Mercurial > repos > bgruening > stacking_ensemble_models

comparison utils.py @ 0:8e93241d5d28 draft default tip

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
planemo upload for repository https://github.com/bgruening/galaxytools/tree/master/tools/sklearn commit c0a3a186966888e5787335a7628bf0a4382637e7
author bgruening
date Tue, 14 May 2019 18:04:46 -0400
parents
children
comparison
equal deleted inserted replaced
-1:000000000000 0:8e93241d5d28
1 import ast
2 import json
3 import imblearn
4 import numpy as np
5 import pandas
6 import pickle
7 import re
8 import scipy
9 import sklearn
10 import skrebate
11 import sys
12 import warnings
13 import xgboost
14
15 from collections import Counter
16 from asteval import Interpreter, make_symbol_table
17 from imblearn import under_sampling, over_sampling, combine
18 from imblearn.pipeline import Pipeline as imbPipeline
19 from mlxtend import regressor, classifier
20 from scipy.io import mmread
21 from sklearn import (
22 cluster, compose, decomposition, ensemble, feature_extraction,
23 feature_selection, gaussian_process, kernel_approximation, metrics,
24 model_selection, naive_bayes, neighbors, pipeline, preprocessing,
25 svm, linear_model, tree, discriminant_analysis)
26
27 try:
28 import iraps_classifier
29 except ImportError:
30 pass
31
32 try:
33 import model_validations
34 except ImportError:
35 pass
36
37 try:
38 import feature_selectors
39 except ImportError:
40 pass
41
42 try:
43 import preprocessors
44 except ImportError:
45 pass
46
47 # handle pickle white list file
48 WL_FILE = __import__('os').path.join(
49 __import__('os').path.dirname(__file__), 'pk_whitelist.json')
50
51 N_JOBS = int(__import__('os').environ.get('GALAXY_SLOTS', 1))
52
53
54 class _SafePickler(pickle.Unpickler, object):
55 """
56 Used to safely deserialize scikit-learn model objects
57 Usage:
58 eg.: _SafePickler.load(pickled_file_object)
59 """
60 def __init__(self, file):
61 super(_SafePickler, self).__init__(file)
62 # load global white list
63 with open(WL_FILE, 'r') as f:
64 self.pk_whitelist = json.load(f)
65
66 self.bad_names = (
67 'and', 'as', 'assert', 'break', 'class', 'continue',
68 'def', 'del', 'elif', 'else', 'except', 'exec',
69 'finally', 'for', 'from', 'global', 'if', 'import',
70 'in', 'is', 'lambda', 'not', 'or', 'pass', 'print',
71 'raise', 'return', 'try', 'system', 'while', 'with',
72 'True', 'False', 'None', 'eval', 'execfile', '__import__',
73 '__package__', '__subclasses__', '__bases__', '__globals__',
74 '__code__', '__closure__', '__func__', '__self__', '__module__',
75 '__dict__', '__class__', '__call__', '__get__',
76 '__getattribute__', '__subclasshook__', '__new__',
77 '__init__', 'func_globals', 'func_code', 'func_closure',
78 'im_class', 'im_func', 'im_self', 'gi_code', 'gi_frame',
79 '__asteval__', 'f_locals', '__mro__')
80
81 # unclassified good globals
82 self.good_names = [
83 'copy_reg._reconstructor', '__builtin__.object',
84 '__builtin__.bytearray', 'builtins.object',
85 'builtins.bytearray', 'keras.engine.sequential.Sequential',
86 'keras.engine.sequential.Model']
87
88 # custom module in Galaxy-ML
89 self.custom_modules = [
90 '__main__', 'keras_galaxy_models', 'feature_selectors',
91 'preprocessors', 'iraps_classifier', 'model_validations']
92
93 # override
94 def find_class(self, module, name):
95 # balack list first
96 if name in self.bad_names:
97 raise pickle.UnpicklingError("global '%s.%s' is forbidden"
98 % (module, name))
99
100 # custom module in Galaxy-ML
101 if module in self.custom_modules:
102 cutom_module = sys.modules.get(module, None)
103 if cutom_module:
104 return getattr(cutom_module, name)
105 else:
106 raise pickle.UnpicklingError("Module %s' is not imported"
107 % module)
108
109 # For objects from outside libraries, it's necessary to verify
110 # both module and name. Currently only a blacklist checker
111 # is working.
