-
Notifications
You must be signed in to change notification settings - Fork 1
/
evaluation.py
399 lines (319 loc) · 16.4 KB
/
evaluation.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
import argparse
from datetime import datetime
from typing import Dict, List, Any, Tuple
import numpy as np
from joblib import Parallel, delayed
from sklearn.metrics import (accuracy_score, roc_curve, auc, precision_recall_curve,
average_precision_score, precision_score)
from sklearn.metrics._ranking import _binary_clf_curve
from sklearn.utils.extmath import softmax
from tqdm import tqdm
from common import load_pickle, load_json_file, to_pickle, ScoresAttributes, extract_pattern_type
from data.patterns import PatternMatrix, PatternType
from configuration import Config
NUM_JOBS = 4
A = ScoresAttributes.get()
def _results_dict() -> Dict:
return {
# {'SNR_0': {'method_0': [roc_auc_0, roc_auc_1, ...], ...}, ...}
# roc_auc_k is {'fpr': [], 'tpr': [], 'auc': []}
'roc_auc': dict(),
'precision_based_scores': dict(),
# {'SNR_0': {'method_0': [explanation_0, explanation_1, ...], ...}, ...}
# explanation_k is either R^{n_validation_data x n_features} or R^{n_features}
A.explanations: dict(),
A.method_names: list()
}
def _generate_empty_evaluation_results_dict() -> Dict:
return {
A.global_based: _results_dict(),
A.sample_based: _results_dict(),
# {'SNR_0': {'train': [acc_data_set_0, acc_data_set_1, ...], 'val': [...]},
# 'SNR_1': {'train': [...], ...}
A.model_accuracies: dict(),
# {'SNR_0': {'logistic_regression': [model_weight_0, model_weight_1, ...],
# 'neural_net':[...]},
# 'SNR_1': {[...]}, ...}
A.model_weights: dict(),
A.data_weights: list(),
}
def _is_keras_model(model: Any) -> bool:
return isinstance(model, list)
def _calculate_model_accuracies(data: List, results_per_snr: List) -> Dict:
def predict_with_keras_model(model_weights: List, x: np.ndarray) -> np.ndarray:
try:
pred = softmax(np.dot(x, model_weights[0]) + model_weights[1])
except IndexError as e:
pred = softmax(np.dot(x, model_weights[0]))
return np.argmax(pred, axis=1)
accuracies = {'train': list(), 'val': list()}
for k in range(len(results_per_snr)):
model = results_per_snr[k]['model']
if _is_keras_model(model=model):
pred_train = predict_with_keras_model(model_weights=model, x=data[k]['train']['x'])
pred_val = predict_with_keras_model(model_weights=model, x=data[k]['val']['x'])
else:
pred_train = model.predict(X=data[k]['train']['x'])
pred_val = model.predict(X=data[k]['val']['x'])
accuracies['val'] += [accuracy_score(y_true=data[k]['val']['y'], y_pred=pred_val)]
accuracies['train'] += [accuracy_score(y_true=data[k]['train']['y'], y_pred=pred_train)]
return accuracies
def precision_curves(y_true, probas_pred, *, pos_label=None,
sample_weight=None):
"""
Minor adaption of corresponding scikit-learn function
"""
fps, tps, thresholds = _binary_clf_curve(y_true, probas_pred,
pos_label=pos_label,
sample_weight=sample_weight)
precision = tps / (tps + fps)
precision[np.isnan(precision)] = 0
recall = tps / tps[-1]
specificity = 1 - fps / fps[-1]
last_ind = tps.searchsorted(tps[-1])
sl = slice(last_ind, None, -1)
return np.r_[precision[sl], 1], np.r_[recall[sl], 0], np.r_[specificity[sl], 1]
def _assemble_explanations(method_names: List, results_per_snr: List) -> Dict:
explanations_dict = dict()
for method in method_names:
explanations_list = list()
for k in range(len(results_per_snr)):
