comparison keras_train_and_eval.py @ 0:af2624d5ab32 draft

"planemo upload for repository https://github.com/bgruening/galaxytools/tree/master/tools/sklearn commit ea12f973df4b97a2691d9e4ce6bf6fae59d57717"
author bgruening
date Sat, 01 May 2021 01:24:32 +0000
parents
children 9349ed2749c6
comparison
equal deleted inserted replaced
-1:000000000000 0:af2624d5ab32
1 import argparse
2 import json
3 import os
4 import pickle
5 import warnings
6 from itertools import chain
7
8 import joblib
9 import numpy as np
10 import pandas as pd
11 from galaxy_ml.externals.selene_sdk.utils import compute_score
12 from galaxy_ml.keras_galaxy_models import _predict_generator
13 from galaxy_ml.model_validations import train_test_split
14 from galaxy_ml.utils import (clean_params, get_main_estimator,
15 get_module, get_scoring, load_model, read_columns,
16 SafeEval, try_get_attr)
17 from scipy.io import mmread
18 from sklearn.metrics.scorer import _check_multimetric_scoring
19 from sklearn.model_selection import _search, _validation
20 from sklearn.model_selection._validation import _score
21 from sklearn.pipeline import Pipeline
22 from sklearn.utils import indexable, safe_indexing
23
24 _fit_and_score = try_get_attr("galaxy_ml.model_validations", "_fit_and_score")
25 setattr(_search, "_fit_and_score", _fit_and_score)
26 setattr(_validation, "_fit_and_score", _fit_and_score)
27
28 N_JOBS = int(os.environ.get("GALAXY_SLOTS", 1))
29 CACHE_DIR = os.path.join(os.getcwd(), "cached")
30 del os
31 NON_SEARCHABLE = ("n_jobs", "pre_dispatch", "memory", "_path", "nthread", "callbacks")
32 ALLOWED_CALLBACKS = (
33 "EarlyStopping",
34 "TerminateOnNaN",
35 "ReduceLROnPlateau",
36 "CSVLogger",
37 "None",
38 )
39
40
41 def _eval_swap_params(params_builder):
42 swap_params = {}
43
44 for p in params_builder["param_set"]:
45 swap_value = p["sp_value"].strip()
46 if swap_value == "":
47 continue
48
49 param_name = p["sp_name"]
50 if param_name.lower().endswith(NON_SEARCHABLE):
51 warnings.warn(
52 "Warning: `%s` is not eligible for search and was "
53 "omitted!" % param_name
54 )
55 continue
56
57 if not swap_value.startswith(":"):
58 safe_eval = SafeEval(load_scipy=True, load_numpy=True)
59 ev = safe_eval(swap_value)
60 else:
61 # Have `:` before search list, asks for estimator evaluatio
62 safe_eval_es = SafeEval(load_estimators=True)
63 swap_value = swap_value[1:].strip()
64 # TODO maybe add regular express check
65 ev = safe_eval_es(swap_value)
66
67 swap_params[param_name] = ev
68
69 return swap_params
70
71
72 def train_test_split_none(*arrays, **kwargs):
73 """extend train_test_split to take None arrays
74 and support split by group names.
75 """
76 nones = []
77 new_arrays = []
78 for idx, arr in enumerate(arrays):
79 if arr is None:
80 nones.append(idx)
81 else:
82 new_arrays.append(arr)
83
84 if kwargs["shuffle"] == "None":
85 kwargs["shuffle"] = None
86
87 group_names = kwargs.pop("group_names", None)
88
89 if group_names is not None and group_names.strip():
90 group_names = [name.strip() for name in group_names.split(",")]
91 new_arrays = indexable(*new_arrays)
92 groups = kwargs["labels"]
93 n_samples = new_arrays[0].shape[0]
94 index_arr = np.arange(n_samples)
95 test = index_arr[np.isin(groups, group_names)]
96 train = index_arr[~np.isin(groups, group_names)]
97 rval = list(
98 chain.from_iterable(
99 (safe_indexing(a, train), safe_indexing(a, test)) for a in new_arrays
100 )
101 )
102 else:
103 rval = train_test_split(*new_arrays, **kwargs)
104
105 for pos in nones:
106 rval[pos * 2: 2] = [None, None]
107
108 return rval
109
110
111 def _evaluate(y_true, pred_probas, scorer, is_multimetric=True):
112 """output scores based on input scorer
113
114 Parameters
115 ----------
116 y_true : array
117 True label or target values
118 pred_probas : array
119 Prediction values, probability for classification problem
120 scorer : dict
121 dict of `sklearn.metrics.scorer.SCORER`
122 is_multimetric : bool, default is True
123 """
124 if y_true.ndim == 1 or y_true.shape[-1] == 1:
125 pred_probas = pred_probas.ravel()
126 pred_labels = (pred_probas > 0.5).astype("int32")
127 targets = y_true.ravel().astype("int32")
128 if not is_multimetric:
129 preds = (
130 pred_labels
131 if scorer.__class__.__name__ == "_PredictScorer"
132 else pred_probas
133 )
134 score = scorer._score_func(targets, preds, **scorer._kwargs)
135
136 return score
137 else:
138 scores = {}