112 # TODO: replace with a whitelist checker.
113 good_names = self.good_names
114 pk_whitelist = self.pk_whitelist
115 if re.match(r'^[a-zA-Z_][a-zA-Z0-9_]*$', name):
116 fullname = module + '.' + name
117 if (fullname in good_names)\
118 or (module.startswith(('sklearn.', 'xgboost.', 'skrebate.',
119 'imblearn.', 'mlxtend.', 'numpy.'))
120 or module == 'numpy'):
121 if fullname not in (pk_whitelist['SK_NAMES'] +
122 pk_whitelist['SKR_NAMES'] +
123 pk_whitelist['XGB_NAMES'] +
124 pk_whitelist['NUMPY_NAMES'] +
125 pk_whitelist['IMBLEARN_NAMES'] +
126 pk_whitelist['MLXTEND_NAMES'] +
127 good_names):
128 # raise pickle.UnpicklingError
129 print("Warning: global %s is not in pickler whitelist "
130 "yet and will loss support soon. Contact tool "
131 "author or leave a message at github.com" % fullname)
132 mod = sys.modules[module]
133 return getattr(mod, name)
134
135 raise pickle.UnpicklingError("global '%s' is forbidden" % fullname)
136
137
138 def load_model(file):
139 """Load pickled object with `_SafePicker`
140 """
141 return _SafePickler(file).load()
142
143
144 def read_columns(f, c=None, c_option='by_index_number',
145 return_df=False, **args):
146 """Return array from a tabular dataset by various columns selection
147 """
148 data = pandas.read_csv(f, **args)
149 if c_option == 'by_index_number':
150 cols = list(map(lambda x: x - 1, c))
151 data = data.iloc[:, cols]
152 if c_option == 'all_but_by_index_number':
153 cols = list(map(lambda x: x - 1, c))
154 data.drop(data.columns[cols], axis=1, inplace=True)
155 if c_option == 'by_header_name':
156 cols = [e.strip() for e in c.split(',')]
157 data = data[cols]
158 if c_option == 'all_but_by_header_name':
159 cols = [e.strip() for e in c.split(',')]
160 data.drop(cols, axis=1, inplace=True)
161 y = data.values
162 if return_df:
163 return y, data
164 else:
165 return y
166
167
168 def feature_selector(inputs, X=None, y=None):