explanations_list += [results_per_snr[k]['explanations'][method]]
explanations_dict[method] = explanations_list
return explanations_dict
def _generate_true_feature_importance(pattern_type: int = 0) -> np.ndarray:
pattern = PatternMatrix(pattern_type=pattern_type)
binary_mask_signal = np.array(1 * (pattern.matrix[:, pattern.dim_of_signal] > 0))
binary_mask_signal[(pattern.matrix[:, pattern.dim_of_signal] < 0)] = -1
return binary_mask_signal
def _compute_roc_auc(y_true: np.ndarray, y_score: np.ndarray,
use_abs: bool = True) -> float:
if use_abs:
y_true, y_score = np.abs(y_true), np.abs(y_score)
fpr, tpr, _ = roc_curve(y_true=y_true, y_score=y_score)
auc_value = auc(x=fpr, y=tpr)
return auc_value
def _assemble_roc_auc_values(y_true: np.ndarray, y_score: List,
idx: int, use_absolute_values: bool = True) -> Dict:
if 1 < len(y_score[idx].shape):
fpr_list, tpr_list, auc_value_list = list(), list(), list()
for k in range(y_score[idx].shape[0]):
auc_value = _compute_roc_auc(
y_true=y_true, y_score=y_score[idx][k, :], use_abs=use_absolute_values)
auc_value_list += [auc_value]
output = {'auc': auc_value_list}
else:
auc_value = _compute_roc_auc(
y_true=y_true, y_score=y_score[idx], use_abs=use_absolute_values)
output = {'auc': [auc_value]}
return output
def _precision_specificity_score(precision: np.ndarray, specificity: np.ndarray,
threshold: float = 0.9) -> float:
score = 0.0
if any(threshold < specificity[:-1]):
score = precision[:-1][threshold < specificity[:-1]][0]
return score
def _compute_precision_based_scores(y_true: np.ndarray, y_score: np.ndarray,
use_abs: bool = True) -> Tuple:
if use_abs is True:
y_true, y_score = np.abs(y_true), np.abs(y_score)
precision, recall, specificity = precision_curves(y_true=y_true, probas_pred=y_score)
pr_auc = auc(x=recall, y=precision)
avg_prec_score = average_precision_score(y_true=y_true, y_score=y_score)
avg_npv_specificity = average_precision_score(y_true=1 - y_true, y_score=1 - y_score)
prec_spec_score = _precision_specificity_score(precision=precision, specificity=specificity)
return pr_auc, avg_prec_score, avg_npv_specificity, prec_spec_score
def _assemble_precision_based_scores(y_true: np.ndarray, y_score: List,
idx: int, use_absolute_values: bool = True) -> Dict:
if 1 < len(y_score[idx].shape):
pr_auc_list, avg_list, avg_npv_specificity_list, max_prec_list = list(), list(), list(), list()
for k in range(y_score[idx].shape[0]):
auc, avg_prec_score, avg_npv_specificity, max_prec = _compute_precision_based_scores(
y_true=y_true, y_score=y_score[idx][k, :], use_abs=use_absolute_values)
avg_list += [avg_prec_score]
pr_auc_list += [auc]
max_prec_list += [max_prec]
avg_npv_specificity_list += [avg_npv_specificity]
output = {'pr_auc': pr_auc_list, 'avg_precision': avg_list,
'avg_npv_specificity': avg_npv_specificity_list, 'max_precision': max_prec_list}
else:
auc, avg_prec_score, avg_npv_specificity, max_prec = _compute_precision_based_scores(
y_true=y_true, y_score=y_score[idx], use_abs=use_absolute_values)
output = {'pr_auc': [auc], 'avg_npv_specificity': [avg_npv_specificity],
'avg_precision': [avg_prec_score], 'max_precision': [max_prec]}
return output
def _roc_analysis(explanations_by_methods: Dict, pattern_type: int) -> Dict:
roc_auc_values = dict()
true_importance = _generate_true_feature_importance(pattern_type=pattern_type)
for method, explanations_list in explanations_by_methods.items():
roc_auc_values[method] = Parallel(n_jobs=NUM_JOBS)(
