-
Notifications
You must be signed in to change notification settings - Fork 5
/
run_diagnostics_bert.py
647 lines (600 loc) · 28 KB
/
run_diagnostics_bert.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
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
import re
import os
import argparse
import random
import copy
import numpy as np
import access_model as tp
import scipy
import scipy.stats
from io import open
from collections import Counter
import matplotlib
matplotlib.use('agg')
import matplotlib.pyplot as plt
import itertools
def process_fk(infile):
hldict = {}
inputlist = []
inputlist_shuf = []
inputlist_nw = []
inputlist_shufnw = []
tgtlist = []
cleancsv = []
with open(infile) as f:
i = 0
for line in f:
if len(line.strip('\t')) < 2: continue
it,sent1,sent2,exp,wc,bc,constraint = line.strip().split('\t')
if it == 'item': continue
sent1l = re.sub('\.','',sent1).split()
sent2l = sent2.split()
random.shuffle(sent1l)
sent2l = sent2l[-2:]
origsent = ' '.join([sent1,sent2])
shuf = ' '.join(sent1l) + '. ' + sent2
nw = sent1 + ' ' + ' '.join(sent2l)
shufnw = ' '.join(sent1l) + '. ' + ' '.join(sent2l)
for tgt,cond in ((exp,'exp'),(wc,'wc'),(bc,'bc')):
hldict[i] = {}
inputlist.append(origsent + ' [MASK]')
inputlist_shuf.append(shuf + ' [MASK]')
inputlist_nw.append(nw + ' [MASK]')
inputlist_shufnw.append(shufnw + ' [MASK]')
tgtlist.append(tgt)
hldict[i]['constraint'] = constraint
hldict[i]['tgt'] = tgt
hldict[i]['sent'] = {}
hldict[i]['sent']['orig'] = origsent
hldict[i]['sent']['shuf'] = shuf
hldict[i]['sent']['trunc'] = nw
hldict[i]['sent']['shuftrunc'] = shufnw
hldict[i]['item'] = it
hldict[i]['exp'] = [exp]
hldict[i]['cond'] = cond
i += 1
return hldict,inputlist,inputlist_shuf,inputlist_nw,inputlist_shufnw,tgtlist
def process_rr(csvfile,gen_obj=False,gen_subj=False):
inputlist = []
tgtlist = []
clozedict = {}
clozelist = []
i = 0
with open(csvfile,'r') as f:
for line in f:
linesplit = line.strip().split('\t')
item = linesplit[0]
if item == 'item' or len(linesplit) < 1:
continue
itemnum,condition = item.split('-')
sent = linesplit[1]
exp = linesplit[2].split('|')
maxcloze = float(linesplit[3])
tgt = linesplit[4].strip().split()[0]
tgtcloze = float(linesplit[5])
tgtcloze_strict = float(linesplit[6])
if gen_obj:
sent = re.sub('which .* the','which one the',sent)
if gen_subj:
sent = re.sub('which (.*) the .* had','which \g<1> the other had',sent)
masked_sent = sent + ' [MASK]'
clozedict[i] = {}
if masked_sent in inputlist:
for item in clozedict:
if clozedict[item]['sent'] == masked_sent:
clozedict[i]['maxcloze'] = maxcloze
break
else:
clozedict[i]['maxcloze'] = maxcloze
inputlist.append(masked_sent)
tgtlist.append(tgt)
clozedict[i]['sent'] = masked_sent
clozedict[i]['tgt'] = tgt
clozedict[i]['cond'] = condition
clozedict[i]['item'] = itemnum
clozedict[i]['tgtcloze'] = tgtcloze
clozedict[i]['tgtcloze_strict'] = tgtcloze_strict
clozedict[i]['exp'] = exp