169 """generate an instance of sklearn.feature_selection classes
170
171 Parameters
172 ----------
173 inputs : dict
174 From galaxy tool parameters.
175 X : array
176 Containing training features.
177 y : array or list
178 Target values.
179 """
180 selector = inputs['selected_algorithm']
181 if selector != 'DyRFECV':
182 selector = getattr(sklearn.feature_selection, selector)
183 options = inputs['options']
184
185 if inputs['selected_algorithm'] == 'SelectFromModel':
186 if not options['threshold'] or options['threshold'] == 'None':
187 options['threshold'] = None
188 else:
189 try:
190 options['threshold'] = float(options['threshold'])
191 except ValueError:
192 pass
193 if inputs['model_inputter']['input_mode'] == 'prefitted':
194 model_file = inputs['model_inputter']['fitted_estimator']
195 with open(model_file, 'rb') as model_handler:
196 fitted_estimator = load_model(model_handler)
197 new_selector = selector(fitted_estimator, prefit=True, **options)
198 else:
199 estimator_json = inputs['model_inputter']['estimator_selector']
200 estimator = get_estimator(estimator_json)
201 check_feature_importances = try_get_attr(
202 'feature_selectors', 'check_feature_importances')
203 estimator = check_feature_importances(estimator)
204 new_selector = selector(estimator, **options)
205
206 elif inputs['selected_algorithm'] == 'RFE':
207 step = options.get('step', None)
208 if step and step >= 1.0:
209 options['step'] = int(step)
210 estimator = get_estimator(inputs["estimator_selector"])
211 check_feature_importances = try_get_attr(
212 'feature_selectors', 'check_feature_importances')
213 estimator = check_feature_importances(estimator)
214 new_selector = selector(estimator, **options)
215
216 elif inputs['selected_algorithm'] == 'RFECV':
217 options['scoring'] = get_scoring(options['scoring'])
218 options['n_jobs'] = N_JOBS
219 splitter, groups = get_cv(options.pop('cv_selector'))
220 if groups is None:
221 options['cv'] = splitter
222 else:
223 options['cv'] = list(splitter.split(X, y, groups=groups))
224 step = options.get('step', None)
225 if step and step >= 1.0:
226 options['step'] = int(step)
227 estimator = get_estimator(inputs['estimator_selector'])
228 check_feature_importances = try_get_attr(
229 'feature_selectors', 'check_feature_importances')
230 estimator = check_feature_importances(estimator)
231 new_selector = selector(estimator, **options)
232
233 elif inputs['selected_algorithm'] == 'DyRFECV':
234 options['scoring'] = get_scoring(options['scoring'])
235 options['n_jobs'] = N_JOBS
236 splitter, groups = get_cv(options.pop('cv_selector'))
237 if groups is None:
238 options['cv'] = splitter
239 else:
240 options['cv'] = list(splitter.split(X, y, groups=groups))
241 step = options.get('step')
242 if not step or step == 'None':
243 step = None
244 else:
245 step = ast.literal_eval(step)
246 options['step'] = step
247 estimator = get_estimator(inputs["estimator_selector"])
248 check_feature_importances = try_get_attr(
249 'feature_selectors', 'check_feature_importances')
250 estimator = check_feature_importances(estimator)
251 DyRFECV = try_get_attr('feature_selectors', 'DyRFECV')
252
253 new_selector = DyRFECV(estimator, **options)
254
255 elif inputs['selected_algorithm'] == 'VarianceThreshold':
256 new_selector = selector(**options)
257
258 else:
259 score_func = inputs['score_func']
260 score_func = getattr(sklearn.feature_selection, score_func)
261 new_selector = selector(score_func, **options)
262
263 return new_selector
264
265
266 def get_X_y(params, file1, file2):
267 """Return machine learning inputs X, y from tabluar inputs
268 """
269 input_type = (params['selected_tasks']['selected_algorithms']
270 ['input_options']['selected_input'])
271 if input_type == 'tabular':
272 header = 'infer' if (params['selected_tasks']['selected_algorithms']
273 ['input_options']['header1']) else None
274 column_option = (params['selected_tasks']['selected_algorithms']
275 ['input_options']['column_selector_options_1']
276 ['selected_column_selector_option'])
277 if column_option in ['by_index_number', 'all_but_by_index_number',
278 'by_header_name', 'all_but_by_header_name']:
279 c = (params['selected_tasks']['selected_algorithms']
280 ['input_options']['column_selector_options_1']['col1'])
281 else:
282 c = None
283 X = read_columns(
284 file1,
285 c=c,
286 c_option=column_option,
287 sep='\t',
288 header=header,
289 parse_dates=True).astype(float)
290 else:
291 X = mmread(file1)
292
293 header = 'infer' if (params['selected_tasks']['selected_algorithms']
294 ['input_options']['header2']) else None
295 column_option = (params['selected_tasks']['selected_algorithms']
296 ['input_options']['column_selector_options_2']
297 ['selected_column_selector_option2'])
298 if column_option in ['by_index_number', 'all_but_by_index_number',
299 'by_header_name', 'all_but_by_header_name']:
300 c = (params['selected_tasks']['selected_algorithms']
301 ['input_options']['column_selector_options_2']['col2'])
302 else:
303 c = None
304 y = read_columns(
305 file2,
306 c=c,
307 c_option=column_option,
308 sep='\t',
309 header=header,
310 parse_dates=True)
311 y = y.ravel()
312
313 return X, y
314
315
316 class SafeEval(Interpreter):
317 """Customized symbol table for safely literal eval
318 """
319 def __init__(self, load_scipy=False, load_numpy=False,
320 load_estimators=False):
321
322 # File opening and other unneeded functions could be dropped
323 unwanted = ['open', 'type', 'dir', 'id', 'str', 'repr']
324
325 # Allowed symbol table. Add more if needed.