delayed(_assemble_roc_auc_values)(true_importance, explanations_list, idx)
for idx in range(len(explanations_list)))
return roc_auc_values
def _precision_based_analysis(explanations_by_methods: Dict, pattern_type: int) -> Dict:
pr_scores = dict()
true_importance = _generate_true_feature_importance(pattern_type=pattern_type)
for method, explanations_list in explanations_by_methods.items():
pr_scores[method] = Parallel(n_jobs=NUM_JOBS)(
delayed(_assemble_precision_based_scores)(true_importance, explanations_list, idx)
for idx in range(len(explanations_list)))
return pr_scores
def _get_model_weights(results_per_snr: List) -> List:
model_weights = list()
for k in range(len(results_per_snr)):
model = results_per_snr[k]['model']
if _is_keras_model(model=model):
model_weights += [model[0]]
else:
model_weights += [model.coef_]
return model_weights
def _is_saliency(method_names: List) -> bool:
return True if 'deep_taylor' in method_names else False
def _is_sample_based_agnostic(method_names: List) -> bool:
return True if 'lime' in method_names else False
def _get_data_weights(result: Dict) -> List:
return list(result['results'].keys())
def _extract_explanations(results: List[Dict], aggregate: bool, use_abs: bool = True) -> Dict:
output = dict()
weights = _get_data_weights(result=results[0])
method_names = list()
for k, r in enumerate(results):
method_names += [(m, k) for m in r['method_names']]
for w in weights:
results_per_method = dict()
for m, j in method_names:
results_per_experiment = list()
for r in results[j]['results'][w]:
explanation = np.abs(r['explanations'][m]) if use_abs else r['explanations'][m]
if aggregate and 1 < len(explanation.shape):
results_per_experiment += [np.mean(explanation, axis=0)]
else:
results_per_experiment += [explanation]
results_per_method[m] = results_per_experiment
output[w] = results_per_method
return output
def _is_sample_based_method(names: List[str]) -> bool:
return _is_saliency(method_names=names) or _is_sample_based_agnostic(method_names=names)
def _get_method_names(results: List, sample_based: bool) -> List[str]:
output = list()
for result in results:
method_names = result['method_names']
if sample_based and _is_sample_based_method(names=method_names):
output += method_names
elif not sample_based:
output += method_names
return output
def _assemble_explanations2(results: List, scores: Dict) -> Tuple[Dict, Dict]:
sample_based_results = list()
for result in results:
if _is_sample_based_method(names=result['method_names']):
sample_based_results += [result]
global_explanations = _extract_explanations(results=results, aggregate=True)
sample_based_explanations = _extract_explanations(results=sample_based_results, aggregate=False)
for key in global_explanations.keys():
global_explanations[key]['llr'] = get_weights_of_logistic_regression(key=key, scores=scores)
return global_explanations, sample_based_explanations
def _load_results(result_paths: List[str]) -> List:
output = list()
for p in result_paths:
output += [load_pickle(file_path=p)]
return output
def _assemble_model_accuracies(results: List, data: Dict) -> Dict:
output = dict()
for w, data_list in data.items():
logistic_regression_counter = 0
results_per_model_type = dict()
for result in results:
if _is_saliency(method_names=result['method_names']):
results_per_model_type[A.neural_net] = _calculate_model_accuracies(
data=data_list, results_per_snr=result['results'][w])
