly) cs.add_condition(coef0_condition) cs.add_condition(gamma_condition) return cs
def test_get_hyperparameters_topological_sort(self): for iteration in range(10): cs = ConfigurationSpace() hp1 = CategoricalHyperparameter("parent", [0, 1]) cs.add_hyperparameter(hp1) hp2 = UniformIntegerHyperparameter("child", 0, 10) cs.add_hyperparameter(hp2) cond1 = EqualsCondition(hp2, hp1, 0) cs.add_condition(cond1) # This automatically checks the configuration! Configuration(cs, dict(parent=0, child=5)) # and now for something more complicated cs = ConfigurationSpace() hp1 = CategoricalHyperparameter("input1", [0, 1]) hp2 = CategoricalHyperparameter("input2", [0, 1]) hp3 = CategoricalHyperparameter("input3", [0, 1]) hp4 = CategoricalHyperparameter("input4", [0, 1]) hp5 = CategoricalHyperparameter("input5", [0, 1]) hp6 = Constant("AND", "True") # More top-level hyperparameters hp7 = CategoricalHyperparameter("input7", [0, 1]) hps = [hp1, hp2, hp3, hp4, hp5, hp6, hp7] random.shuffle(hps) for hp in hps: cs.add_hyperparameter(hp) cond1 = EqualsCondition(hp6, hp1, 1) cond2 = NotEqualsCondition(hp6, hp2, 1) cond3 = InCondition(hp6, hp3, [1]) cond4 = EqualsCondition(hp5, hp3, 1) cond5 = EqualsCondition(hp4, hp5, 1) cond6 = EqualsCondition(hp6, hp4, 1) cond7 = EqualsCondition(hp6, hp5, 1) conj1 = AndConjunction(cond1, cond2) conj2 = OrConjunction(conj1, cond3) conj3 = AndConjunction(conj2, cond6, cond7) cs.add_condition(cond4) cs.add_condition(cond5) cs.add_condition(conj3) hps = cs.get_hyperparameters() self.assertEqual(hps.index(hp1), 0) self.assertEqual(hps.index(hp2), 1) self.assertEqual(hps.index(hp3), 2) self.assertEqual(hps.index(hp7), 3) self.assertEqual(hps.index(hp5), 4) self.assertEqual(hps.index(hp4), 5) self.assertEqual(hps.index(hp6), 6)
def test_add_configuration_space(self): cs = ConfigurationSpace() hp1 = cs.add_hyperparameter(CategoricalHyperparameter( "input1", [0, 1])) forb1 = cs.add_forbidden_clause(ForbiddenEqualsClause(hp1, 1)) hp2 = cs.add_hyperparameter( UniformIntegerHyperparameter("child", 0, 10)) cond = cs.add_condition(EqualsCondition(hp2, hp1, 0)) cs2 = ConfigurationSpace() cs2.add_configuration_space('prefix', cs, delimiter='__') self.assertEqual( str(cs2), '''Configuration space object: Hyperparameters: prefix__child, Type: UniformInteger, Range: [0, 10], Default: 5 prefix__input1, Type: Categorical, Choices: {0, 1}, Default: 0 Conditions: prefix__child | prefix__input1 == 0 Forbidden Clauses: Forbidden: prefix__input1 == 1 ''')
def test_get_parent_and_children_of(self): cs = ConfigurationSpace() hp1 = CategoricalHyperparameter("parent", [0, 1]) cs.add_hyperparameter(hp1) hp2 = UniformIntegerHyperparameter("child", 0, 10) cs.add_hyperparameter(hp2) cond1 = EqualsCondition(hp2, hp1, 0) cs.add_condition(cond1) self.assertEqual([hp1], cs.get_parents_of(hp2.name)) self.assertEqual([hp1], cs.get_parents_of(hp2)) self.assertEqual([hp2], cs.get_children_of(hp1.name)) self.assertEqual([hp2], cs.get_children_of(hp1)) self.assertRaisesRegexp( KeyError, "Hyperparameter 'Foo' does not exist in this " "configuration space.", cs.get_parents_of, "Foo") self.assertRaisesRegexp( KeyError, "Hyperparameter 'Foo' does not exist in this " "configuration space.", cs.get_children_of, "Foo")
def test_repr(self): cs1 = ConfigurationSpace() retval = cs1.__str__() self.assertEqual("Configuration space object:\n Hyperparameters:\n", retval) hp1 = CategoricalHyperparameter("parent", [0, 1]) cs1.add_hyperparameter(hp1) retval = cs1.__str__() self.assertEqual( "Configuration space object:\n Hyperparameters:\n" " %s\n" % str(hp1), retval) hp2 = UniformIntegerHyperparameter("child", 0, 10) cond1 = EqualsCondition(hp2, hp1, 0) cs1.add_hyperparameter(hp2) cs1.add_condition(cond1) retval = cs1.__str__() self.assertEqual( "Configuration space object:\n Hyperparameters:\n" " %s\n %s\n Conditions:\n %s\n" % (str(hp2), str(hp1), str(cond1)), retval)