326 new_syms = {
327 'np_arange': getattr(np, 'arange'),
328 'ensemble_ExtraTreesClassifier':
329 getattr(ensemble, 'ExtraTreesClassifier')
330 }
331
332 syms = make_symbol_table(use_numpy=False, **new_syms)
333
334 if load_scipy:
335 scipy_distributions = scipy.stats.distributions.__dict__
336 for k, v in scipy_distributions.items():
337 if isinstance(v, (scipy.stats.rv_continuous,
338 scipy.stats.rv_discrete)):
339 syms['scipy_stats_' + k] = v
340
341 if load_numpy:
342 from_numpy_random = [
343 'beta', 'binomial', 'bytes', 'chisquare', 'choice',
344 'dirichlet', 'division', 'exponential', 'f', 'gamma',
345 'geometric', 'gumbel', 'hypergeometric', 'laplace',
346 'logistic', 'lognormal', 'logseries', 'mtrand',
347 'multinomial', 'multivariate_normal', 'negative_binomial',
348 'noncentral_chisquare', 'noncentral_f', 'normal', 'pareto',
349 'permutation', 'poisson', 'power', 'rand', 'randint',
350 'randn', 'random', 'random_integers', 'random_sample',
351 'ranf', 'rayleigh', 'sample', 'seed', 'set_state',
352 'shuffle', 'standard_cauchy', 'standard_exponential',
353 'standard_gamma', 'standard_normal', 'standard_t',
354 'triangular', 'uniform', 'vonmises', 'wald', 'weibull', 'zipf']
355 for f in from_numpy_random:
356 syms['np_random_' + f] = getattr(np.random, f)
357
358 if load_estimators:
359 estimator_table = {
360 'sklearn_svm': getattr(sklearn, 'svm'),
361 'sklearn_tree': getattr(sklearn, 'tree'),
362 'sklearn_ensemble': getattr(sklearn, 'ensemble'),
363 'sklearn_neighbors': getattr(sklearn, 'neighbors'),
364 'sklearn_naive_bayes': getattr(sklearn, 'naive_bayes'),
365 'sklearn_linear_model': getattr(sklearn, 'linear_model'),
366 'sklearn_cluster': getattr(sklearn, 'cluster'),
367 'sklearn_decomposition': getattr(sklearn, 'decomposition'),
368 'sklearn_preprocessing': getattr(sklearn, 'preprocessing'),
369 'sklearn_feature_selection':
370 getattr(sklearn, 'feature_selection'),
371 'sklearn_kernel_approximation':
372 getattr(sklearn, 'kernel_approximation'),
373 'skrebate_ReliefF': getattr(skrebate, 'ReliefF'),
374 'skrebate_SURF': getattr(skrebate, 'SURF'),
375 'skrebate_SURFstar': getattr(skrebate, 'SURFstar'),
376 'skrebate_MultiSURF': getattr(skrebate, 'MultiSURF'),
377 'skrebate_MultiSURFstar': getattr(skrebate, 'MultiSURFstar'),
378 'skrebate_TuRF': getattr(skrebate, 'TuRF'),
379 'xgboost_XGBClassifier': getattr(xgboost, 'XGBClassifier'),
380 'xgboost_XGBRegressor': getattr(xgboost, 'XGBRegressor'),