139 for name, one_scorer in scorer.items():
140 preds = (
141 pred_labels
142 if one_scorer.__class__.__name__ == "_PredictScorer"
143 else pred_probas
144 )
145 score = one_scorer._score_func(targets, preds, **one_scorer._kwargs)
146 scores[name] = score
147
148 # TODO: multi-class metrics
149 # multi-label
150 else:
151 pred_labels = (pred_probas > 0.5).astype("int32")
152 targets = y_true.astype("int32")
153 if not is_multimetric:
154 preds = (
155 pred_labels
156 if scorer.__class__.__name__ == "_PredictScorer"
157 else pred_probas
158 )
159 score, _ = compute_score(preds, targets, scorer._score_func)
160 return score
161 else:
162 scores = {}
163 for name, one_scorer in scorer.items():
164 preds = (
165 pred_labels
166 if one_scorer.__class__.__name__ == "_PredictScorer"
167 else pred_probas
168 )
169 score, _ = compute_score(preds, targets, one_scorer._score_func)
170 scores[name] = score
171
172 return scores
173
174
175 def main(
176 inputs,
177 infile_estimator,
178 infile1,
179 infile2,
180 outfile_result,
181 outfile_object=None,
182 outfile_weights=None,
183 outfile_y_true=None,
184 outfile_y_preds=None,
185 groups=None,
186 ref_seq=None,
187 intervals=None,
188 targets=None,
189 fasta_path=None,
190 ):
191 """
192 Parameter
193 ---------
194 inputs : str
195 File path to galaxy tool parameter
196
197 infile_estimator : str
198 File path to estimator
199
200 infile1 : str
201 File path to dataset containing features
202
203 infile2 : str
204 File path to dataset containing target values
205
206 outfile_result : str
207 File path to save the results, either cv_results or test result
208
209 outfile_object : str, optional
210 File path to save searchCV object
211
212 outfile_weights : str, optional
213 File path to save deep learning model weights
214
215 outfile_y_true : str, optional
216 File path to target values for prediction
217
218 outfile_y_preds : str, optional
219 File path to save deep learning model weights
220
221 groups : str
222 File path to dataset containing groups labels
223
224 ref_seq : str
225 File path to dataset containing genome sequence file
226
227 intervals : str
228 File path to dataset containing interval file
229
230 targets : str
231 File path to dataset compressed target bed file
232
233 fasta_path : str
234 File path to dataset containing fasta file
235 """
236 warnings.simplefilter("ignore")
237
238 with open(inputs, "r") as param_handler:
239 params = json.load(param_handler)
240
241 # load estimator
242 with open(infile_estimator, "rb") as estimator_handler:
243 estimator = load_model(estimator_handler)
244
245 estimator = clean_params(estimator)
246
247 # swap hyperparameter
248 swapping = params["experiment_schemes"]["hyperparams_swapping"]
249 swap_params = _eval_swap_params(swapping)
250 estimator.set_params(**swap_params)
251
252 estimator_params = estimator.get_params()
253
254 # store read dataframe object
255 loaded_df = {}
256
257 input_type = params["input_options"]["selected_input"]
258 # tabular input
259 if input_type == "tabular":
260 header = "infer" if params["input_options"]["header1"] else None
261 column_option = params["input_options"]["column_selector_options_1"][
262 "selected_column_selector_option"
263 ]
264 if column_option in [
265 "by_index_number",
266 "all_but_by_index_number",
267 "by_header_name",
268 "all_but_by_header_name",
269 ]:
270 c = params["input_options"]["column_selector_options_1"]["col1"]
271 else:
272 c = None
273
274 df_key = infile1 + repr(header)
275 df = pd.read_csv(infile1, sep="\t", header=header, parse_dates=True)
276 loaded_df[df_key] = df
277
278 X = read_columns(df, c=c, c_option=column_option).astype(float)
279 # sparse input
280 elif input_type == "sparse":
281 X = mmread(open(infile1, "r"))
282
283 # fasta_file input
284 elif input_type == "seq_fasta":
285 pyfaidx = get_module("pyfaidx")
286 sequences = pyfaidx.Fasta(fasta_path)
287 n_seqs = len(sequences.keys())
288 X = np.arange(n_seqs)[:, np.newaxis]
289 for param in estimator_params.keys():
290 if param.endswith("fasta_path"):
291 estimator.set_params(**{param: fasta_path})
292 break
293 else:
294 raise ValueError(
295 "The selected estimator doesn't support "
296 "fasta file input! Please consider using "
297 "KerasGBatchClassifier with "
298 "FastaDNABatchGenerator/FastaProteinBatchGenerator "
299 "or having GenomeOneHotEncoder/ProteinOneHotEncoder "
300 "in pipeline!"
