-
Notifications
You must be signed in to change notification settings - Fork 14
/
Copy pathmodel_selection.py
436 lines (378 loc) · 19.5 KB
/
model_selection.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
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
import numpy as np
from subpopbench.utils.misc import safe_load
class SelectionMethod:
"""
Abstract class whose subclasses implement strategies for model selection across hparams & steps
"""
def __init__(self):
raise TypeError
@classmethod
def run_acc(cls, run_records):
"""
Given records from a run, return a {val_acc, test_acc, ...} dict representing
the best val-acc, corresponding test-acc and other test metrics for that run.
"""
raise NotImplementedError
@classmethod
def hparams_accs(cls, records):
"""
Given all records from a single (dataset, algorithm) pair,
return a sorted list of (run_acc, records) tuples.
"""
return (records.group('args.hparams_seed').map(
lambda _, run_records: (
cls.run_acc(run_records),
run_records
)
).filter(lambda x: x[0] is not None).sorted(key=lambda x: x[0]['val_acc'])[::-1])
@classmethod
def sweep_acc(cls, records):
"""
Given all records from a single (dataset, algorithm) pair,
return the mean test acc of the k runs with the top val accs.
"""
_hparams_accs = cls.hparams_accs(records)
if len(_hparams_accs):
return _hparams_accs[0][0]['test_acc']
else:
return None
@classmethod
def sweep_acc_worst(cls, records):
_hparams_accs = cls.hparams_accs(records)
if len(_hparams_accs):
return _hparams_accs[0][0]['test_acc_worst']
else:
return None
@classmethod
def sweep_precision(cls, records):
_hparams_accs = cls.hparams_accs(records)
if len(_hparams_accs):
return _hparams_accs[0][0]['test_precision']
else:
return None
@classmethod
def sweep_precision_worst(cls, records):
_hparams_accs = cls.hparams_accs(records)
if len(_hparams_accs):
return _hparams_accs[0][0]['test_precision_worst']
else:
return None
@classmethod
def sweep_f1(cls, records):
_hparams_accs = cls.hparams_accs(records)
if len(_hparams_accs):
return _hparams_accs[0][0]['test_f1']
else:
return None
@classmethod
def sweep_f1_worst(cls, records):
_hparams_accs = cls.hparams_accs(records)
if len(_hparams_accs):
return _hparams_accs[0][0]['test_f1_worst']
else:
return None
@classmethod
def sweep_acc_adjusted(cls, records):
_hparams_accs = cls.hparams_accs(records)
if len(_hparams_accs):
return _hparams_accs[0][0]['test_acc_adjusted']
else:
return None
@classmethod
def sweep_acc_balanced(cls, records):
_hparams_accs = cls.hparams_accs(records)
if len(_hparams_accs):
return _hparams_accs[0][0]['test_acc_balanced']
else:
return None
@classmethod
def sweep_auroc(cls, records):
_hparams_accs = cls.hparams_accs(records)
if len(_hparams_accs):
return _hparams_accs[0][0]['test_auroc']
else:
return None
@classmethod
def sweep_worst_auroc(cls, records):
_hparams_accs = cls.hparams_accs(records)
if len(_hparams_accs):
return _hparams_accs[0][0]['test_worst_auroc']
else:
return None
@classmethod
def sweep_auprc(cls, records):
_hparams_accs = cls.hparams_accs(records)
if len(_hparams_accs):
return _hparams_accs[0][0]['test_auprc']
else:
return None
@classmethod
def sweep_ece(cls, records):
_hparams_accs = cls.hparams_accs(records)
if len(_hparams_accs):
return _hparams_accs[0][0]['test_ece']
else:
return None
@classmethod
def get_test_split(cls, record):
if record['args']['dataset'] == 'ImagenetBG':
return 'mixed_rand'
elif record['args']['dataset'] == 'Living17':
return 'zs'
else:
return 'te'
class OracleMeanAcc(SelectionMethod):
"""Picks argmax(mean(grp_test_acc for grp in all_groups))"""
name = "test set mean accuracy (oracle)"
@classmethod
def _step_acc(cls, record):
"""Given a single record, return a {val_acc, test_acc, ...} dict."""
