def train_decoupled_tree(info): assert info['method_name'] in DECOUPLED_TREE_METHODS data, cvindices = load_processed_data(info['data_file']) data = split_data_by_cvindices(data, cvindices, fold_id=info['fold_id'], fold_num=1, fold_num_test=2) data, groups = split_groups_from_data(data=data, group_names=data['partitions']) data = convert_remaining_groups_to_rules(data) data = cast_numeric_fields(data) # specify training handle if info['method_name'] == 'tree_01': data = add_intercept(data) training_handle = lambda data: train_zero_one_linear_model( data, time_limit=info['max_runtime']) elif info['method_name'] == 'tree_lr': training_handle = lambda data: train_logreg( data, settings=None, normalize_variables=False) else: training_handle = lambda data: train_svm( data, settings=None, normalize_variables=False) # train tree tree = DecouplingTree(data, groups, training_handle=training_handle) tree.grow(print_flag=True, log_flag=True) ### prune tree tree.prune(alpha=0.1, test_correction=True, atomic_groups=True, data=data, groups=groups, stat_field='test') tree.render() ### transform to classifier tree clf_set = DecoupledClassifierSet( data=data, groups=groups, pooled_model=tree.root.model, decoupled_models=[l.model for l in tree.partition.leaves], groups_to_models=build_model_assignment_map(tree.partition)) ### save results info.update({ 'clf_set': clf_set, 'debug': { 'tree': tree }, }) return info
def train_decoupled_models(info): assert info['method_name'] in DECOUPLED_CLF_METHODS data, cvindices = load_processed_data(info['data_file']) data = split_data_by_cvindices(data, cvindices, fold_id=info['fold_id'], fold_num=1, fold_num_test=-1) data, groups = split_groups_from_data(data=data, group_names=data['partitions']) data = convert_remaining_groups_to_rules(data) data = cast_numeric_fields(data) # group stuff splits = groups_to_splits(groups, drop_missing=True) group_names, group_values = groups_to_group_data(groups, stat_field='train') split_values, group_indicators = np.unique(group_values, axis=0, return_inverse=True) # set method if info['method_name'] == 'dcp_svm': training_handle = lambda data: train_svm( data, settings=None, normalize_variables=False) else: #info['method_name'] == 'dcp_lr': training_handle = lambda data: train_logreg( data, settings=None, normalize_variables=False) if info['attr_id'] == 'none': S = np.zeros_like(group_indicators) elif info['attr_id'] == 'all': S = group_indicators sensitive_splits = split_values else: matched_attr = np.flatnonzero( [info['attr_id'].lower() == g.lower() for g in group_names]) sensitive_splits, S = np.unique(group_values[:, matched_attr], axis=0, return_inverse=True) # train pooled classifier pooled_model = training_handle({ 'X': data['X'], 'Y': data['Y'], 'variable_names': data['variable_names'] }) # train decoupled classifiers model_ids = np.unique(S) groups_to_models = {} model_dict = {} if info['attr_id'] == 'none': model_dict[0] = deepcopy(pooled_model) for k, s in enumerate(splits): groups_to_models[s] = 0 elif info['attr_id'] == 'all': for group_id in model_ids: idx = S == group_id split = tuple(zip(group_names, sensitive_splits[group_id])) groups_to_models[split] = group_id model = training_handle({ 'X': data['X'][idx, :], 'Y': data['Y'][idx], 'variable_names': data['variable_names'] }) model_dict[group_id] = model else: for model_id, s in enumerate(sensitive_splits): idx = S == model_id model_dict[model_id] = training_handle({ 'X': data['X'][idx, :], 'Y': data['Y'][idx], 'variable_names': data['variable_names'] }) assignment_idx = np.isin(split_values[:, matched_attr], s).flatten() matched_full_values = split_values[assignment_idx, :] for vals in matched_full_values: split = tuple([(g, z) for g, z in zip(group_names, vals)]) groups_to_models[split] = model_id decoupled_models = [model_dict[k] for k in range(len(model_dict))] assert check_model_assignment(groups_to_models, groups, decoupled_models) clf_set = DecoupledClassifierSet(data=data, groups=groups, pooled_model=pooled_model, decoupled_models=decoupled_models, groups_to_models=groups_to_models) info.update({ 'clf_set': clf_set, }) return info