def test_eq(self): # Compare empty configuration spaces cs1 = ConfigurationSpace() cs2 = ConfigurationSpace() self.assertEqual(cs1, cs2) # Compare to something which isn't a configuration space self.assertTrue(not (cs1 == "ConfigurationSpace")) # Compare to equal configuration spaces hp1 = CategoricalHyperparameter("parent", [0, 1]) hp2 = UniformIntegerHyperparameter("child", 0, 10) hp3 = UniformIntegerHyperparameter("friend", 0, 5) cond1 = EqualsCondition(hp2, hp1, 0) cs1.add_hyperparameter(hp1) cs1.add_hyperparameter(hp2) cs1.add_condition(cond1) cs2.add_hyperparameter(hp1) cs2.add_hyperparameter(hp2) cs2.add_condition(cond1) self.assertEqual(cs1, cs2) cs1.add_hyperparameter(hp3) self.assertFalse(cs1 == cs2)
def get_hyperparameter_search_space(dataset_properties=None): max_num_layers = 7 # Maximum number of layers coded # Hacky way to condition layers params based on the number of layers # 'c'=1, 'd'=2, 'e'=3 ,'f'=4', g ='5', h='6' + output_layer layer_choices = [ chr(i) for i in range(ord('c'), ord('b') + max_num_layers) ] batch_size = UniformIntegerHyperparameter("batch_size", 32, 4096, log=True, default=32) number_epochs = UniformIntegerHyperparameter("number_epochs", 2, 80, default=5) num_layers = CategoricalHyperparameter("num_layers", choices=layer_choices, default='c') lr = UniformFloatHyperparameter("learning_rate", 1e-6, 1.0, log=True, default=0.01) l2 = UniformFloatHyperparameter("lambda2", 1e-7, 1e-2, log=True, default=1e-4) dropout_output = UniformFloatHyperparameter("dropout_output", 0.0, 0.99, default=0.5) # Define basic hyperparameters and define the config space # basic means that are independent from the number of layers cs = ConfigurationSpace() cs.add_hyperparameter(number_epochs) cs.add_hyperparameter(batch_size) cs.add_hyperparameter(num_layers) cs.add_hyperparameter(lr) cs.add_hyperparameter(l2) cs.add_hyperparameter(dropout_output) # Define parameters with different child parameters and conditions solver_choices = [ "adam", "adadelta", "adagrad", "sgd", "momentum", "nesterov", "smorm3s" ] solver = CategoricalHyperparameter(name="solver", choices=solver_choices, default="smorm3s") beta1 = UniformFloatHyperparameter("beta1", 1e-4, 0.1, log=True, default=0.1) beta2 = UniformFloatHyperparameter("beta2", 1e-4, 0.1, log=True, default=0.01) rho = UniformFloatHyperparameter("rho", 0.05, 0.99, log=True, default=0.95) momentum = UniformFloatHyperparameter("momentum", 0.3, 0.999, default=0.9) # TODO: Add policy based on this sklearn sgd policy_choices = ['fixed', 'inv', 'exp', 'step'] lr_policy = CategoricalHyperparameter(name="lr_policy", choices=policy_choices, default='fixed') gamma = UniformFloatHyperparameter(name="gamma", lower=1e-3, upper=1e-1, default=1e-2) power = UniformFloatHyperparameter("power", 0.0, 1.0, default=0.5) epoch_step = UniformIntegerHyperparameter("epoch_step", 2, 20, default=5) cs.add_hyperparameter(solver) cs.add_hyperparameter(beta1) cs.add_hyperparameter(beta2) cs.add_hyperparameter(momentum) cs.add_hyperparameter(rho) cs.add_hyperparameter(lr_policy) cs.add_hyperparameter(gamma) cs.add_hyperparameter(power) cs.add_hyperparameter(epoch_step) # Define parameters that are needed it for each layer output_activation_choices = ['softmax', 'sigmoid', 'softplus', 'tanh'] activations_choices = [ 'sigmoid', 'tanh', 'scaledTanh', 'elu', 'relu', 'leaky', 'linear' ] weight_choices = [ 'constant', 'normal', 'uniform', 'glorot_normal', 'glorot_uniform', 'he_normal', 'he_uniform', 'ortogonal', 'sparse' ] # Iterate over parameters that are used in each layer for i in range(1, max_num_layers): layer_units = UniformIntegerHyperparameter("num_units_layer_" + str(i), 64, 4096, log=True, default=128) cs.add_hyperparameter(layer_units) layer_dropout = UniformFloatHyperparameter("dropout_layer_" + str(i), 0.0, 0.99, default=0.5) cs.add_hyperparameter(layer_dropout) weight_initialization = CategoricalHyperparameter( 'weight_init_' + str(i), choices=weight_choices, default='he_normal') cs.add_hyperparameter(weight_initialization) layer_std = UniformFloatHyperparameter("std_layer_" + str(i), 1e-6, 0.1, log=True, default=0.005) cs.add_hyperparameter(layer_std) layer_activation = CategoricalHyperparameter( "activation_layer_" + str(i), choices=activations_choices, default="relu") cs.add_hyperparameter(layer_activation) layer_leakiness = UniformFloatHyperparameter('leakiness_layer_' + str(i), 0.01, 0.99, default=0.3) cs.add_hyperparameter(layer_leakiness) layer_tanh_alpha = UniformFloatHyperparameter('tanh_alpha_layer_' + str(i), 0.5, 1.0, default=2. / 3.) cs.add_hyperparameter(layer_tanh_alpha) layer_tanh_beta = UniformFloatHyperparameter('tanh_beta_layer_' + str(i), 1.1, 3.0, log=True, default=1.7159) cs.add_hyperparameter(layer_tanh_beta) # TODO: Could be in a function in a new module for i in range(2, max_num_layers): # Condition layers parameter on layer choice layer_unit_param = cs.get_hyperparameter("num_units_layer_" + str(i)) layer_cond = InCondition(child=layer_unit_param, parent=num_layers, values=[l for l in layer_choices[i - 1:]]) cs.add_condition(layer_cond) # Condition dropout parameter on layer choice layer_dropout_param = cs.get_hyperparameter("dropout_layer_" + str(i)) layer_cond = InCondition(child=layer_dropout_param, parent=num_layers, values=[l for l in layer_choices[i - 1:]]) cs.add_condition(layer_cond) # Condition weight initialization on layer choice layer_weight_param = cs.get_hyperparameter("weight_init_" + str(i)) layer_cond = InCondition(child=layer_weight_param, parent=num_layers, values=[l for l in layer_choices[i - 1:]]) cs.add_condition(layer_cond) # Condition std parameter on weight layer initialization choice layer_std_param = cs.get_hyperparameter("std_layer_" + str(i)) weight_cond = EqualsCondition(child=layer_std_param, parent=layer_weight_param, value='normal') cs.add_condition(weight_cond) # Condition activation parameter on layer choice layer_activation_param = cs.get_hyperparameter( "activation_layer_" + str(i)) layer_cond = InCondition(child=layer_activation_param, parent=num_layers, values=[l for l in layer_choices[i - 1:]]) cs.add_condition(layer_cond) # Condition leakiness on activation choice layer_leakiness_param = cs.get_hyperparameter("leakiness_layer_" + str(i)) activation_cond = EqualsCondition(child=layer_leakiness_param, parent=layer_activation_param, value='leaky') cs.add_condition(activation_cond) # Condition tanh on activation choice layer_tanh_alpha_param = cs.get_hyperparameter( "tanh_alpha_layer_" + str(i)) activation_cond = EqualsCondition(child=layer_tanh_alpha_param, parent=layer_activation_param, value='scaledTanh') cs.add_condition(activation_cond) layer_tanh_beta_param = cs.get_hyperparameter("tanh_beta_layer_" + str(i)) activation_cond = EqualsCondition(child=layer_tanh_beta_param, parent=layer_activation_param, value='scaledTanh') cs.add_condition(activation_cond) # Conditioning on solver momentum_depends_on_solver = InCondition( momentum, solver, values=["momentum", "nesterov"]) beta1_depends_on_solver = EqualsCondition(beta1, solver, "adam") beta2_depends_on_solver = EqualsCondition(beta2, solver, "adam") rho_depends_on_solver = EqualsCondition(rho, solver, "adadelta") cs.add_condition(momentum_depends_on_solver) cs.add_condition(beta1_depends_on_solver) cs.add_condition(beta2_depends_on_solver) cs.add_condition(rho_depends_on_solver) # Conditioning on learning rate policy lr_policy_depends_on_solver = InCondition( lr_policy, solver, ["adadelta", "adagrad", "sgd", "momentum", "nesterov"]) gamma_depends_on_policy = InCondition(child=gamma, parent=lr_policy, values=["inv", "exp", "step"]) power_depends_on_policy = EqualsCondition(power, lr_policy, "inv") epoch_step_depends_on_policy = EqualsCondition(epoch_step, lr_policy, "step") cs.add_condition(lr_policy_depends_on_solver) cs.add_condition(gamma_depends_on_policy) cs.add_condition(power_depends_on_policy) cs.add_condition(epoch_step_depends_on_policy) return cs
def get_hyperparameter_search_space(cls, dataset_properties, default=None, include=None, exclude=None): cs = ConfigurationSpace() # Compile a list of legal preprocessors for this problem available_preprocessors = cls.get_available_components( data_prop=dataset_properties, include=include, exclude=exclude) if len(available_preprocessors) == 0: raise ValueError( "No preprocessors found, please add NoPreprocessing") if default is None: defaults = ['no_preprocessing', 'select_percentile', 'pca', 'truncatedSVD'] for default_ in defaults: if default_ in available_preprocessors: default = default_ break preprocessor = CategoricalHyperparameter('__choice__', list( available_preprocessors.keys()), default=default) cs.add_hyperparameter(preprocessor) for