Mercurial > repos > bgruening > sklearn_estimator_attributes
view utils.py @ 0:2ad4c2798be7 draft
planemo upload for repository https://github.com/bgruening/galaxytools/tree/master/tools/sklearn commit c0a3a186966888e5787335a7628bf0a4382637e7
author | bgruening |
---|---|
date | Tue, 14 May 2019 18:12:53 -0400 |
parents | |
children |
line wrap: on
line source
import ast import json import imblearn import numpy as np import pandas import pickle import re import scipy import sklearn import skrebate import sys import warnings import xgboost from collections import Counter from asteval import Interpreter, make_symbol_table from imblearn import under_sampling, over_sampling, combine from imblearn.pipeline import Pipeline as imbPipeline from mlxtend import regressor, classifier from scipy.io import mmread from sklearn import ( cluster, compose, decomposition, ensemble, feature_extraction, feature_selection, gaussian_process, kernel_approximation, metrics, model_selection, naive_bayes, neighbors, pipeline, preprocessing, svm, linear_model, tree, discriminant_analysis) try: import iraps_classifier except ImportError: pass try: import model_validations except ImportError: pass try: import feature_selectors except ImportError: pass try: import preprocessors except ImportError: pass # handle pickle white list file WL_FILE = __import__('os').path.join( __import__('os').path.dirname(__file__), 'pk_whitelist.json') N_JOBS = int(__import__('os').environ.get('GALAXY_SLOTS', 1)) class _SafePickler(pickle.Unpickler, object): """ Used to safely deserialize scikit-learn model objects Usage: eg.: _SafePickler.load(pickled_file_object) """ def __init__(self, file): super(_SafePickler, self).__init__(file) # load global white list with open(WL_FILE, 'r') as f: self.pk_whitelist = json.load(f) self.bad_names = ( 'and', 'as', 'assert', 'break', 'class', 'continue', 'def', 'del', 'elif', 'else', 'except', 'exec', 'finally', 'for', 'from', 'global', 'if', 'import', 'in', 'is', 'lambda', 'not', 'or', 'pass', 'print', 'raise', 'return', 'try', 'system', 'while', 'with', 'True', 'False', 'None', 'eval', 'execfile', '__import__', '__package__', '__subclasses__', '__bases__', '__globals__', '__code__', '__closure__', '__func__', '__self__', '__module__', '__dict__', '__class__', '__call__', '__get__', '__getattribute__', '__subclasshook__', '__new__', '__init__', 'func_globals', 'func_code', 'func_closure', 'im_class', 'im_func', 'im_self', 'gi_code', 'gi_frame', '__asteval__', 'f_locals', '__mro__') # unclassified good globals self.good_names = [ 'copy_reg._reconstructor', '__builtin__.object', '__builtin__.bytearray', 'builtins.object', 'builtins.bytearray', 'keras.engine.sequential.Sequential', 'keras.engine.sequential.Model'] # custom module in Galaxy-ML self.custom_modules = [ '__main__', 'keras_galaxy_models', 'feature_selectors', 'preprocessors', 'iraps_classifier', 'model_validations'] # override def find_class(self, module, name): # balack list first if name in self.bad_names: raise pickle.UnpicklingError("global '%s.%s' is forbidden" % (module, name)) # custom module in Galaxy-ML if module in self.custom_modules: cutom_module = sys.modules.get(module, None) if cutom_module: return getattr(cutom_module, name) else: raise pickle.UnpicklingError("Module %s' is not imported" % module) # For objects from outside libraries, it's necessary to verify # both module and name. Currently only a blacklist checker # is working. # TODO: replace with a whitelist checker. good_names = self.good_names pk_whitelist = self.pk_whitelist if re.match(r'^[a-zA-Z_][a-zA-Z0-9_]*$', name): fullname = module + '.' + name if (fullname in