def train_onehot_model(info): """ trains linear classifier with a one-hot encoding :param info: :return: """ assert info['method_name'] in ONEHOT_CLF_METHODS data, cvindices = load_processed_data(info['data_file']) data = split_data_by_cvindices(data, cvindices, fold_id=info['fold_id'], fold_num=1, fold_num_test=-1) data, groups = split_groups_from_data(data=data, group_names=data['partitions']) data = convert_remaining_groups_to_rules(data) data = cast_numeric_fields(data) # group stuff splits = groups_to_splits(groups, drop_missing=True) group_names, group_values = groups_to_group_data(groups, stat_field='train') split_values, group_indicators = np.unique(group_values, axis=0, return_inverse=True) if info['method_name'] == 'onehot_svm': training_handle = lambda data: train_svm( data, settings=None, normalize_variables=False) else: #info['method_name'] == 'onehot_lr': training_handle = lambda data: train_logreg( data, settings=None, normalize_variables=False) if info['attr_id'] == 'none': S = np.zeros_like(group_indicators) elif info['attr_id'] == 'all': S = group_indicators training_splits = split_values else: matched_attr = np.flatnonzero( [info['attr_id'].lower() == g.lower() for g in group_names]) training_splits, S = np.unique(group_values[:, matched_attr], axis=0, return_inverse=True) # train blind classifier pooled_model = training_handle({ 'X': data['X'], 'Y': data['Y'], 'variable_names': data['variable_names'] }) # train group-specific models model_dict = {} groups_to_models = {} if info['attr_id'] == 'none': model_dict[0] = deepcopy(pooled_model) for k, s in enumerate(splits): groups_to_models[s] = 0 else: coefficient_idx = get_variable_indices(data, include_intercept=False) # find majority group model_ids, subgroup_counts = np.unique(S, return_counts=True) majority_group_id = np.argmax(subgroup_counts) minority_group_ids = np.delete(model_ids, majority_group_id) # create group indicator variables Z = np.vstack([np.isin(S, s) for s in model_ids]).transpose().astype('float') assert np.isclose(Z.sum(axis=1), 1.0).all() Z = Z[:, minority_group_ids] Z_names = [ 'Subgroup_is_%s' % ''.join(training_splits[s]) for s in minority_group_ids ] ZX = np.concatenate((Z, data['X']), axis=1) ZX_names = Z_names + data['variable_names'] # setup parameters for model object baseline_model = training_handle({ 'X': ZX, 'Y': data['Y'], 'variable_names': ZX_names }) baseline_intercept = float(baseline_model._intercept) baseline_coefficients = np.array( baseline_model._coefficients).flatten() coef_idx = np.arange(len(ZX_names)) Z_coef_idx = np.arange(len(Z_names)) X_coef_idx = np.setdiff1d(coef_idx, Z_coef_idx) baseline_coefs = np.copy(baseline_coefficients[X_coef_idx]) subgroup_coefs = np.copy(baseline_coefficients[Z_coef_idx]) subgroup_coefs = np.insert(arr=subgroup_coefs, obj=majority_group_id, values=0.0) if info['attr_id'] == 'all': subgroup_intercepts = np.array(subgroup_coefs) else: subgroup_intercepts = np.repeat(np.nan, len(splits)) for model_id, s in enumerate(training_splits): matched_subgroups = np.flatnonzero( np.isin(split_values[:, matched_attr], s).flatten()) if model_id == majority_group_id: subgroup_intercepts[matched_subgroups] = 0.0 else: subgroup_intercepts[matched_subgroups] = subgroup_coefs[ model_id] assert np.isfinite(subgroup_intercepts).all() assert np.any(subgroup_intercepts == 0.0) # create new models for each group for k, s in enumerate(splits): model_info = { 'intercept': float(baseline_intercept + subgroup_intercepts[k]), 'coefficients': np.copy(baseline_coefs), 'coefficient_idx': coefficient_idx, } training_info = dict(baseline_model.training_info) training_info['split'] = s model = ClassificationModel(predict_handle=lambda X: np.sign( X.dot(model_info['coefficients']) + model_info['intercept']), model_type=ClassificationModel. LINEAR_MODEL_TYPE, model_info=model_info, training_info=training_info) model_dict[k] = model groups_to_models[s] = k decoupled_models = [model_dict[k] for k in range(len(model_dict))] assert check_model_assignment(groups_to_models, groups, decoupled_models) clf_set = DecoupledClassifierSet(data=data, groups=groups, pooled_model=pooled_model, decoupled_models=decoupled_models, groups_to_models=groups_to_models) info.update({'clf_set': clf_set}) return info
