-
Notifications
You must be signed in to change notification settings - Fork 0
/
multicalibrator.py
460 lines (397 loc) · 16.4 KB
/
multicalibrator.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
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
"""
Proportional Multicalibration Post-processor
copyright William La Cava
License: GNU GPL3
"""
import ipdb
import numpy as np
import pandas as pd
from tqdm import tqdm
from sklearn.base import BaseEstimator, ClassifierMixin, TransformerMixin
from sklearn.model_selection import train_test_split
from sklearn.utils.validation import check_X_y, check_array, check_is_fitted
from sklearn.utils.multiclass import unique_labels
from sklearn.utils import resample
from sklearn.metrics import mean_squared_error as mse
from sklearn.metrics import r2_score
from copy import copy
import pmc.utils as utils
from pmc.metrics import (multicalibration_score,
proportional_multicalibration_score)
import logging
logging.basicConfig(level=logging.DEBUG)
logger = logging.getLogger(__name__)
class MultiCalibrator(ClassifierMixin, BaseEstimator):
""" A classifier post-processor that updates a model to satisfy different
notions of fairness.
Parameters
----------
estimator : Probabilistic Classifier
A pre-trained classifier that outputs probabilities.
auditor_type: Classifier or callable
Method that returns a subset of sample from the data, belonging to a
specific group.
metric: 'MC' or 'PMC', default: PMC
alpha: float, default: 0.01
tolerance for calibration error per group.
n_bins: int, default: 10
used to discretize probabilities.
bin_scaling: str, default: 'linear'
how to space the bins; linear or log
gamma: float, default: 0.1
the minimum probability of a group occuring in the data.
rho: float, default: 0.1
the minimum risk prediction to attempt to adjust.
relevant for proportional multicalibration.
max_iters: int, default: None
maximum iterations. Will terminate whether or not alpha is achieved.
random_state: int, default: 0
random seed.
Attributes
----------
X_ : ndarray, shape (n_samples, n_features)
The input passed during :meth:`fit`.
y_ : ndarray, shape (n_samples,)
The labels passed during :meth:`fit`.
classes_ : ndarray, shape (n_classes,)
The classes seen at :meth:`fit`.
"""
def __init__(self,
estimator=None,
auditor_type=None,
metric='PMC',
alpha=0.01,
n_bins=10,
bin_scaling='standard',
gamma=0.01,
rho=0.1,
eta=1.0,
max_iters=100,
random_state=0,
verbosity=0,
iter_sample=None,
split=0.5
):
self.estimator=estimator
self.auditor_type=auditor_type
self.metric=metric
self.alpha=alpha
self.n_bins=n_bins
self.bin_scaling=bin_scaling
self.gamma=gamma
self.rho=rho
self.eta=eta
self.max_iters=max_iters
self.random_state=random_state
self.verbosity=verbosity
self.iter_sample=iter_sample
self.split=split
def __name__(self):
if self.metric=='PMC':
return 'Proportional Multicalibrator'
return 'MultiCalibrator'
def fit(self, X, y):
"""A reference implementation of a fitting function for a classifier.
Parameters
----------
X : array-like, shape (n_samples, n_features)
The training input samples.
y : array-like, shape (n_samples,)
The target values. An array of int.
Returns
-------
self : object
Returns self.
