/
utils.py
500 lines (408 loc) · 18.2 KB
/
utils.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
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
"""
Different helpful functions, objects, methods are collected here.
"""
from __future__ import division, print_function, absolute_import
from collections import OrderedDict
import time
import numpy
import pandas
from sklearn.utils.validation import column_or_1d
from sklearn.metrics import roc_curve
def weighted_quantile(array, quantiles, sample_weight=None, array_sorted=False, old_style=False):
"""Computing quantiles of array. Unlike the numpy.percentile, this function supports weights,
but it is inefficient and performs complete sorting.
:param array: distribution, array of shape [n_samples]
:param quantiles: floats from range [0, 1] with quantiles of shape [n_quantiles]
:param sample_weight: optional weights of samples, array of shape [n_samples]
:param array_sorted: if True, the sorting step will be skipped
:param old_style: if True, will correct output to be consistent with numpy.percentile.
:return: array of shape [n_quantiles]
Example:
>>> weighted_quantile([1, 2, 3, 4, 5], [0.5])
Out: array([ 3.])
>>> weighted_quantile([1, 2, 3, 4, 5], [0.5], sample_weight=[3, 1, 1, 1, 1])
Out: array([ 2.])
"""
array = numpy.array(array)
quantiles = numpy.array(quantiles)
sample_weight = check_sample_weight(array, sample_weight)
assert numpy.all(quantiles >= 0) and numpy.all(quantiles <= 1), 'Percentiles should be in [0, 1]'
if not array_sorted:
array, sample_weight = reorder_by_first(array, sample_weight)
weighted_quantiles = numpy.cumsum(sample_weight) - 0.5 * sample_weight
if old_style:
# To be convenient with numpy.percentile
weighted_quantiles -= weighted_quantiles[0]
weighted_quantiles /= weighted_quantiles[-1]
else:
weighted_quantiles /= numpy.sum(sample_weight)
return numpy.interp(quantiles, weighted_quantiles, array)
def reorder_by_first(*arrays):
"""
Applies the same permutation to all passed arrays,
permutation sorts the first passed array
"""
arrays = check_arrays(*arrays)
order = numpy.argsort(arrays[0])
return [arr[order] for arr in arrays]
def check_sample_weight(y_true, sample_weight):
"""Checks the weights, if None, returns array.
:param y_true: labels (or any array of length [n_samples])
:param sample_weight: None or array of length [n_samples]
:return: numpy.array of shape [n_samples]
"""
if sample_weight is None:
return numpy.ones(len(y_true), dtype=numpy.float)
else:
sample_weight = numpy.array(sample_weight, dtype=numpy.float)
assert len(y_true) == len(sample_weight), \
"The length of weights is different: not {0}, but {1}".format(len(y_true), len(sample_weight))
return sample_weight
class Flattener(object):
def __init__(self, data, sample_weight=None):
"""
Prepares normalization function for some set of values
transforms it to uniform distribution from [0, 1].
:param data: predictions
:type data: list or numpy.array
:param sample_weight: weights
:type sample_weight: None or list or numpy.array
:return func: normalization function
Example of usage:
>>> normalizer = Flattener(signal)
>>> hist(normalizer(background))
>>> hist(normalizer(signal))
"""
sample_weight = check_sample_weight(data, sample_weight=sample_weight)
data = column_or_1d(data)
assert numpy.all(sample_weight >= 0.), 'sample weight must be non-negative'
self.data, sample_weight = reorder_by_first(data, sample_weight)
self.predictions = numpy.cumsum(sample_weight) / numpy.sum(sample_weight)
def __call__(self, data):
return numpy.interp(data, self.data, self.predictions)
class Binner(object):
def __init__(self, values, bins_number):
"""
Binner is a class that helps to split the values into several bins.
Initially an array of values is given, which is then splitted into 'bins_number' equal parts,
and thus we are computing limits (boundaries of bins).
"""
percentiles = [i * 100.0 / bins_number for i in range(1, bins_number)]
self.limits = numpy.percentile(values, percentiles)
def get_bins(self, values):
"""Given the values of feature, compute the index of bin
:param values: array of shape [n_samples]
:return: array of shape [n_samples]
"""
return numpy.searchsorted(self.limits, values)
def set_limits(self, limits):
"""Change the thresholds inside bins."""
self.limits = limits
@property
def bins_number(self):
""":return: number of bins"""
return len(self.limits) + 1
def split_into_bins(self, *arrays):
"""
:param arrays: data to be splitted, the first array corresponds
:return: sequence of length [n_bins] with values corresponding to each bin.