good_names)\ or (module.startswith(('sklearn.', 'xgboost.', 'skrebate.', 'imblearn.', 'mlxtend.', 'numpy.')) or module == 'numpy'): if fullname not in (pk_whitelist['SK_NAMES'] + pk_whitelist['SKR_NAMES'] + pk_whitelist['XGB_NAMES'] + pk_whitelist['NUMPY_NAMES'] + pk_whitelist['IMBLEARN_NAMES'] + pk_whitelist['MLXTEND_NAMES'] + good_names): # raise pickle.UnpicklingError print("Warning: global %s is not in pickler whitelist " "yet and will loss support soon. Contact tool " "author or leave a message at github.com" % fullname) mod = sys.modules[module] return getattr(mod, name) raise pickle.UnpicklingError("global '%s' is forbidden" % fullname) def load_model(file): """Load pickled object with `_SafePicker` """ return _SafePickler(file).load() def read_columns(f, c=None, c_option='by_index_number', return_df=False, **args): """Return array from a tabular dataset by various columns selection """ data = pandas.read_csv(f, **args) if c_option == 'by_index_number': cols = list(map(lambda x: x - 1, c)) data = data.iloc[:, cols] if c_option == 'all_but_by_index_number': cols = list(map(lambda x: x - 1, c)) data.drop(data.columns[cols], axis=1, inplace=True) if c_option == 'by_header_name': cols = [e.strip() for e in c.split(',')] data = data[cols] if c_option == 'all_but_by_header_name': cols = [e.strip() for e in c.split(',')] data.drop(cols, axis=1, inplace=True) y = data.values if return_df: return y, data else: return y def feature_selector(inputs, X=None, y=None): """generate an instance of sklearn.feature_selection classes Parameters ---------- inputs : dict From galaxy tool parameters. X : array Containing training features. y : array or list Target values. """ selector = inputs['selected_algorithm'] if selector != 'DyRFECV': selector = getattr(sklearn.feature_selection, selector) options = inputs['options'] if inputs['selected_algorithm'] == 'SelectFromModel': if not options['threshold'] or options['threshold'] == 'None': options['threshold'] = None else: try: options['threshold'] = float(options['threshold']) except ValueError: pass if inputs['model_inputter']['input_mode'] == 'prefitted': model_file = inputs['model_inputter']['fitted_estimator'] with open(model_file, 'rb') as model_handler: fitted_estimator = load_model(model_handler) new_selector = selector(fitted_estimator, prefit=True, **options) else: estimator_json = inputs['model_inputter']['estimator_selector'] estimator = get_estimator(estimator_json) check_feature_importances = try_get_attr( 'feature_selectors', 'check_feature_importances') estimator = check_feature_importances(estimator) new_selector = selector(estimator, **options) elif inputs['selected_algorithm'] == 'RFE': step = options.get('step', None) if step and step >= 1.0: options['step'] = int(step) estimator = get_estimator(inputs["estimator_selector"]) check_feature_importances = try_get_attr( 'feature_selectors', 'check_feature_importances') estimator = check_feature_importances(estimator) new_selector = selector(estimator, **options) elif inputs['selected_algorithm'] == 'RFECV': options['scoring'] = get_scoring(options['scoring']) options['n_jobs'] = N_JOBS splitter, groups = get_cv(options.pop('cv_selector')) if groups is None: options['cv'] = splitter else: options['cv'] = list(splitter.split(X, y, groups=groups)) step = options.get('step', None) if step and step >= 1.0: options['step'] = int(step) estimator = get_estimator(inputs['estimator_selector']) check_feature_importances = try_get_attr( 'feature_selectors', 'check_feature_importances') estimator = check_feature_importances(estimator) new_selector = selector(estimator, **options) elif inputs['selected_algorithm'] == 