te = cls.get_test_split(record)
return {'val_acc': record[te]['overall']['accuracy'],
'test_acc': record[te]['overall']['accuracy'],
'test_acc_worst': record[te]['min_group']['accuracy'],
'test_precision': record[te]['overall']['macro_avg']['precision'],
'test_precision_worst': np.min(
[record[te]['per_class'][i]['precision'] for i in record[te]['per_class']]),
'test_f1': record[te]['overall']['macro_avg']['f1-score'],
'test_f1_worst': np.min([record[te]['per_class'][i]['f1-score'] for i in record[te]['per_class']]),
'test_acc_adjusted': record[te]['adjusted_accuracy'],
'test_acc_balanced': record[te]['overall']['balanced_acc'],
'test_auroc': safe_load(record[te]['overall']['AUROC']),
'test_worst_auroc': safe_load(record[te]['min_attr']['AUROC']),
'test_ece': record[te]['overall']['ECE']}
@classmethod
def run_acc(cls, run_records):
return run_records.map(cls._step_acc).argmax('test_acc')
class OracleWorstAcc(SelectionMethod):
"""Picks argmax(min(grp_test_acc for grp in all_groups))"""
name = "test set worst accuracy (oracle)"
@classmethod
def _step_acc(cls, record):
"""Given a single record, return a {val_acc, test_acc, ...} dict."""
te = cls.get_test_split(record)
return {'val_acc': record[te]['min_group']['accuracy'],
'test_acc': record[te]['overall']['accuracy'],
'test_acc_worst': record[te]['min_group']['accuracy'],
'test_precision': record[te]['overall']['macro_avg']['precision'],
'test_precision_worst': np.min(
[record[te]['per_class'][i]['precision'] for i in record[te]['per_class']]),
'test_f1': record[te]['overall']['macro_avg']['f1-score'],
'test_f1_worst': np.min([record[te]['per_class'][i]['f1-score'] for i in record[te]['per_class']]),
'test_acc_adjusted': record[te]['adjusted_accuracy'],
'test_acc_balanced': record[te]['overall']['balanced_acc'],
'test_auroc': safe_load(record[te]['overall']['AUROC']),
'test_worst_auroc': safe_load(record[te]['min_attr']['AUROC']),
'test_ece': record[te]['overall']['ECE']}
@classmethod
def run_acc(cls, run_records):
return run_records.map(cls._step_acc).argmax('test_acc_worst')
class ValMeanAcc(SelectionMethod):
# attribute agnostic
name = "validation set mean accuracy"
@classmethod
def _step_acc(cls, record):
te = cls.get_test_split(record)
return {'val_acc': record['va']['overall']['accuracy'],
'test_acc': record[te]['overall']['accuracy'],
'test_acc_worst': record[te]['min_group']['accuracy'],
'test_precision': record[te]['overall']['macro_avg']['precision'],
'test_precision_worst': np.min(
[record[te]['per_class'][i]['precision'] for i in record[te]['per_class']]),
'test_f1': record[te]['overall']['macro_avg']['f1-score'],
'test_f1_worst': np.min([record[te]['per_class'][i]['f1-score'] for i in record[te]['per_class']]),
'test_acc_adjusted': record[te]['adjusted_accuracy'],
'test_acc_balanced': record[te]['overall']['balanced_acc'],
'test_auroc': safe_load(record[te]['overall']['AUROC']),
'test_worst_auroc': safe_load(record[te]['min_attr']['AUROC']),
'test_ece': record[te]['overall']['ECE']}
@classmethod
def run_acc(cls, run_records):
if not len(run_records):
return None
return run_records.map(cls._step_acc).argmax('val_acc')
class ValWorstAccAttributeYes(ValMeanAcc):
"""Picks argmax(min(grp_val_acc for grp in all_groups))"""
name = "validation set worst accuracy (with attributes)"
@classmethod
def _step_acc(cls, record):
te = cls.get_test_split(record)
return {'val_acc': record['va']['min_group']['accuracy'],
'test_acc': record[te]['overall']['accuracy'],
'test_acc_worst': record[te]['min_group']['accuracy'],
'test_precision': record[te]['overall']['macro_avg']['precision'],
'test_precision_worst': np.min(
[record[te]['per_class'][i]['precision'] for i in record[te]['per_class']]),
'test_f1': record[te]['overall']['macro_avg']['f1-score'],
'test_f1_worst': np.min([record[te]['per_class'][i]['f1-score'] for i in record[te]['per_class']]),
'test_acc_adjusted': record[te]['adjusted_accuracy'],
'test_acc_balanced': record[te]['overall']['balanced_acc'],
'test_auroc': safe_load(record[te]['overall']['AUROC']),
'test_worst_auroc': safe_load(record[te]['min_attr']['AUROC']),
'test_ece': record[te]['overall']['ECE']}
class ValWorstAccAttributeNo(ValMeanAcc):