"""
values = arrays[0]
for array in arrays:
assert len(array) == len(values), "passed arrays have different length"
bins = self.get_bins(values)
result = []
for bin in range(len(self.limits) + 1):
indices = bins == bin
result.append([numpy.array(array)[indices] for array in arrays])
return result
def calc_ROC(prediction, signal, sample_weight=None, max_points=10000):
"""
Calculate roc curve, returns limited number of points.
This is needed for interactive plots, which suffer
:param prediction: predictions
:type prediction: numpy.ndarray or list
:param signal: true labels
:type signal: array or list
:param sample_weight: weights
:type sample_weight: None or array or list
:param int max_points: maximum of used points on roc curve
:return: (tpr, tnr), (err_tnr, err_tpr), thresholds
"""
sample_weight = numpy.ones(len(signal)) if sample_weight is None else sample_weight
prediction, signal, sample_weight = check_arrays(prediction, signal, sample_weight)
assert set(signal) == {0, 1}, "the labels should be 0 and 1, labels are " + str(set(signal))
fpr, tpr, thresholds = roc_curve(signal, prediction, sample_weight=sample_weight)
tpr = numpy.insert(tpr, 0, [0.])
fpr = numpy.insert(fpr, 0, [0.])
thresholds = numpy.insert(thresholds, 0, [thresholds[0] + 1.])
tnr = 1 - fpr
weight_bck = sample_weight[signal == 0]
weight_sig = sample_weight[signal == 1]
err_tnr = numpy.sqrt(tnr * (1 - tnr) * numpy.sum(weight_bck ** 2)) / numpy.sum(weight_bck)
err_tpr = numpy.sqrt(tpr * (1 - tpr) * numpy.sum(weight_sig ** 2)) / numpy.sum(weight_sig)
if len(prediction) > max_points:
sum_weights = numpy.cumsum((fpr + tpr) / 2.)
sum_weights /= sum_weights[-1]
positions = numpy.searchsorted(sum_weights, numpy.linspace(0, 1, max_points))
tpr, tnr = tpr[positions], tnr[positions]
err_tnr, err_tpr = err_tnr[positions], err_tpr[positions]
thresholds = thresholds[positions]
return (tpr, tnr), (err_tnr, err_tpr), thresholds
def calc_feature_correlation_matrix(df, weights=None):
"""
Calculate correlation matrix
:param pandas.DataFrame df: data of shape [n_samples, n_features]
:param weights: weights of shape [n_samples] (optional)
:return: correlation matrix for dataFrame of shape [n_features, n_features]
:rtype: numpy.ndarray
"""
values = numpy.array(df)
weights = check_sample_weight(df, sample_weight=weights)
means = numpy.average(values, weights=weights, axis=0)
values -= means
covariation = values.T.dot(values * weights[:, None])
diag = covariation.diagonal()
return covariation / numpy.sqrt(diag)[:, None] / numpy.sqrt(diag)[None, :]
def calc_hist_with_errors(x, weight=None, bins=60, normed=True, x_range=None, ignored_sideband=0.0):
"""
Calculate data for error bar (for plot pdf with errors)
:param x: data
:type x: list or numpy.array
:param weight: weights
:type weight: None or list or numpy.array
:return: tuple (x-points (list), y-points (list), y points errors (list), x points errors (list))
"""
weight = numpy.ones(len(x)) if weight is None else weight
x, weight = check_arrays(x, weight)
if x_range is None:
x_range = numpy.percentile(x, [100 * ignored_sideband, 100 * (1 - ignored_sideband)])
ans, bins = numpy.histogram(x, bins=bins, normed=normed, weights=weight, range=x_range)
yerr = []
normalization = 1.0
if normed:
normalization = float(len(bins) - 1) / float(sum(weight)) / (x_range[1] - x_range[0])
for i in range(len(bins) - 1):
weight_bin = weight[(x > bins[i]) * (x <= bins[i + 1])]
yerr.append(numpy.sqrt(sum(weight_bin * weight_bin)) * normalization)
bins_mean = [0.5 * (bins[i] + bins[i + 1]) for i in range(len(ans))]
xerr = [0.5 * (bins[i + 1] - bins[i]) for i in range(len(ans))]
return bins_mean, ans, yerr, xerr
def get_efficiencies(prediction, spectator, sample_weight=None, bins_number=20,
thresholds=None, errors=False, ignored_sideband=0.0):
"""
Construct efficiency function dependent on spectator for each threshold
Different score functions available: Efficiency, Precision, Recall, F1Score,
and other things from sklearn.metrics
:param prediction: list of probabilities
:param spectator: list of spectator's values
:param bins_number: int, count of bins for plot
:param thresholds: list of prediction's threshold
(default=prediction's cuts for which efficiency will be [0.2, 0.4, 0.5, 0.6, 0.8])
:return:
if errors=False
OrderedDict threshold -> (x_values, y_values)
if errors=True
OrderedDict threshold -> (x_values, y_values, y_err, x_err)
All the parts: x_values, y_values, y_err, x_err are numpy.arrays of the same length.