'DyRFECV': options['scoring'] = get_scoring(options['scoring']) options['n_jobs'] = N_JOBS splitter, groups = get_cv(options.pop('cv_selector')) if groups is None: options['cv'] = splitter else: options['cv'] = list(splitter.split(X, y, groups=groups)) step = options.get('step') if not step or step == 'None': step = None else: step = ast.literal_eval(step) options['step'] = step estimator = get_estimator(inputs["estimator_selector"]) check_feature_importances = try_get_attr( 'feature_selectors', 'check_feature_importances') estimator = check_feature_importances(estimator) DyRFECV = try_get_attr('feature_selectors', 'DyRFECV') new_selector = DyRFECV(estimator, **options) elif inputs['selected_algorithm'] == 'VarianceThreshold': new_selector = selector(**options) else: score_func = inputs['score_func'] score_func = getattr(sklearn.feature_selection, score_func) new_selector = selector(score_func, **options) return new_selector def get_X_y(params, file1, file2): """Return machine learning inputs X, y from tabluar inputs """ input_type = (params['selected_tasks']['selected_algorithms'] ['input_options']['selected_input']) if input_type == 'tabular': header = 'infer' if (params['selected_tasks']['selected_algorithms'] ['input_options']['header1']) else None column_option = (params['selected_tasks']['selected_algorithms'] ['input_options']['column_selector_options_1'] ['selected_column_selector_option']) if column_option in ['by_index_number', 'all_but_by_index_number', 'by_header_name', 'all_but_by_header_name']: c = (params['selected_tasks']['selected_algorithms'] ['input_options']['column_selector_options_1']['col1']) else: c = None X = read_columns( file1, c=c, c_option=column_option, sep='\t', header=header, parse_dates=True).astype(float) else: X = mmread(file1) header = 'infer' if (params['selected_tasks']['selected_algorithms'] ['input_options']['header2']) else None column_option = (params['selected_tasks']['selected_algorithms'] ['input_options']['column_selector_options_2'] ['selected_column_selector_option2']) if column_option in ['by_index_number', 'all_but_by_index_number', 'by_header_name', 'all_but_by_header_name']: c = (params['selected_tasks']['selected_algorithms'] ['input_options']['column_selector_options_2']['col2']) else: c = None y = read_columns( file2, c=c, c_option=column_option, sep='\t', header=header, parse_dates=True) y = y.ravel() return X, y class SafeEval(Interpreter): """Customized symbol table for safely literal eval """ def __init__(self, load_scipy=False, load_numpy=False, load_estimators=False): # File opening and other unneeded functions could be dropped unwanted = ['open', 'type', 'dir', 'id', 'str', 'repr'] # Allowed symbol table. Add more if needed. new_syms = { 'np_arange': getattr(np, 'arange'), 'ensemble_ExtraTreesClassifier': getattr(ensemble, 'ExtraTreesClassifier') } syms = make_symbol_table(use_numpy=False, **new_syms) if load_scipy: scipy_distributions = scipy.stats.distributions.__dict__ for k, v in scipy_distributions.items(): if isinstance(v, (scipy.stats.rv_continuous, scipy.stats.rv_discrete)): syms['scipy_stats_' + k] = v if load_numpy: from_numpy_random = [ 'beta', 'binomial', 'bytes', 'chisquare', 'choice', 'dirichlet', 'division', 'exponential', 'f', 'gamma', 'geometric', 'gumbel', 'hypergeometric', 'laplace', 'logistic', 'lognormal', 'logseries', 'mtrand', 'multinomial', 'multivariate_normal', 'negative_binomial', 'noncentral_chisquare', 'noncentral_f', 'normal', 'pareto', 'permutation', 'poisson', 'power', 'rand', 'randint', 'randn', 'random', 'random_integers', 'random_sample', 'ranf', 'rayleigh', 'sample', 'seed', 'set_state', 