-
Notifications
You must be signed in to change notification settings - Fork 2
/
experiments.py
537 lines (410 loc) · 18.5 KB
/
experiments.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
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
from data import texts_vote_lists_truths_by_topic_id
import numpy as np
from itertools import izip, ifilter, chain, imap
import random
from plots import plot_learning_curve, plot_lines
from scipy.stats import nanmean
from sklearn.externals.joblib import Parallel, delayed
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity
import datetime
from scipy.stats import ttest_ind
import sys
from scipy.special import logit, expit
from sklearn import gaussian_process
import gc
from scipy import sparse, io
import subprocess
import datetime
import shutil
N_CORES = 1
MATLAB_TEMP_DIR = '/scratch/' # Local scratch folder that gets deleted automatically when the job is done
def get_accuracy(estimates, truths):
"""
This gets boolean lists of estimates and truths with corresponding
positions and returns a fraction of matching items
If any of the pair (estimate, truth) is None, it is disregarded
Symmetric w.r.t. argument order
"""
pairs = izip(estimates, truths)
pairs_without_Nones = ifilter(lambda x: None not in x, pairs)
matching = [x == y for (x, y) in pairs_without_Nones]
if not matching:
return None
else:
return np.mean(matching)
unit_to_bool_random = lambda x: random.choice([True, False]) if (x == 0.5 or x is None) else (x > 0.5)
get_mean_vote = lambda vote_list: np.mean(vote_list) if vote_list else None
def p_majority_vote(texts, vote_lists):
""" This is how all confidence functions should look like
Return value in [0, 1] means certainty in document's relevance
"""
return imap(get_mean_vote, vote_lists)
def est_majority_vote(texts, vote_lists, X, text_similarity):
""" This is how all estimator functions should look like
"""
return ( unit_to_bool_random(conf) for conf in p_majority_vote(texts, vote_lists) )
def copy_and_shuffle_sublists(list_of_lists):
""" Get a copy with all lists shuffled
Use this to draw 'random' votes with .pop()
"""
return [sorted(l, key=lambda x: random.random()) for l in list_of_lists]
def get_accuracy_sequence(estimator, n_votes_to_sample, texts,
vote_lists, truths, X, text_similarity, idx=None, return_final=False, *args):
""" Randomly sample votes and re-calculate estimates.
"""
random.seed() # This is using system time
unknown_votes = copy_and_shuffle_sublists(vote_lists)
known_votes = [ [] for _ in unknown_votes ]
estimates = [None for _ in vote_lists]
accuracy_sequence = [None] * n_votes_to_sample
for index in xrange(n_votes_to_sample):
# Draw one vote for a random document
updated_doc_idx = random.randrange(len(vote_lists))
if not unknown_votes[updated_doc_idx]:
# We ran out of votes for this document, diregard this sequence
return None
vote = unknown_votes[updated_doc_idx].pop()
known_votes[updated_doc_idx].append(vote)
# Recalculate all the estimates for the sake of consistency
estimates = estimator(texts, known_votes, X, text_similarity, *args)
# Calucate the accuracy_sequence
accuracy_sequence[index] = get_accuracy(estimates, truths)
return accuracy_sequence
def get_indexes_of_sublists_smaller_than(length, list_of_lists):
"""
>>> get_indexes_of_sublists_smaller_than(4, [[1,1], [2,2,2,2], [3,3,3]])
[0, 2]
"""
return [index for index, element in enumerate(list_of_lists) if len(element) < length]
def boolean_slice(l, take_bool):
"""
>>> boolean_slice(range(5), [False, False, False, False, True])