"""
prediction, spectator, sample_weight = \
check_arrays(prediction, spectator, sample_weight)
spectator_min, spectator_max = weighted_quantile(spectator, [ignored_sideband, (1. - ignored_sideband)])
mask = (spectator >= spectator_min) & (spectator <= spectator_max)
spectator = spectator[mask]
prediction = prediction[mask]
bins_number = min(bins_number, len(prediction))
sample_weight = sample_weight if sample_weight is None else numpy.array(sample_weight)[mask]
if thresholds is None:
thresholds = [weighted_quantile(prediction, quantiles=1 - eff, sample_weight=sample_weight)
for eff in [0.2, 0.4, 0.5, 0.6, 0.8]]
binner = Binner(spectator, bins_number=bins_number)
if sample_weight is None:
sample_weight = numpy.ones(len(prediction))
bins_data = binner.split_into_bins(spectator, prediction, sample_weight)
bin_edges = numpy.array([spectator_min] + list(binner.limits) + [spectator_max])
xerr = numpy.diff(bin_edges) / 2.
result = OrderedDict()
for threshold in thresholds:
x_values = []
y_values = []
N_in_bin = []
for num, (masses, probabilities, weights) in enumerate(bins_data):
y_values.append(numpy.average(probabilities > threshold, weights=weights))
N_in_bin.append(numpy.sum(weights))
if errors:
x_values.append((bin_edges[num + 1] + bin_edges[num]) / 2.)
else:
x_values.append(numpy.mean(masses))
x_values, y_values, N_in_bin = check_arrays(x_values, y_values, N_in_bin)
if errors:
result[threshold] = (x_values, y_values, numpy.sqrt(y_values * (1 - y_values) / N_in_bin), xerr)
else:
result[threshold] = (x_values, y_values)
return result
def train_test_split(*arrays, **kw_args):
"""
Does the same thing as sklearn.cross_validation.train_test_split.
Additionally has 'allow_none' parameter.
:param arrays: arrays to split with same first dimension
:type arrays: list[numpy.array] or list[pandas.DataFrame]
:param bool allow_none: default False, is set to True, allows
non-first arguments to be None (in this case, both resulting train and test parts are None).
"""
from sklearn import cross_validation
allow_none = kw_args.pop('allow_none', False)
assert len(arrays) > 0, "at least one array should be passed"
length = len(arrays[0])
for array in arrays:
assert len(array) == length, "different size"
train_indices, test_indices = cross_validation.train_test_split(range(length), **kw_args)
result = []
for array in arrays:
if isinstance(array, pandas.DataFrame):
result.append(array.iloc[train_indices, :])
result.append(array.iloc[test_indices, :])
elif (array is None) and allow_none:
# specially for checking weights
result.append(None)
result.append(None)
else:
result.append(numpy.array(array)[train_indices])
result.append(numpy.array(array)[test_indices])
return result
def train_test_split_group(group_column, *arrays, **kw_args):
"""
Modification of :class:`train_test_split` which alters splitting rule.
:param group_column: array-like of shape [n_samples] with indices of groups,
events from one group will be kept together (all events in train or all events in test).