381 'imblearn_over_sampling': getattr(imblearn, 'over_sampling'),
382 'imblearn_combine': getattr(imblearn, 'combine')
383 }
384 syms.update(estimator_table)
385
386 for key in unwanted:
387 syms.pop(key, None)
388
389 super(SafeEval, self).__init__(
390 symtable=syms, use_numpy=False, minimal=False,
391 no_if=True, no_for=True, no_while=True, no_try=True,
392 no_functiondef=True, no_ifexp=True, no_listcomp=False,
393 no_augassign=False, no_assert=True, no_delete=True,
394 no_raise=True, no_print=True)
395
396
397 def get_estimator(estimator_json):
398 """Return a sklearn or compatible estimator from Galaxy tool inputs
399 """
400 estimator_module = estimator_json['selected_module']
401
402 if estimator_module == 'custom_estimator':
403 c_estimator = estimator_json['c_estimator']
404 with open(c_estimator, 'rb') as model_handler:
405 new_model = load_model(model_handler)
406 return new_model
407
408 if estimator_module == "binarize_target":
409 wrapped_estimator = estimator_json['wrapped_estimator']
410 with open(wrapped_estimator, 'rb') as model_handler:
411 wrapped_estimator = load_model(model_handler)
412 options = {}
413 if estimator_json['z_score'] is not None:
414 options['z_score'] = estimator_json['z_score']
415 if estimator_json['value'] is not None:
416 options['value'] = estimator_json['value']
417 options['less_is_positive'] = estimator_json['less_is_positive']
418 if estimator_json['clf_or_regr'] == 'BinarizeTargetClassifier':
419 klass = try_get_attr('iraps_classifier',
420 'BinarizeTargetClassifier')
421 else:
422 klass = try_get_attr('iraps_classifier',
423 'BinarizeTargetRegressor')
424 return klass(wrapped_estimator, **options)
425
426 estimator_cls = estimator_json['selected_estimator']
427
428 if estimator_module == 'xgboost':
429 klass = getattr(xgboost, estimator_cls)
430 else:
431 module = getattr(sklearn, estimator_module)
432 klass = getattr(module, estimator_cls)
433
434 estimator = klass()
435
436 estimator_params = estimator_json['text_params'].strip()
437 if estimator_params != '':
438 try:
439 safe_eval = SafeEval()
440 params = safe_eval('dict(' + estimator_params + ')')
441 except ValueError:
442 sys.exit("Unsupported parameter input: `%s`" % estimator_params)
443 estimator.set_params(**params)
444 if 'n_jobs' in estimator.get_params():
445 estimator.set_params(n_jobs=N_JOBS)
446
447 return estimator
448
449
450 def get_cv(cv_json):