"""
logger = logging.getLogger(__name__)
logger.setLevel({
0:logging.WARN,
1:logging.INFO,
2:logging.DEBUG
}
[self.verbosity]
)
# clear statistics from previous calls to fit
if hasattr(self, 'stats_'):
del self.stats_
# Check that X and y have correct shape
# X, y = check_X_y(X, y)
# Store the classes seen during fit
self.classes_ = unique_labels(y)
assert len(self.classes_) == 2, "Only binary classification supported"
# assert self.split > 0.0 and self.split <= 1.0
if self.split == 0.0 or self.split == 1.0:
train_X = X
test_X = X
train_y = y
test_y = y
else:
train_X,test_X,train_y,test_y = \
train_test_split(X,
y,
train_size=self.split,
test_size=1-self.split,
shuffle=False,
random_state=self.random_state
)
self.est_ = self.estimator.fit(train_X, train_y)
self.X_ = test_X
self.y_ = test_y.astype(float)
if not isinstance(self.X_, pd.DataFrame):
self.X_ = pd.DataFrame(self.X_)
if not isinstance(self.y_, pd.Series):
self.y_ = pd.Series(self.y_)
self.X_ = self.X_.set_index(self.y_.index)
assert hasattr(self.est_, 'predict_proba'), ("Classifier has no"
"'predict_proba' method")
self.auditor_ = copy(self.auditor_type)
for att in vars(self):
if hasattr(self.auditor_, att):
setattr(self.auditor_, att, getattr(self,att))
# map groups to adjustments
self.adjustments_ = []
iters, n_updates = 0, 0
updated = True
# predictions
y_init = self.est_.predict_proba(self.X_)[:,1]
y_init = pd.Series(y_init, index=self.X_.index)
y_adjusted = copy(y_init)
MSE = mse(self.y_, y_init)
########################################
# initialize categories and loss metric
categories = self.auditor_.make_categories(self.X_, y_init)
# bootstrap sample self.X_,y
Xs, ys = self.X_, self.y_
init_cal_loss, _, _ = self.auditor_.loss(
self.y_,
y_adjusted,
self.X_
)
smallest_cat = len(Xs)
########################################
# main boosting loop
for i in range(self.max_iters):
if self.iter_sample == 'bootstrap':
Xs, ys, ys_pred = resample(self.X_, self.y_, y_adjusted,
random_state=self.random_state
)
else:
Xs, ys, ys_pred = self.X_, self.y_, y_adjusted
MSE = mse(ys, ys_pred)
cal_loss, p_worst_c, p_worst_idx, cats = \
self.auditor_.loss(ys, ys_pred, Xs, return_cat=True)
other_metric = 'PMC' if self.metric=='MC' else 'MC'
stats = {
'iteration':i,
'# categories': len(categories),
'smallest category': smallest_cat,
'# updates': n_updates,
self.metric: cal_loss,
'MSE': MSE,
other_metric: self.auditor_.loss(ys, ys_pred, Xs, metric=other_metric)[0],
'worst category':p_worst_c,
}
self.update_stats(stats)
logger.info(', '.join([ f'{k}: {v:.3f}' if isinstance(v,float) else f'{k}: {v}' for k,v in stats.items() ]))
# make an iterable over groups, intervals
categories = self.auditor_.categorize(Xs, ys_pred)
if self.iter_sample == None:
assert utils.category_diff(categories, cats), \
"categories don't match"
assert p_worst_c in categories.keys()
Mworst_delta = 0
pmc_adjust = 1
smallest_cat = len(Xs)
if self.verbosity > 0:
iterator = tqdm(categories.items(),
desc='updating categories',
leave=False)
else:
iterator = categories.items()
updated=False
########################################
# loop through categories (group, interval pairs)
for category, idx in iterator:
if len(idx) < smallest_cat:
smallest_cat = len(idx)
# calc average predicted risk for the group
rbar = ys_pred.loc[idx].mean()
# calc actual average risk for the group
ybar = ys.loc[idx].mean()
# delta
delta = ybar - rbar
# set alpha
alpha = self.alpha
# set the PMC adjustment if needed
if self.metric=='PMC':
pmc_adjust = max(ybar,self.rho)
alpha *= pmc_adjust
logger.debug(
f'category:{category}, '
f'rbar:{rbar:3f}, '
f'ybar:{ybar:3f}, '
f'delta:{delta:3f}, '
f'alpha:{alpha:3f}, '
f'delta/pmc_adjust:{np.abs(delta)/pmc_adjust:.3f}'
)
if ((self.metric=='MC' and np.abs(delta) > Mworst_delta)
or (self.metric=='PMC'
and np.abs(delta)/pmc_adjust > Mworst_delta)
):
Mworst_delta=np.abs(delta)
Mworst_c = category
Mworst_idx = idx
if self.metric=='PMC':
Mworst_delta /= pmc_adjust
if np.abs(delta) > alpha:
update = self.eta*delta
logger.debug(f'Updating category:{category}')
# update estimates
y_adjusted.loc[idx] += update
if updated == False:
# initialize adjustment list
self.adjustments_.append({})
# store adjustment
self.adjustments_[-1][category] = update
updated=True
n_updates += 1
# make sure update was good
assert not any(y_adjusted.isna())
iters += 1
if iters >= self.max_iters:
logger.info('max iters reached')
break
# constrain adjusted output between 0 and 1
y_adjusted = utils.squash_series(y_adjusted)
assert y_adjusted.max() <= 1.0 and y_adjusted.min() >= 0.0
new_cal_loss, worst_c, worst_idx = self.auditor_.loss(
ys,
ys_pred,
Xs
)
logger.debug(f'worst category from multicalibrator: '
f'{Mworst_c}, alpha = {Mworst_delta}')
logger.debug(f'worst category from auditor: '
f'{worst_c}, alpha = {new_cal_loss}')
if iters >= self.max_iters:
logger.warn('max_iters was reached before alpha termination'
' criterion was satisfied.')