If `group_column` is used, train_size and test_size will refer to number of groups, not events
:param arrays: arrays to split
:type arrays: list[numpy.array] or list[pandas.DataFrame]
:param bool allow_none: default False
(useful for sample_weight - after splitting train and test of `None` are again `None`)
"""
from sklearn import cross_validation
allow_none = kw_args.pop('allow_none', None)
assert len(arrays) > 0, "at least one array should be passed"
length = len(arrays[0])
for array in arrays:
assert len(array) == length, "different size"
initial_data = numpy.array(group_column)
assert len(initial_data) == length, "group column must have the same length"
group_ids = numpy.unique(initial_data)
train_indices, test_indices = cross_validation.train_test_split(group_ids, **kw_args)
train_indices = numpy.in1d(initial_data, train_indices)
test_indices = numpy.in1d(initial_data, test_indices)
result = []
for array in arrays:
if isinstance(array, pandas.DataFrame):
result.append(array.iloc[train_indices, :])
result.append(array.iloc[test_indices, :])
elif (array is None) and allow_none:
# specially for checking weights
result.append(None)
result.append(None)
else:
result.append(numpy.array(array)[train_indices])
result.append(numpy.array(array)[test_indices])
return result
def get_columns_dict(columns):
"""
Get (new column: old column) dict expressions.
This function is used to process names of features, which can contain expressions.
:param list[str] columns: columns names
:rtype: dict
"""
result = OrderedDict()
for column in columns:
column_split = column.split(':')
assert len(column_split) < 3, 'Error in parsing feature expression {}'.format(column)
if len(column_split) == 2:
result[column_split[0].strip()] = column_split[1].strip()
else:
result[column] = column
return result
def get_columns_in_df(df, columns):
"""
Get columns in data frame using *numexpr* evaluation
:param pandas.DataFrame df: data
:param columns: necessary columns
:param columns: None or list[str]
:return: data frame with pointed columns
"""
if columns is None:
return df
columns_dict = get_columns_dict(columns)
df_new = OrderedDict()
for column_new, column in columns_dict.items():
if column in df.columns:
df_new[column_new] = df[column]
else:
# warning - this thing is known to fail in threads
# numexpr.evaluate(column, local_dict=df)
# thus we are using python engine, which is slow :(
df_new[column_new] = df.eval(column, engine='python')
return pandas.DataFrame(df_new)
def check_arrays(*arrays):
"""
Left for consistency, version of `sklearn.validation.check_arrays`
:param list[iterable] arrays: arrays with same length of first dimension.
"""
assert len(arrays) > 0, 'The number of array must be greater than zero'
checked_arrays = []
shapes = []
for arr in arrays:
if arr is not None:
checked_arrays.append(numpy.array(arr))
shapes.append(checked_arrays[-1].shape[0])
else:
checked_arrays.append(None)
assert numpy.sum(numpy.array(shapes) == shapes[0]) == len(shapes), 'Different shapes of the arrays {}'.format(
shapes)
return checked_arrays
def fit_metric(metric, *args, **kargs):
"""
Metric can implement one of two interfaces (function or object).
This function fits metrics, if it is required (by simply checking presence of fit method).
:param metric: metric function, following REP conventions
"""
if hasattr(metric, 'fit'):
metric.fit(*args, **kargs)
class Stopwatch(object):
"""
Simple tool to measure time.
If your internet connection is reliable, use %time magic.
>>> with Stopwatch() as timer:
>>> # do something here
>>> classifier.fit(X, y)
>>> # print how much time was spent
>>> print(timer)
"""
def __enter__(self):
self.start = time.time()
return self
def __exit__(self, err_type, err_value, err_traceback):
self.stop = time.time()
self.err_type = err_type
self.err_value = err_value
self.err_traceback = err_traceback
@property
def elapsed(self):
return self.stop - self.start
def __repr__(self):
result = "interval: {:.2f} sec".format(self.elapsed)
if self.err_type is not None:
message = "\nError {error} of type {error_type} was raised"
result += message.format(error=repr(self.err_value), error_type=self.err_type)
return result
def take_last(sequence):
"""
Returns the last element in sequence or raises an error
"""
empty = True
for element in sequence:
empty = False
if empty:
raise IndexError('The sequence is empty.')
else:
return element