[4]
>>> boolean_slice(['a','b','c','d','e'], [False, True, False, False, True])
['b', 'e']
"""
return [el for el, take in zip(l, take_bool) if take]
def get_indexes_with_neighborhood_votes_less_than(votes_required, vote_lists,
text_similarity, sufficient_similarity):
""" For every document, if all documents which are closer than sufficient_similarity
have cumulatively less than votes_required votes, it's index is added to resulting list
"""
result_idx = []
for idx in xrange(len(vote_lists)):
if len(vote_lists[idx]) < votes_required:
# Join votes of all documents closer than sufficient_similarity and see if it's enough
similarities = text_similarity[:, idx]
similarities[idx] = 0.0
# Boolean indexes of neighbors are: similarities > sufficient_similarity
lengths = [len(vote_list) for vote_list in boolean_slice(vote_lists, similarities > sufficient_similarity)]
if sum(lengths) < votes_required:
result_idx.append(idx)
return result_idx
def get_accuracy_sequence_active(estimator, n_votes_to_sample, texts,
vote_lists, truths, text_similarity, active_pars, idx=None, return_final=False, *args):
""" Active version of the function above
"""
unknown_votes = copy_and_shuffle_sublists(vote_lists)
known_votes = [ [] for _ in unknown_votes ]
estimates = [None for _ in vote_lists]
accuracy_sequence = [None] * n_votes_to_sample
(votes_required, sufficient_similarity) = active_pars
for index in xrange(n_votes_to_sample):
if sufficient_similarity:
# Count all sufficiently similar documents' votes together
interesting_idx = get_indexes_with_neighborhood_votes_less_than(votes_required,
vote_lists, text_similarity, sufficient_similarity)
else:
# Just get votes_required votes per document
interesting_idx = get_indexes_of_sublists_smaller_than(votes_required, known_votes)
if interesting_idx:
# There are still documents to fill, pick a random
updated_doc_idx = random.choice(interesting_idx)
else:
# All documents have required number of votes, pick random from all
updated_doc_idx = random.randrange(len(vote_lists))
if not unknown_votes[updated_doc_idx]:
# We ran out of votes for this document, diregard this sequence
return None
vote = unknown_votes[updated_doc_idx].pop()
known_votes[updated_doc_idx].append(vote)
# Recalculate all the estimates for the sake of consistency
estimates = estimator(texts, known_votes, X, text_similarity, *args)
# Calucate the accuracy_sequence
accuracy_sequence[index] = get_accuracy(estimates, truths)
return accuracy_sequence
def index_sublist_items(list_of_lists):
"""
>>> a = [[1, 2], [65, 66], [12, 13, 14]]
>>> list(index_sublist_items(a))
[(0, 1), (0, 2), (1, 65), (1, 66), (2, 12), (2, 13), (2, 14)]
"""
indexed_items = [ [ (idx, list_el) for list_el in l ]
for idx, l in enumerate(list_of_lists) ]
return list(chain(*indexed_items))
def plot_learning_curves_for_topic(topic_id, n_runs, votes_per_doc, estimators_dict, comment=None):
texts, vote_lists, truths = texts_vote_lists_truths_by_topic_id[topic_id]
n_documents = len(texts)
vectorizer = TfidfVectorizer()
X = vectorizer.fit_transform(texts)
text_similarity = cosine_similarity(X)
min_votes_per_doc, max_votes_per_doc = votes_per_doc
start_idx, stop_idx = int(min_votes_per_doc * n_documents), int(max_votes_per_doc * n_documents)
x = np.arange(float(start_idx), float(stop_idx)) / n_documents
estimator_y = {}
for estimator_name, estimator_and_args in estimators_dict.iteritems():
print 'Calculating for %s' % estimator_name
estimator, args, active_pars = estimator_and_args
if active_pars is None:
sequences = Parallel(n_jobs=N_CORES)( delayed(get_accuracy_sequence)(estimator, stop_idx, texts,
vote_lists, truths, X, text_similarity, idx, False, *args) for idx in xrange(n_runs) )
else:
sequences = Parallel(n_jobs=N_CORES)( delayed(get_accuracy_sequence_active)(estimator, stop_idx, texts,
vote_lists, truths, text_similarity, active_pars, idx, False, *args) for idx in xrange(n_runs) )
good_slices = [ s[start_idx:] for s in sequences if s is not None ]
if good_slices:
results = np.vstack(good_slices)
# Pickling is not necessary yet
'''
begin_accuracies = results[:, 0]
middle_accuracies = results[:, int(results.shape[1] / 2)]
end_accuracies = results[:, -1]
begin_accuracies.dump("pickles/%s-%s-begin-accuracies---.pkl" % (topic_id, estimator_name) )
'''
estimator_y[estimator_name] = np.mean(results, axis=0)
else:
print 'Query %s is not represented with estimator %s' % (topic_id, estimator_name)
if comment:
title = 'Query %s, %s runs, %s' % (topic_id, n_runs, comment)
else:
title = 'Query %s, %s runs' % (topic_id, n_runs)
plot_learning_curve(title, x, estimator_y, 'Votes per document', 'Accuracy')
def plot_learning_curves_across_topics(n_runs, start_idx, stop_idx, estimators_dict, comment=None):
"""
TODO Most probably buggy
"""
for topic_id, data in texts_vote_lists_truths_by_topic_id.iteritems():
print 'Loading topic %s' % topic_id
texts, vote_lists, truths = data
n_documents = len(texts)
vectorizer = TfidfVectorizer()
tfidf = vectorizer.fit_transform(texts)
text_similarity = cosine_similarity(tfidf)
x = np.arange(start_idx, stop_idx)
y_by_estimator = dict( (estimator, []) for estimator in estimators_dict.keys() )
for estimator_name, estimator_and_args in estimators_dict.iteritems():
print 'Calculating for %s' % estimator_name
estimator, args, active_pars = estimator_and_args
if active_pars is None:
sequences = Parallel(n_jobs=4)( delayed(get_accuracy_sequence)(estimator, stop_idx, texts,
vote_lists, truths, text_similarity, idx, False, *args) for idx in xrange(n_runs) )
else:
sequences = Parallel(n_jobs=4)( delayed(get_accuracy_sequence_active)(estimator, stop_idx, texts,
vote_lists, truths, text_similarity, active_pars, idx, False, *args) for idx in xrange(n_runs) )
good_slices = [ s[start_idx:] for s in sequences if s is not None ]
if good_slices:
results = np.vstack(good_slices)
begin_accuracies = results[:, 0]
end_accuracies = results[:, -1]
begin_accuracies.dump("pickles/%s-%s-begin-accuracies--.pkl" % (topic_id, estimator_name) )
end_accuracies.dump("pickles/%s-%s-end-accuracies--.pkl" % (topic_id, estimator_name))
# We will then need to vstack and avg though all the topic accuracies for each estimator
y_by_estimator[estimator_name].append( np.mean(results, axis=0) )
else:
print 'Topic %s is not represented with estimator %s' % (topic_id, estimator_name)
result_by_estimator = {}
for estimator_name, mean_accuracy_sequences in y_by_estimator.iteritems():
if mean_accuracy_sequences:
to_avg = np.vstack(mean_accuracy_sequences)
result_by_estimator[estimator_name] = np.mean(to_avg, axis=0)
else:
print "Nope"
if comment:
title = 'Across topics, %s runs, %s' % (n_runs, comment)
else:
title = 'Across topics, %s runs' % topic_id
plot_learning_curve(title, x, result_by_estimator, 'Votes sampled', 'Accuracy')
def t_test_accuracy(topic_id, n_runs, estimator_params_votes_per_doc_tuples):
""" Test if accuracy for estimators with given parameters is
significantly better than that of the first estimator in the tuple
"""
texts, vote_lists, truths = texts_vote_lists_truths_by_topic_id[topic_id]