def train_coupled_model(info): assert info['method_name'] in COUPLED_CLF_METHODS data, cvindices = load_processed_data(info['data_file']) data = split_data_by_cvindices(data, cvindices, fold_id=info['fold_id'], fold_num=1, fold_num_test=-1) data, groups = split_groups_from_data(data=data, group_names=data['partitions']) data = convert_remaining_groups_to_rules(data) data = cast_numeric_fields(data) # group stuff splits = groups_to_splits(groups, drop_missing=True) group_names, group_values = groups_to_group_data(groups, stat_field='train') split_values, group_indicators = np.unique(group_values, axis=0, return_inverse=True) # add intercept data = add_intercept(data) method_specs = info['method_name'].split('_') pooled_model_type = method_specs[1] if method_specs[2] == 'svm': loss_function = CoupledRiskMinimizer.LOSS_SVM else: loss_function = CoupledRiskMinimizer.LOSS_LOGISTIC pooled_params = { 'EPS': 1e-3, 'cons_type': CoupledRiskMinimizer.CONSTRAINT_PARITY if pooled_model_type == 'parity' else CoupledRiskMinimizer.CONSTRAINT_NONE, } decoupled_params = cons_params = { 'cons_type': CoupledRiskMinimizer.CONSTRAINT_PREFERED_BOTH, 'tau': 0.1, 'print_flag': True } if info['attr_id'] == 'all': S = group_indicators sensitive_splits = split_values elif info['attr_id'] == 'none': S = np.zeros_like(group_indicators) else: matched_attr = np.flatnonzero( [info['attr_id'].lower() == g.lower() for g in group_names]) sensitive_splits, S = np.unique(group_values[:, matched_attr], axis=0, return_inverse=True) # train pooled classifier pooled_clf = CoupledRiskMinimizer(loss_function, lam=1e-5, train_multiple=False, sparse_formulation=True) pooled_clf.fit(X=data['X'], y=data['Y'], x_sensitive=S, cons_params=pooled_params) pooled_model = pooled_clf.classifier() debug = {'pooled_clf': pooled_clf} # train preference based classifier groups_to_models = {} if info['attr_id'] == 'none': for s in splits: groups_to_models[tuple(s)] = 0 decoupled_models = [deepcopy(pooled_model)] else: _, dist_dict = pooled_clf.get_distance_boundary(X=data['X'], x_sensitive=S) switch_margins = pooled_clf.switch_margins(dist_dict) lam = {k: 1e-3 for k in np.unique(S)} decoupled_params['s_val_to_cons_sum'] = switch_margins clf = CoupledRiskMinimizer(loss_function, lam=lam, train_multiple=True, sparse_formulation=True) clf.fit(X=data['X'], y=data['Y'], x_sensitive=S, cons_params=cons_params) group_models = clf.classifier() if info['attr_id'] == 'all': for group_id in group_models.keys(): split = tuple(zip(group_names, sensitive_splits[group_id])) groups_to_models[split] = group_id else: for s in split_values: # determine split split = tuple([(g, z) for g, z in zip(group_names, s)]) # find model sensitive_values = s[matched_attr] model_id = int( np.flatnonzero(sensitive_values == sensitive_splits)) groups_to_models[split] = model_id # build classifier sets decoupled_models = [group_models[i] for i in range(len(group_models))] for k, w in clf.w.items(): assert np.isclose(decoupled_models[k].coefficients, w[1:]).all() assert np.isclose(decoupled_models[k].intercept, w[0]) debug.update({'clf': clf}) assert check_model_assignment(groups_to_models, groups, decoupled_models) clf_set = DecoupledClassifierSet(data=data, groups=groups, pooled_model=pooled_model, decoupled_models=decoupled_models, groups_to_models=groups_to_models) info.update({'clf_set': clf_set, 'debug': debug}) return info
from dcptree.data_io import load_processed_data from dcptree.cross_validation import filter_data_to_fold from dcptree.group_helper import * from dcptree.classification_models import * from dcptree.tree import * #directories data_name = 'adult' format_label = 'envyfree' random_seed = 1337 data_dir = repo_dir + 'data/' selected_groups = ['Sex'] ## load data data_file = '%s%s_processed.pickle' % (data_dir, data_name) data, cvindices = load_processed_data(data_file) # filter to fold data = filter_data_to_fold(data, cvindices, fold_id='K05N01', fold_num=0, include_validation=True) # remove selected groups data, groups = split_groups_from_data(data=data, group_names=selected_groups) data = convert_remaining_groups_to_rules(data) data = cast_numeric_fields(data) training_handle = lambda data: train_logreg( data, settings=None, normalize_variables=False) #tree = EnvyFreeTree(data, groups, training_handle = training_handle, splitting_handle = split_test)