"""Picks argmax(min(grp_val_acc for grp in all_groups))"""
name = "validation set worst accuracy (without attributes)"
@classmethod
def _step_acc(cls, record):
# class becomes the minimum group
te = cls.get_test_split(record)
return {'val_acc': np.min([record['va']['per_class'][i]['recall'] for i in record['va']['per_class']]),
'test_acc': record[te]['overall']['accuracy'],
'test_acc_worst': record[te]['min_group']['accuracy'],
'test_precision': record[te]['overall']['macro_avg']['precision'],
'test_precision_worst': np.min(
[record[te]['per_class'][i]['precision'] for i in record[te]['per_class']]),
'test_f1': record[te]['overall']['macro_avg']['f1-score'],
'test_f1_worst': np.min([record[te]['per_class'][i]['f1-score'] for i in record[te]['per_class']]),
'test_acc_adjusted': record[te]['adjusted_accuracy'],
'test_acc_balanced': record[te]['overall']['balanced_acc'],
'test_auroc': safe_load(record[te]['overall']['AUROC']),
'test_worst_auroc': safe_load(record[te]['min_attr']['AUROC']),
'test_ece': record[te]['overall']['ECE']}
class ValMeanPrecision(ValMeanAcc):
"""Picks argmax(mean(cls_val_precision for cls in all_classes))"""
name = "validation set mean precision"
@classmethod
def _step_acc(cls, record):
te = cls.get_test_split(record)
return {'val_acc': record['va']['overall']['macro_avg']['precision'],
'test_acc': record[te]['overall']['accuracy'],
'test_acc_worst': record[te]['min_group']['accuracy'],
'test_precision': record[te]['overall']['macro_avg']['precision'],
'test_precision_worst': np.min(
[record[te]['per_class'][i]['precision'] for i in record[te]['per_class']]),
'test_f1': record[te]['overall']['macro_avg']['f1-score'],
'test_f1_worst': np.min([record[te]['per_class'][i]['f1-score'] for i in record[te]['per_class']]),
'test_acc_adjusted': record[te]['adjusted_accuracy'],
'test_acc_balanced': record[te]['overall']['balanced_acc'],
'test_auroc': safe_load(record[te]['overall']['AUROC']),
'test_worst_auroc': safe_load(record[te]['min_attr']['AUROC']),
'test_ece': record[te]['overall']['ECE']}
class ValWorstPrecision(ValMeanAcc):
"""Picks argmax(min(cls_val_precision for cls in all_classes))"""
name = "validation set worst precision"
@classmethod
def _step_acc(cls, record):
te = cls.get_test_split(record)
return {'val_acc': np.min([record['va']['per_class'][i]['precision'] for i in record['va']['per_class']]),
'test_acc': record[te]['overall']['accuracy'],
'test_acc_worst': record[te]['min_group']['accuracy'],
'test_precision': record[te]['overall']['macro_avg']['precision'],
'test_precision_worst': np.min(
[record[te]['per_class'][i]['precision'] for i in record[te]['per_class']]),
'test_f1': record[te]['overall']['macro_avg']['f1-score'],
'test_f1_worst': np.min([record[te]['per_class'][i]['f1-score'] for i in record[te]['per_class']]),
'test_acc_adjusted': record[te]['adjusted_accuracy'],
'test_acc_balanced': record[te]['overall']['balanced_acc'],
'test_auroc': safe_load(record[te]['overall']['AUROC']),
'test_worst_auroc': safe_load(record[te]['min_attr']['AUROC']),
'test_ece': record[te]['overall']['ECE']}
class ValMeanF1(ValMeanAcc):
"""Picks argmax(mean(cls_val_f1 for cls in all_classes))"""
name = "validation set mean f1-score"
@classmethod
def _step_acc(cls, record):
te = cls.get_test_split(record)
return {'val_acc': record['va']['overall']['macro_avg']['f1-score'],
'test_acc': record[te]['overall']['accuracy'],
'test_acc_worst': record[te]['min_group']['accuracy'],
'test_precision': record[te]['overall']['macro_avg']['precision'],
'test_precision_worst': np.min(
[record[te]['per_class'][i]['precision'] for i in record[te]['per_class']]),
'test_f1': record[te]['overall']['macro_avg']['f1-score'],
'test_f1_worst': np.min([record[te]['per_class'][i]['f1-score'] for i in record[te]['per_class']]),
'test_acc_adjusted': record[te]['adjusted_accuracy'],
'test_acc_balanced': record[te]['overall']['balanced_acc'],
'test_auroc': safe_load(record[te]['overall']['AUROC']),
'test_worst_auroc': safe_load(record[te]['min_attr']['AUROC']),
'test_ece': record[te]['overall']['ECE']}
class ValWorstF1(ValMeanAcc):
"""Picks argmax(min(cls_val_f1 for cls in all_classes))"""
name = "validation set worst f1-score"