vectorizer = TfidfVectorizer()
text_similarity = cosine_similarity(vectorizer.fit_transform(texts))
accuracy_arrays = []
for estimator, args, votes_per_doc in estimator_params_votes_per_doc_tuples:
stop_idx = votes_per_doc * len(texts)
# Now get n_runs accuracies and put then into numpy arrays
accuracies = Parallel(n_jobs=4)( delayed(get_accuracy_sequence)(estimator, stop_idx, texts,
vote_lists, truths, text_similarity, idx, True, *args) for idx in xrange(n_runs) )
accuracy_arrays.append( np.array( filter(lambda x: x is not None, accuracies) ) )
# Baseline
result_row = []
result_row.append( "%0.2f" % np.mean(accuracy_arrays[0]) )
# T-tests
for accuracy_array in accuracy_arrays[1:]:
_, pval = ttest_ind(accuracy_array, accuracy_arrays[0], equal_var=False)
significance_indicator = lambda p: "*" if p < 0.01 else " "
is_better = "$" if np.mean(accuracy_array) > np.mean(accuracy_arrays[0]) else " "
result_row.append( "%0.2f %s %s" % (np.mean(accuracy_array), significance_indicator(pval), is_better))
return "|".join(result_row)
def get_p_and_var(vote_list):
if not vote_list:
return None, None
p = get_mean_vote(vote_list)
if p is None:
return None, None
n = len(vote_list)
# Variance is None if there is only one vote
var = p * (1 - p) / n if n > 1 else None
return p, var
def is_doc_variance_better(doc_var, neighbor_var):
""" Returns True if the document variance is less than leighbor variance
"""
if neighbor_var is None:
return True
else:
if doc_var is None:
return False
else:
return (doc_var < neighbor_var)
def get_sufficient_similarity(n):
return 1 - 1 / float(n - 1) if n > 1 else 0
def p_majority_vote_or_nn(texts, vote_lists, text_similarity, sufficient_similarity):
""" If the nearest neighbor's similarity to you is bigger than sufficient_similarity
and variance smaller than yours, take neighbor's conf instead of yours
if sufficient_similarity is None it's selected by number of votes
"""
result_p = []
for doc_index, vote_list in enumerate(vote_lists):
doc_p, doc_var = get_p_and_var(vote_list)
similarities = text_similarity[:, doc_index]
similarities[doc_index] = 0
nn_similarity = similarities.max()
if sufficient_similarity is None:
# Select similarity threshold depending on amount of votes
sufficient_similarity = get_sufficient_similarity(len(vote_list))
if nn_similarity > sufficient_similarity:
nn_index = similarities.argmax()
nn_p, nn_var = get_p_and_var(vote_lists[nn_index])
p = doc_p if is_doc_variance_better(doc_var, nn_var) else nn_p
else:
p = doc_p
result_p.append(p)
return result_p
def est_majority_vote_or_nn(texts, vote_lists, X, text_similarity, sufficient_similarity):
return ( unit_to_bool_random(p) for p
in p_majority_vote_or_nn(texts, vote_lists, text_similarity, sufficient_similarity) )
def p_majority_vote_with_nn(texts, vote_lists, text_similarity, sufficient_similarity):
result_p = []
for doc_index, doc_vote_list in enumerate(vote_lists):
similarities = text_similarity[:, doc_index]
similarities[doc_index] = 0
nn_similarity = similarities.max()
if nn_similarity > sufficient_similarity:
# Join their votes
nn_vote_list = vote_lists[ similarities.argmax() ]
joint_vote_list = doc_vote_list + nn_vote_list
p, var = get_p_and_var(joint_vote_list)
else:
p, var = get_p_and_var(doc_vote_list)
result_p.append(p)
return result_p
def est_majority_vote_with_nn(texts, vote_lists, X, text_similarity, sufficient_similarity):
return ( unit_to_bool_random(p) for p
in p_majority_vote_with_nn(texts, vote_lists, text_similarity, sufficient_similarity) )