name in available_preprocessors: preprocessor_configuration_space = available_preprocessors[name]. \ get_hyperparameter_search_space(dataset_properties) for parameter in preprocessor_configuration_space.get_hyperparameters(): new_parameter = copy.deepcopy(parameter) new_parameter.name = "%s:%s" % (name, new_parameter.name) cs.add_hyperparameter(new_parameter) # We must only add a condition if the hyperparameter is not # conditional on something else if len(preprocessor_configuration_space. get_parents_of(parameter)) == 0: condition = EqualsCondition(new_parameter, preprocessor, name) cs.add_condition(condition) for condition in available_preprocessors[name]. \ get_hyperparameter_search_space( dataset_properties).get_conditions(): if not isinstance(condition, AbstractConjunction): dlcs = [condition] else: dlcs = condition.get_descendent_literal_conditions() for dlc in dlcs: if not dlc.child.name.startswith(name): dlc.child.name = "%s:%s" % (name, dlc.child.name) if not dlc.parent.name.startswith(name): dlc.parent.name = "%s:%s" % (name, dlc.parent.name) cs.add_condition(condition) for forbidden_clause in available_preprocessors[name]. \ get_hyperparameter_search_space( dataset_properties).forbidden_clauses: dlcs = forbidden_clause.get_descendant_literal_clauses() for dlc in dlcs: if not dlc.hyperparameter.name.startswith(name): dlc.hyperparameter.name = "%s:%s" % (name, dlc.hyperparameter.name) cs.add_forbidden_clause(forbidden_clause) return cs
def get_hyperparameter_search_space(cls, dataset_properties, default=None, include=None, exclude=None): if include is not None and exclude is not None: raise ValueError("The argument include and exclude cannot be used together.") cs = ConfigurationSpace() # Compile a list of all estimator objects for this problem available_estimators = cls.get_available_components( data_prop=dataset_properties, include=include, exclude=exclude) if len(available_estimators) == 0: raise ValueError("No regressors found") if default is None: defaults = ['random_forest', 'support_vector_regression'] + \ list(available_estimators.keys()) for default_ in defaults: if default_ in available_estimators: if include is not None and default_ not in include: continue if exclude is not None and default_ in exclude: continue default = default_ break estimator = CategoricalHyperparameter('__choice__', list(available_estimators.keys()), default=default) cs.add_hyperparameter(estimator) for estimator_name in available_estimators.keys(): # We have to retrieve the configuration space every time because # we change the objects it returns. If we reused it, we could not # retrieve the conditions further down # TODO implement copy for hyperparameters and forbidden and # conditions! estimator_configuration_space = available_estimators[ estimator_name]. \ get_hyperparameter_search_space(dataset_properties) for parameter in estimator_configuration_space.get_hyperparameters(): new_parameter = copy.deepcopy(parameter) new_parameter.name = "%s:%s" % ( estimator_name, new_parameter.name) cs.add_hyperparameter(new_parameter) # We must only add a condition if the hyperparameter is not # conditional on something else if len(estimator_configuration_space. get_parents_of(parameter)) == 0: condition = EqualsCondition(new_parameter, estimator, estimator_name) cs.add_condition(condition) for condition in available_estimators[estimator_name]. \ get_hyperparameter_search_space( dataset_properties).get_conditions(): dlcs = condition.get_descendant_literal_conditions() for dlc in dlcs: if not dlc.child.name.startswith(estimator_name): dlc.child.name = "%s:%s" % ( estimator_name, dlc.child.name) if not dlc.parent.name.startswith(estimator_name): dlc.parent.name = "%s:%s" % ( estimator_name, dlc.parent.name) cs.add_condition(condition) for forbidden_clause in available_estimators[estimator_name]. \ get_hyperparameter_search_space( dataset_properties).forbidden_clauses: dlcs = forbidden_clause.get_descendant_literal_clauses() for dlc in dlcs: if not dlc.hyperparameter.name.startswith(estimator_name): dlc.hyperparameter.name = "%s:%s" % (estimator_name, dlc.hyperparameter.name) cs.add_forbidden_clause(forbidden_clause) return cs