301 )
302
303 elif input_type == "refseq_and_interval":
304 path_params = {
305 "data_batch_generator__ref_genome_path": ref_seq,
306 "data_batch_generator__intervals_path": intervals,
307 "data_batch_generator__target_path": targets,
308 }
309 estimator.set_params(**path_params)
310 n_intervals = sum(1 for line in open(intervals))
311 X = np.arange(n_intervals)[:, np.newaxis]
312
313 # Get target y
314 header = "infer" if params["input_options"]["header2"] else None
315 column_option = params["input_options"]["column_selector_options_2"][
316 "selected_column_selector_option2"
317 ]
318 if column_option in [
319 "by_index_number",
320 "all_but_by_index_number",
321 "by_header_name",
322 "all_but_by_header_name",
323 ]:
324 c = params["input_options"]["column_selector_options_2"]["col2"]
325 else:
326 c = None
327
328 df_key = infile2 + repr(header)
329 if df_key in loaded_df:
330 infile2 = loaded_df[df_key]
331 else:
332 infile2 = pd.read_csv(infile2, sep="\t", header=header, parse_dates=True)
333 loaded_df[df_key] = infile2
334
335 y = read_columns(
336 infile2, c=c, c_option=column_option, sep="\t", header=header, parse_dates=True
337 )
338 if len(y.shape) == 2 and y.shape[1] == 1:
339 y = y.ravel()
340 if input_type == "refseq_and_interval":
341 estimator.set_params(data_batch_generator__features=y.ravel().tolist())
342 y = None
343 # end y
344
345 # load groups
346 if groups:
347 groups_selector = (
348 params["experiment_schemes"]["test_split"]["split_algos"]
349 ).pop("groups_selector")
350
351 header = "infer" if groups_selector["header_g"] else None
352 column_option = groups_selector["column_selector_options_g"][
353 "selected_column_selector_option_g"
354 ]
355 if column_option in [
356 "by_index_number",
357 "all_but_by_index_number",
358 "by_header_name",
359 "all_but_by_header_name",
360 ]:
361 c = groups_selector["column_selector_options_g"]["col_g"]
362 else:
363 c = None
364
365 df_key = groups + repr(header)
366 if df_key in loaded_df:
367 groups = loaded_df[df_key]
368
369 groups = read_columns(
370 groups,
371 c=c,
372 c_option=column_option,
373 sep="\t",
374 header=header,
375 parse_dates=True,
376 )
377 groups = groups.ravel()
378
379 # del loaded_df
380 del loaded_df
381
382 # cache iraps_core fits could increase search speed significantly
383 memory = joblib.Memory(location=CACHE_DIR, verbose=0)
384 main_est = get_main_estimator(estimator)
385 if main_est.__class__.__name__ == "IRAPSClassifier":
386 main_est.set_params(memory=memory)
387
388 # handle scorer, convert to scorer dict
389 scoring = params["experiment_schemes"]["metrics"]["scoring"]
390 if scoring is not None:
391 # get_scoring() expects secondary_scoring to be a comma separated string (not a list)
392 # Check if secondary_scoring is specified
393 secondary_scoring = scoring.get("secondary_scoring", None)
394 if secondary_scoring is not None:
395 # If secondary_scoring is specified, convert the list into comman separated string
396 scoring["secondary_scoring"] = ",".join(scoring["secondary_scoring"])
397
398 scorer = get_scoring(scoring)
399 scorer, _ = _check_multimetric_scoring(estimator, scoring=scorer)
400
401 # handle test (first) split
402 test_split_options = params["experiment_schemes"]["test_split"]["split_algos"]
403
404 if test_split_options["shuffle"] == "group":
405 test_split_options["labels"] = groups
406 if test_split_options["shuffle"] == "stratified":
407 if y is not None:
408 test_split_options["labels"] = y
409 else:
410 raise ValueError(
411 "Stratified shuffle split is not " "applicable on empty target values!"
412 )
413
414 (
415 X_train,
416 X_test,
417 y_train,
418 y_test,
419 groups_train,
420 _groups_test,
421 ) = train_test_split_none(X, y, groups, **test_split_options)
422
423 exp_scheme = params["experiment_schemes"]["selected_exp_scheme"]
424
425 # handle validation (second) split
426 if exp_scheme == "train_val_test":
427 val_split_options = params["experiment_schemes"]["val_split"]["split_algos"]
428
429 if val_split_options["shuffle"] == "group":
430 val_split_options["labels"] = groups_train
431 if val_split_options["shuffle"] == "stratified":
432 if y_train is not None:
433 val_split_options["labels"] = y_train
434 else:
435 raise ValueError(
436 "Stratified shuffle split is not "
437 "applicable on empty target values!"