def p_merge_enough_votes(texts, vote_lists, text_similarity, votes_required):
""" Merge votes from nearest neighbors until a sufficient amount of votes
is reached
"""
result_p = []
for doc_index, doc_vote_list in enumerate(vote_lists):
if len(doc_vote_list) >= votes_required:
p, var = get_p_and_var(doc_vote_list)
else:
# Gather votes around from neighbors
similarities = text_similarity[:, doc_index]
similarities[doc_index] = 0
decreasing_order_idx = np.argsort(similarities)[::-1]
# Fill the vote list until it's big enough
vote_list = doc_vote_list[:]
for neighbor_idx in decreasing_order_idx:
vote_list += vote_lists[neighbor_idx]
if len(vote_list) >= votes_required:
break
# Derive estimate from that vote list
p, var = get_p_and_var(vote_list)
result_p.append(p)
return result_p
def est_merge_enough_votes(texts, vote_lists, X, text_similarity, votes_required):
return ( unit_to_bool_random(p) for p
in p_merge_enough_votes(texts, vote_lists, text_similarity, votes_required) )
def p_gp(texts, vote_lists, X, text_similarity):
""" Smooth estimates with Gaussian Processes using linear correlation function
Extrapolate to get estimates for unknown values as well
"""
# for every vote in a vote list we have to get a vector of features
labels = []
feature_vectors = []
bool_to_plus_minus_one = lambda b: 1.0 if b else -1.0
for doc_idx, vote_list in enumerate(vote_lists):
for vote in vote_list:
labels.append( bool_to_plus_minus_one(vote) )
feature_vectors.append( X[doc_idx, :] )
X_new = sparse.vstack(feature_vectors)
y = np.array(labels, dtype=np.float64)[np.newaxis].T
# Preparing a temp folder for running MATLAB
random.seed()
folder_id = random.randint(0, sys.maxint)
matlab_folder_name = MATLAB_TEMP_DIR + 'matlab_' + str(folder_id)
shutil.copytree('matlab', matlab_folder_name)
io.savemat(matlab_folder_name + '/train.mat', mdict = {'x' : X_new, 'y' : y})
io.savemat(matlab_folder_name + '/test.mat', mdict = {'t' : X })
print 'Running MATLAB, started %s' % str(datetime.datetime.now())
code = subprocess.call(['matlab/run_in_dir.sh', matlab_folder_name])
if code != 0:
raise OSError('MATLAB code couldn\'t run')
print 'Finished %s' % str(datetime.datetime.now())
print 'Getting the matrix'
# Loads a `prob` vector
prob_location = matlab_folder_name + '/prob.mat'
print 'Loading prob vector from %s' % prob_location
mat_objects = io.loadmat(prob_location)
prob = mat_objects['prob']
result = prob[:, 0]
"""
print 'prob.shape'
print prob.shape
print 'y[30:]'
print y[30:]
print 'prob[30:]'
print prob[30:]
print 'X.shape'
print X.shape
print 'X_new.shape'
print X_new.shape
print 'result'
print result
"""
# Remove the temp folder
# Not necessary for local scratch
# shutil.rmtree(matlab_folder_name)
return result
def est_gp(texts, vote_lists, X, text_similarity):
return ( unit_to_bool_random(p) for p
in p_gp(texts, vote_lists, X, text_similarity) )
if __name__ == "__main__":
loser_topics = ['20644','20922']
print "started job at %s" % datetime.datetime.now()
for topic_id in ['20910']:
print 'topic %s' % topic_id
plot_learning_curves_for_topic(topic_id, 1000, (1.0, 3.0), {
'MajorityVote' : (est_majority_vote, [], None),
# 'MajorityVote,Active(3)' : (est_majority_vote, [], [ 3, None ]),
# 'MergeEnoughVotes(1),Active(1)' : (est_merge_enough_votes, [ 1 ], [ 1, None ]),
# 'MergeEnoughVotes(1)' : (est_merge_enough_votes, [ 1 ], None),
# 'GP(1)' : (est_gp, [ 1 ], None),
'GP' : (est_gp, [ None ], None),
}, comment="")
print "finished job at %s" % datetime.datetime.now()