451 """ Return CV splitter from Galaxy tool inputs
452
453 Parameters
454 ----------
455 cv_json : dict
456 From Galaxy tool inputs.
457 e.g.:
458 {
459 'selected_cv': 'StratifiedKFold',
460 'n_splits': 3,
461 'shuffle': True,
462 'random_state': 0
463 }
464 """
465 cv = cv_json.pop('selected_cv')
466 if cv == 'default':
467 return cv_json['n_splits'], None
468
469 groups = cv_json.pop('groups_selector', None)
470 if groups is not None:
471 infile_g = groups['infile_g']
472 header = 'infer' if groups['header_g'] else None
473 column_option = (groups['column_selector_options_g']
474 ['selected_column_selector_option_g'])
475 if column_option in ['by_index_number', 'all_but_by_index_number',
476 'by_header_name', 'all_but_by_header_name']:
477 c = groups['column_selector_options_g']['col_g']
478 else:
479 c = None
480 groups = read_columns(
481 infile_g,
482 c=c,
483 c_option=column_option,
484 sep='\t',
485 header=header,
486 parse_dates=True)
487 groups = groups.ravel()
488
489 for k, v in cv_json.items():
490 if v == '':
491 cv_json[k] = None
492
493 test_fold = cv_json.get('test_fold', None)
494 if test_fold:
495 if test_fold.startswith('__ob__'):
496 test_fold = test_fold[6:]
497 if test_fold.endswith('__cb__'):
498 test_fold = test_fold[:-6]
499 cv_json['test_fold'] = [int(x.strip()) for x in test_fold.split(',')]
500
501 test_size = cv_json.get('test_size', None)
502 if test_size and test_size > 1.0:
503 cv_json['test_size'] = int(test_size)
504
505 if cv == 'OrderedKFold':
506 cv_class = try_get_attr('model_validations', 'OrderedKFold')
507 elif cv == 'RepeatedOrderedKFold':
508 cv_class = try_get_attr('model_validations', 'RepeatedOrderedKFold')
509 else:
510 cv_class = getattr(model_selection, cv)
511 splitter = cv_class(**cv_json)
512
513 return splitter, groups
514
515
516 # needed when sklearn < v0.20
517 def balanced_accuracy_score(y_true, y_pred):
518 """Compute balanced accuracy score, which is now available in
519 scikit-learn from v0.20.0.
520 """
521 C = metrics.confusion_matrix(y_true, y_pred)
522 with np.errstate(divide='ignore', invalid='ignore'):
523 per_class = np.diag(C) / C.sum(axis=1)
524 if np.any(np.isnan(per_class)):
525 warnings.warn('y_pred contains classes not in y_true')
526 per_class = per_class[~np.isnan(per_class)]
527 score = np.mean(per_class)
528 return score
529
530
531 def get_scoring(scoring_json):
532 """Return single sklearn scorer class
533 or multiple scoers in dictionary
534 """
535 if scoring_json['primary_scoring'] == 'default':
536 return None
537
538 my_scorers = metrics.SCORERS
539 my_scorers['binarize_auc_scorer'] =\
540 try_get_attr('iraps_classifier', 'binarize_auc_scorer')
541 my_scorers['binarize_average_precision_scorer'] =\
542 try_get_attr('iraps_classifier', 'binarize_average_precision_scorer')
543 if 'balanced_accuracy' not in my_scorers:
544 my_scorers['balanced_accuracy'] =\
545 metrics.make_scorer(balanced_accuracy_score)
546
547 if scoring_json['secondary_scoring'] != 'None'\
548 and scoring_json['secondary_scoring'] !=\
549 scoring_json['primary_scoring']:
550 return_scoring = {}
551 primary_scoring = scoring_json['primary_scoring']
552 return_scoring[primary_scoring] = my_scorers[primary_scoring]
553 for scorer in scoring_json['secondary_scoring'].split(','):
554 if scorer != scoring_json['primary_scoring']:
555 return_scoring[scorer] = my_scorers[scorer]
556 return return_scoring
557
558 return my_scorers[scoring_json['primary_scoring']]
559
560
561 def get_search_params(estimator):
562 """Format the output of `estimator.get_params()`
563 """
564 params = estimator.get_params()
565 results = []
566 for k, v in params.items():
567 # params below won't be shown for search in the searchcv tool
568 keywords = ('n_jobs', 'pre_dispatch', 'memory', 'steps',
569 'nthread', 'verbose')
570 if k.endswith(keywords):
571 results.append(['*', k, k+": "+repr(v)])
572 else:
573 results.append(['@', k, k+": "+repr(v)])
574 results.append(
575 ["", "Note:",
576 "@, params eligible for search in searchcv tool."])
577
578 return results
579
580
581 def try_get_attr(module, name):
582 """try to get attribute from a custom module
583
584 Parameters
585 ----------
586 module : str
587 Module name
588 name : str
589 Attribute (class/function) name.
590
591 Returns
592 -------
593 class or function
594 """
595 mod = sys.modules.get(module, None)
596 if mod:
597 return getattr(mod, name)
598 else:
599 raise Exception("No module named %s." % module)