438 )
439
440 (
441 X_train,
442 X_val,
443 y_train,
444 y_val,
445 groups_train,
446 _groups_val,
447 ) = train_test_split_none(X_train, y_train, groups_train, **val_split_options)
448
449 # train and eval
450 if hasattr(estimator, "validation_data"):
451 if exp_scheme == "train_val_test":
452 estimator.fit(X_train, y_train, validation_data=(X_val, y_val))
453 else:
454 estimator.fit(X_train, y_train, validation_data=(X_test, y_test))
455 else:
456 estimator.fit(X_train, y_train)
457
458 if hasattr(estimator, "evaluate"):
459 steps = estimator.prediction_steps
460 batch_size = estimator.batch_size
461 generator = estimator.data_generator_.flow(
462 X_test, y=y_test, batch_size=batch_size
463 )
464 predictions, y_true = _predict_generator(
465 estimator.model_, generator, steps=steps
466 )
467 scores = _evaluate(y_true, predictions, scorer, is_multimetric=True)
468
469 else:
470 if hasattr(estimator, "predict_proba"):
471 predictions = estimator.predict_proba(X_test)
472 else:
473 predictions = estimator.predict(X_test)
474
475 y_true = y_test
476 scores = _score(estimator, X_test, y_test, scorer, is_multimetric=True)
477 if outfile_y_true:
478 try:
479 pd.DataFrame(y_true).to_csv(outfile_y_true, sep="\t", index=False)
480 pd.DataFrame(predictions).astype(np.float32).to_csv(
481 outfile_y_preds,
482 sep="\t",
483 index=False,
484 float_format="%g",
485 chunksize=10000,
486 )
487 except Exception as e:
488 print("Error in saving predictions: %s" % e)
489
490 # handle output
491 for name, score in scores.items():
492 scores[name] = [score]
493 df = pd.DataFrame(scores)
494 df = df[sorted(df.columns)]
495 df.to_csv(path_or_buf=outfile_result, sep="\t", header=True, index=False)
496
497 memory.clear(warn=False)
498
499 if outfile_object:
500 main_est = estimator
501 if isinstance(estimator, Pipeline):
502 main_est = estimator.steps[-1][-1]
503
504 if hasattr(main_est, "model_") and hasattr(main_est, "save_weights"):
505 if outfile_weights:
506 main_est.save_weights(outfile_weights)
507 del main_est.model_
508 del main_est.fit_params
509 del main_est.model_class_
510 if getattr(main_est, "validation_data", None):
511 del main_est.validation_data
512 if getattr(main_est, "data_generator_", None):
513 del main_est.data_generator_
514
515 with open(outfile_object, "wb") as output_handler:
516 pickle.dump(estimator, output_handler, pickle.HIGHEST_PROTOCOL)
517
518
519 if __name__ == "__main__":
520 aparser = argparse.ArgumentParser()
521 aparser.add_argument("-i", "--inputs", dest="inputs", required=True)
522 aparser.add_argument("-e", "--estimator", dest="infile_estimator")
523 aparser.add_argument("-X", "--infile1", dest="infile1")
524 aparser.add_argument("-y", "--infile2", dest="infile2")
525 aparser.add_argument("-O", "--outfile_result", dest="outfile_result")
526 aparser.add_argument("-o", "--outfile_object", dest="outfile_object")
527 aparser.add_argument("-w", "--outfile_weights", dest="outfile_weights")
528 aparser.add_argument("-l", "--outfile_y_true", dest="outfile_y_true")
529 aparser.add_argument("-p", "--outfile_y_preds", dest="outfile_y_preds")
530 aparser.add_argument("-g", "--groups", dest="groups")
531 aparser.add_argument("-r", "--ref_seq", dest="ref_seq")
532 aparser.add_argument("-b", "--intervals", dest="intervals")
533 aparser.add_argument("-t", "--targets", dest="targets")
534 aparser.add_argument("-f", "--fasta_path", dest="fasta_path")
535 args = aparser.parse_args()
536
537 main(
538 args.inputs,
539 args.infile_estimator,
540 args.infile1,
541 args.infile2,
542 args.outfile_result,
543 outfile_object=args.outfile_object,
544 outfile_weights=args.outfile_weights,
545 outfile_y_true=args.outfile_y_true,
546 outfile_y_preds=args.outfile_y_preds,
547 groups=args.groups,
548 ref_seq=args.ref_seq,
549 intervals=args.intervals,
550 targets=args.targets,
551 fasta_path=args.fasta_path,
552 )