@classmethod
def _step_acc(cls, record):
te = cls.get_test_split(record)
return {'val_acc': np.min([record['va']['per_class'][i]['f1-score'] for i in record['va']['per_class']]),
'test_acc': record[te]['overall']['accuracy'],
'test_acc_worst': record[te]['min_group']['accuracy'],
'test_precision': record[te]['overall']['macro_avg']['precision'],
'test_precision_worst': np.min(
[record[te]['per_class'][i]['precision'] for i in record[te]['per_class']]),
'test_f1': record[te]['overall']['macro_avg']['f1-score'],
'test_f1_worst': np.min([record[te]['per_class'][i]['f1-score'] for i in record[te]['per_class']]),
'test_acc_adjusted': record[te]['adjusted_accuracy'],
'test_acc_balanced': record[te]['overall']['balanced_acc'],
'test_auroc': safe_load(record[te]['overall']['AUROC']),
'test_worst_auroc': safe_load(record[te]['min_attr']['AUROC']),
'test_ece': record[te]['overall']['ECE']}
class ValBalancedAcc(ValMeanAcc):
"""Picks argmax(balanced_acc)"""
name = "validation set class-balanced accuracy (macro recall)"
@classmethod
def _step_acc(cls, record):
te = cls.get_test_split(record)
return {'val_acc': record['va']['overall']['balanced_acc'],
'test_acc': record[te]['overall']['accuracy'],
'test_acc_worst': record[te]['min_group']['accuracy'],
'test_precision': record[te]['overall']['macro_avg']['precision'],
'test_precision_worst': np.min(
[record[te]['per_class'][i]['precision'] for i in record[te]['per_class']]),
'test_f1': record[te]['overall']['macro_avg']['f1-score'],
'test_f1_worst': np.min([record[te]['per_class'][i]['f1-score'] for i in record[te]['per_class']]),
'test_acc_adjusted': record[te]['adjusted_accuracy'],
'test_acc_balanced': record[te]['overall']['balanced_acc'],
'test_auroc': safe_load(record[te]['overall']['AUROC']),
'test_worst_auroc': safe_load(record[te]['min_attr']['AUROC']),
'test_ece': record[te]['overall']['ECE']}
class ValAUROC(ValMeanAcc):
"""Picks argmax(auroc)"""
name = "validation set AUROC"
@classmethod
def _step_acc(cls, record):
te = cls.get_test_split(record)
return {'val_acc': safe_load(record['va']['overall']['AUROC']),
'test_acc': record[te]['overall']['accuracy'],
'test_acc_worst': record[te]['min_group']['accuracy'],
'test_precision': record[te]['overall']['macro_avg']['precision'],
'test_precision_worst': np.min(
[record[te]['per_class'][i]['precision'] for i in record[te]['per_class']]),
'test_f1': record[te]['overall']['macro_avg']['f1-score'],
'test_f1_worst': np.min([record[te]['per_class'][i]['f1-score'] for i in record[te]['per_class']]),
'test_acc_adjusted': record[te]['adjusted_accuracy'],
'test_acc_balanced': record[te]['overall']['balanced_acc'],
'test_auroc': safe_load(record[te]['overall']['AUROC']),
'test_worst_auroc': safe_load(record[te]['min_attr']['AUROC']),
'test_ece': record[te]['overall']['ECE']}
class ValClassDiff(SelectionMethod):
"""
Minimum class difference as model selection without group annotations
https://openreview.net/pdf?id=TSqRwmrRiOn
"""
name = "minimum class difference"
@classmethod
def _step_acc(cls, record):
class_diff = 0.
per_class_acc = [record['va']['per_class'][i]['recall'] for i in record['va']['per_class']]
for i in range(len(per_class_acc)):
for j in range(i+1, len(per_class_acc)):
class_diff += np.abs(per_class_acc[i] - per_class_acc[j])
te = cls.get_test_split(record)
return {'val_acc': class_diff,
'test_acc': record[te]['overall']['accuracy'],
'test_acc_worst': record[te]['min_group']['accuracy'],
'test_precision': record[te]['overall']['macro_avg']['precision'],
'test_precision_worst': np.min(
[record[te]['per_class'][i]['precision'] for i in record[te]['per_class']]),
'test_f1': record[te]['overall']['macro_avg']['f1-score'],
'test_f1_worst': np.min([record[te]['per_class'][i]['f1-score'] for i in record[te]['per_class']]),
'test_acc_adjusted': record[te]['adjusted_accuracy'],
'test_acc_balanced': record[te]['overall']['balanced_acc'],
'test_auroc': safe_load(record[te]['overall']['AUROC']),
'test_worst_auroc': safe_load(record[te]['min_attr']['AUROC']),
'test_ece': record[te]['overall']['ECE']}
@classmethod
def run_acc(cls, run_records):
if not len(run_records):
return None
return run_records.map(cls._step_acc).argmin('val_acc')