'shuffle', 'standard_cauchy', 'standard_exponential', 'standard_gamma', 'standard_normal', 'standard_t', 'triangular', 'uniform', 'vonmises', 'wald', 'weibull', 'zipf'] for f in from_numpy_random: syms['np_random_' + f] = getattr(np.random, f) if load_estimators: estimator_table = { 'sklearn_svm': getattr(sklearn, 'svm'), 'sklearn_tree': getattr(sklearn, 'tree'), 'sklearn_ensemble': getattr(sklearn, 'ensemble'), 'sklearn_neighbors': getattr(sklearn, 'neighbors'), 'sklearn_naive_bayes': getattr(sklearn, 'naive_bayes'), 'sklearn_linear_model': getattr(sklearn, 'linear_model'), 'sklearn_cluster': getattr(sklearn, 'cluster'), 'sklearn_decomposition': getattr(sklearn, 'decomposition'), 'sklearn_preprocessing': getattr(sklearn, 'preprocessing'), 'sklearn_feature_selection': getattr(sklearn, 'feature_selection'), 'sklearn_kernel_approximation': getattr(sklearn, 'kernel_approximation'), 'skrebate_ReliefF': getattr(skrebate, 'ReliefF'), 'skrebate_SURF': getattr(skrebate, 'SURF'), 'skrebate_SURFstar': getattr(skrebate, 'SURFstar'), 'skrebate_MultiSURF': getattr(skrebate, 'MultiSURF'), 'skrebate_MultiSURFstar': getattr(skrebate, 'MultiSURFstar'), 'skrebate_TuRF': getattr(skrebate, 'TuRF'), 'xgboost_XGBClassifier': getattr(xgboost, 'XGBClassifier'), 'xgboost_XGBRegressor': getattr(xgboost, 'XGBRegressor'), 'imblearn_over_sampling': getattr(imblearn, 'over_sampling'), 'imblearn_combine': getattr(imblearn, 'combine') } syms.update(estimator_table) for key in unwanted: syms.pop(key, None) super(SafeEval, self).__init__( symtable=syms, use_numpy=False, minimal=False, no_if=True, no_for=True, no_while=True, no_try=True, no_functiondef=True, no_ifexp=True, no_listcomp=False, no_augassign=False, no_assert=True, no_delete=True, no_raise=True, no_print=True) def get_estimator(estimator_json): """Return a sklearn or compatible estimator from Galaxy tool inputs """ estimator_module = estimator_json['selected_module'] if estimator_module == 'custom_estimator': c_estimator = estimator_json['c_estimator'] with open(c_estimator, 'rb') as model_handler: new_model = load_model(model_handler) return new_model if estimator_module == "binarize_target": wrapped_estimator = estimator_json['wrapped_estimator'] with open(wrapped_estimator, 'rb') as model_handler: wrapped_estimator = load_model(model_handler) options = {} if estimator_json['z_score'] is not None: options['z_score'] = estimator_json['z_score'] if estimator_json['value'] is not None: options['value'] = estimator_json['value'] options['less_is_positive'] = estimator_json['less_is_positive'] if estimator_json['clf_or_regr'] == 'BinarizeTargetClassifier': klass = try_get_attr('iraps_classifier', 'BinarizeTargetClassifier') else: klass = try_get_attr('iraps_classifier', 'BinarizeTargetRegressor') return klass(wrapped_estimator, **options) estimator_cls = estimator_json['selected_estimator'] if estimator_module == 'xgboost': klass = getattr(xgboost, estimator_cls) else: module = getattr(sklearn, estimator_module) klass = getattr(module, estimator_cls) estimator = klass() estimator_params = estimator_json['text_params'].strip() if estimator_params != '': try: safe_eval = SafeEval() params = safe_eval('dict(' + estimator_params + ')') except ValueError: sys.exit("Unsupported parameter input: `%s`" % estimator_params) estimator.set_params(**params) if 'n_jobs' in estimator.get_params(): estimator.set_params(n_jobs=N_JOBS) return estimator def get_cv(cv_json): """ Return CV splitter from Galaxy tool inputs Parameters ---------- cv_json : dict From Galaxy tool inputs. e.g.: { 'selected_cv': 'StratifiedKFold', 'n_splits': 3, 'shuffle': True, 'random_state': 0 } """ cv = cv_json.pop('selected_cv') if cv == 'default': return cv_json['n_splits'], None groups = cv_json.pop('groups_selector', None) if groups is not None: infile_g = groups['infile_g'] header = 'infer' if groups['header_g'] else None column_option = (groups['column_selector_options_g'] ['selected_column_selector_option_g']) if column_option in ['by_index_number', 'all_but_by_index_number', 'by_header_name', 'all_but_by_header_name']: c = groups['column_selector_options_g']['col_g'] else: c = None groups = read_columns( infile_g, c=c, c_option=column_option, sep='\t', header=header, parse_dates=True) groups = groups.ravel() for k, v in cv_json.items(): if v == '': cv_json[k] = None test_fold = cv_json.get('test_fold', None) if test_fold: if test_fold.startswith('__ob__'): test_fold = test_fold[6:] if test_fold.endswith('__cb__'): test_fold = test_fold[:-6] cv_json['test_fold'] = [int(x.strip()) for x in test_fold.split(',')] test_size = cv_json.get('test_size', None) if test_size and test_size > 1.0: cv_json['test_size'] = int(test_size) if cv == 'OrderedKFold': cv_class = try_get_attr('model_validations', 'OrderedKFold') elif cv == 'RepeatedOrderedKFold': cv_class = try_get_attr('model_validations', 'RepeatedOrderedKFold') else: cv_class = getattr(model_selection, cv) splitter = cv_class(**cv_json) return splitter, groups # needed when sklearn < v0.20 def balanced_accuracy_score(y_true, y_pred): """Compute balanced accuracy score, which is now available in scikit-learn from v0.20.0. """ C = metrics.confusion_matrix(y_true, y_pred) with np.errstate(divide='ignore', invalid='ignore'): per_class = np.diag(C) / C.sum(axis=1) if np.any(np.isnan(per_class)): warnings.warn('y_pred contains classes not in y_true') per_class = per_class[~np.isnan(per_class)] score = np.mean(per_class) return score def get_scoring(scoring_json): """Return single sklearn scorer class or multiple scoers in dictionary """ if scoring_json['primary_scoring'] == 'default': return None my_scorers = metrics.SCORERS my_scorers['binarize_auc_scorer'] =\ try_get_attr('iraps_classifier', 'binarize_auc_scorer') my_scorers['binarize_average_precision_scorer'] =\ try_get_attr('iraps_classifier', 'binarize_average_precision_scorer') if 'balanced_accuracy' not in my_scorers: my_scorers['balanced_accuracy'] =\ metrics.make_scorer(balanced_accuracy_score) if scoring_json['secondary_scoring'] != 'None'\ and scoring_json['secondary_scoring'] !=\ scoring_json['primary_scoring']: return_scoring = {} primary_scoring = scoring_json['primary_scoring'] return_scoring[primary_scoring] = my_scorers[primary_scoring] for scorer in scoring_json['secondary_scoring'].split(','): if scorer != scoring_json['primary_scoring']: return_scoring[scorer] = my_scorers[scorer] return return_scoring return my_scorers[scoring_json['primary_scoring']] def get_search_params(estimator): """Format the output of `estimator.get_params()` """ params = estimator.get_params() results = [] for k, v in params.items(): # params below won't be shown for search in the searchcv tool keywords = ('n_jobs', 'pre_dispatch', 'memory', 'steps', 'nthread', 'verbose') if k.endswith(keywords): results.append(['*', k, k+": "+repr(v)]) else: results.append(['@', k, k+": "+repr(v)]) results.append( ["", "Note:", "@, params eligible for search in searchcv tool."]) return results def try_get_attr(module, name): """try to get attribute from a custom module Parameters ---------- module : str Module name name : str Attribute (class/function) name. Returns ------- class or function """ mod = sys.modules.get(module, None) if mod: return getattr(mod, name) else: raise Exception("No module named %s." % module)