# Since we have two of the same logistic regression models but just
# one neural net model.
elif 0 == logistic_regression_counter:
logistic_regression_counter += 1
results_per_model_type[A.logistic_regression] = _calculate_model_accuracies(
data=data_list, results_per_snr=result['results'][w])
else:
continue
output[w] = results_per_model_type
return output
def _assemble_model_weights(results: List, weights: List) -> Dict:
output = dict()
for w in weights:
logistic_regression_counter = 0
model_weights_per_model_type = dict()
for result in results:
if _is_saliency(method_names=result['method_names']):
model_weights_per_model_type[A.neural_net] = _get_model_weights(
results_per_snr=result['results'][w])
# Since we have two of the same logistic regression models but just
# one neural net model.
elif 0 == logistic_regression_counter:
logistic_regression_counter += 1
model_weights_per_model_type[A.logistic_regression] = _get_model_weights(
results_per_snr=result['results'][w])
else:
continue
output[w] = model_weights_per_model_type
return output
def _assemble_results_roc_analysis(explanations: Dict, weights: List, pattern_type: int) -> Dict:
output = dict()
for w in tqdm(weights):
output[w] = _roc_analysis(explanations_by_methods=explanations[w],
pattern_type=pattern_type)
return output
def _assemble_results_precision_analysis(explanations: Dict,
weights: List, pattern_type: int) -> Dict:
output = dict()
for w in tqdm(weights):
output[w] = _precision_based_analysis(explanations_by_methods=explanations[w],
pattern_type=pattern_type)
return output
def get_weights_of_logistic_regression(key: str, scores: Dict) -> List:
return [c.flatten() for c in scores[A.model_weights][key]['Logistic Regression']]
def evaluate(config: Config) -> List[str]:
data = load_pickle(file_path=config.data_path)
results = _load_results(result_paths=config.result_paths)
print('Assemble explanations, model weights, ...!')
scores = _generate_empty_evaluation_results_dict()
scores[A.global_based][A.method_names] = _get_method_names(results=results, sample_based=False)
scores[A.sample_based][A.method_names] = _get_method_names(results=results, sample_based=True)
scores[A.data_weights] = _get_data_weights(result=load_pickle(file_path=config.result_paths[0]))
scores[A.model_accuracies] = _assemble_model_accuracies(results=results, data=data)
scores[A.model_weights] = _assemble_model_weights(results=results, weights=scores[A.data_weights])
g, s = _assemble_explanations2(results=results, scores=scores)
scores[A.global_based][A.explanations] = g
scores[A.sample_based][A.explanations] = s
print('Calculate scores!')
pattern_type = int(extract_pattern_type(data_path=config.data_path))
scores[A.global_based]['roc_auc'] = _assemble_results_roc_analysis(
explanations=scores[A.global_based][A.explanations],
weights=scores[A.data_weights], pattern_type=pattern_type)
scores[A.global_based]['precision_based_scores'] = _assemble_results_precision_analysis(
explanations=scores[A.global_based][A.explanations],
weights=scores[A.data_weights], pattern_type=pattern_type)
scores[A.sample_based]['roc_auc'] = _assemble_results_roc_analysis(
explanations=scores[A.sample_based][A.explanations],
weights=scores[A.data_weights], pattern_type=pattern_type)
scores[A.sample_based]['precision_based_scores'] = _assemble_results_precision_analysis(
explanations=scores[A.sample_based][A.explanations],
weights=scores[A.data_weights], pattern_type=pattern_type)
print('Save results!')
output_paths = list()
date = datetime.now().strftime("%Y-%m-%d-%H-%M-%S")
pattern_type = f'pattern_type_{extract_pattern_type(data_path=config.data_path)}'
suffix = '_'.join(['evaluation', date, pattern_type])
output_paths += [to_pickle(output_dir=config.output_dir_scores, data=scores, suffix=suffix)]
return output_paths
def get_command_line_arguments() -> argparse.Namespace:
parser = argparse.ArgumentParser()
parser.add_argument('--path', dest='path', required=True,
help='Input file path of json file containing'
'input parameter for the experiment!')
return parser.parse_args()
if __name__ == '__main__':
args = get_command_line_arguments()
try:
evaluate(config=Config.get(input_conf=load_json_file(file_path=args.path)))
except KeyboardInterrupt as e:
print(e)