break
if self.iter_sample=='bootstrap' and not updated:
total_cal_loss, _, _ = self.auditor_.loss(
self.y_,
y_adjusted,
self.X_
)
if total_cal_loss < self.alpha:
break
elif not updated:
logger.info('no updates this round. breaking')
break
else:
cal_diff = cal_loss - new_cal_loss
## end for loop
########################################
logger.info(f'finished. updates: {n_updates}')
y_end = pd.Series(self.predict_proba(self.X_)[:,1], index=self.X_.index)
np.testing.assert_allclose(y_adjusted, y_end, rtol=1e-04)
init_MC = self.auditor_.loss(self.y_, y_init, self.X_, metric='MC')[0]
final_MC = self.auditor_.loss(self.y_, y_end, self.X_, metric='MC')[0]
init_PMC = self.auditor_.loss(self.y_, y_init, self.X_, metric='PMC')[0]
final_PMC = self.auditor_.loss(self.y_, y_end, self.X_, metric='PMC')[0]
logger.info(f'initial multicalibration: {init_MC:.3f}')
logger.info(f'final multicalibration: {final_MC:.3f}')
logger.info(f'initial proportional multicalibration: {init_PMC:.3f}')
logger.info(f'final proportional multicalibration: {final_PMC:.3f}')
self.stats_ = pd.DataFrame(self.stats_)
self.n_updates_ = n_updates
# Return the classifier
return self
def update_stats(self, stats):
if not hasattr(self, 'stats_'):
# self.stats_ = pd.DataFrame()
self.stats_ = [stats]
else:
self.stats_.append(stats)
# self.stats_ = pd.concat([self.stats_,
# pd.DataFrame(stats, index=[stats['iteration']])
# ])
def predict_proba(self, X):
""" A reference implementation of a prediction for a classifier.
Parameters
----------
X : array-like, shape (n_samples, n_features)
The input samples.
Returns
-------
y : ndarray, shape (n_samples,)
The label for each sample is the label of the closest sample
seen during fit.
"""
# Check if fit had been called
# check_is_fitted(self, ['X_', 'y_'])
# Input validation
# X = check_array(X)
# y_pred = self.est_.predict_proba(X)[:,1]
y_pred = pd.Series(self.est_.predict_proba(X)[:,1],
index=X.index)
for adjust_iter in self.adjustments_:
if self.iter_sample == 'bootstrap':
Xs, ys_pred = resample(X, y_pred,
random_state=self.random_state
)
else:
Xs, ys_pred = X, y_pred
categories = self.auditor_.categorize(Xs, ys_pred)
for category, update in adjust_iter.items():
if category in categories.keys():
idx = categories[category]
y_pred.loc[idx] += update
# y_pred.loc[idx] = utils.squash_series(y_pred.loc[idx])
# else:
# logger.warn(f'y_pred missing category {category}')
y_pred = utils.squash_series(y_pred)
# y_pred = utils.squash_series(y_pred)
# ipdb.set_trace()
rety = np.vstack((1-y_pred, y_pred)).T
return rety
def predict(self, X):
""" A reference implementation of a prediction for a classifier.
Parameters
----------
X : array-like, shape (n_samples, n_features)
The input samples.
Returns
-------
y : ndarray, shape (n_samples,)
The label for each sample is the label of the closest sample
seen during fit.
"""
# Check is fit had been called
check_is_fitted(self, ['X_', 'y_'])
# Input validation
# X = check_array(X)
return self.predict_proba(X)[:,1] > 0.5
def score(self, X, y, **kwargs):
"""Return auditor score
Parameters
----------
X : array-like, shape (n_samples, n_features)
The input samples.
y : ndarray, shape (n_samples,)
The label for each sample is the label of the closest sample
seen during fit.
kwargs: dictionary
arguments passed to the scoring function.
Returns
-------
The negative (proportional) multicalibration loss
"""
if 'groups' in kwargs.keys():
groups = kwargs['groups']
else:
groups = self.auditor_.groups
if self.metric=='MC':
return multicalibration_score(self,X,y,groups,**kwargs)
else:
return proportional_multicalibration_score(self,X,y,groups,**kwargs)