clozelist.append(maxcloze)
i += 1
return clozedict,inputlist,tgtlist,clozelist
def process_fischler(infile):
nkdict = {}
inputlist = []
tgtlist = []
it = 0
i = 0
csvclean = []
with open(infile,'r') as f:
for line in f:
it,affsent,negsent,afftgt,negtgt = [e.strip() for e in line.strip().split('\t')]
affsent = re.sub(' \(.+\)','',affsent)
negsent = re.sub(' \(.+\)','',negsent)
if it == 'item': continue
for sent,tgt,cond in [(affsent,afftgt,'TA'),(negsent,afftgt,'FN'),(affsent,negtgt,'FA'),(negsent,negtgt,'TN')]:
nkdict[i] = {}
if re.match('[aeiou]',tgt): det = 'an'
else: det = 'a'
masked_sent = ' '.join([sent,det,'[MASK]'])
nkdict[i]['sent'] = masked_sent
nkdict[i]['tgt'] = tgt
nkdict[i]['item'] = it
nkdict[i]['cond'] = cond
if cond in ('TA','FA'):
nkdict[i]['exp'] = [afftgt]
else:
nkdict[i]['exp'] = [negtgt]
inputlist.append(masked_sent)
tgtlist.append(tgt)
i += 1
return inputlist,nkdict,tgtlist
def process_nk(infile):
nkdict = {}
inputlist = []
tgtlist = []
it = 0
i = 0
with open(infile,'r') as f:
for line in f:
it,affsent,negsent,afftgt,negtgt,lic = [e.strip() for e in line.strip().split('\t')]
if it == 'item': continue
for sent,tgt,cond in [(affsent,afftgt,'TA'),(negsent,afftgt,'FN'),(affsent,negtgt,'FA'),(negsent,negtgt,'TN')]:
nkdict[i] = {}
masked_sent = sent + ' [MASK]'
nkdict[i]['sent'] = masked_sent
nkdict[i]['tgt'] = tgt
nkdict[i]['item'] = it
nkdict[i]['cond'] = cond
nkdict[i]['licensing'] = lic
if cond in ('TA','FA'):
nkdict[i]['exp'] = [afftgt]
else:
nkdict[i]['exp'] = [negtgt]
inputlist.append(masked_sent)
tgtlist.append(tgt)
i += 1
return inputlist,nkdict,tgtlist
def make_conddict(clozedict):
conddict = {}
for k in clozedict:
it = clozedict[k]['item']
co = clozedict[k]['cond']
if it not in conddict:
conddict[it] = {}
conddict[it][co] = {}
conddict[it][co]['sent'] = clozedict[k]['sent']
conddict[it][co]['tgt'] = clozedict[k]['tgt']
if 'tgtcloze' in clozedict[k]:
conddict[it][co]['tgtcloze'] = clozedict[k]['tgtcloze']
if 'constraint' in clozedict[k]:
conddict[it][co]['constraint'] = clozedict[k]['constraint']
if 'licensing' in clozedict[k]:
conddict[it][co]['licensing'] = clozedict[k]['licensing']
if 'maxcloze' in clozedict[k]:
conddict[it][co]['maxcloze'] = clozedict[k]['maxcloze']
conddict[it][co]['tgtprob'] = clozedict[k]['tgtprob']
conddict[it][co]['toppreds'] = clozedict[k]['toppreds']
conddict[it][co]['topprobs'] = clozedict[k]['topprobs']
return conddict
def get_acc(scorelist):
if len(scorelist) == 0:
acc = 0
else:
acc = float(sum(scorelist))/len(scorelist)
return acc
def test_fk_acc(hldict,inputlist,tgtlist,model,tokenizer,setting,fklog,k=5,bert=True):
tok_preds,top_probs = tp.get_predictions(inputlist,model,tokenizer,k=k,bert=bert)
tok_probs = tp.get_probabilities(inputlist,tgtlist,model,tokenizer,bert=bert)
tot_score = []
by_constraint_score = {}
by_constraint_score['H'] = []
by_constraint_score['L'] = []
correct = []
used = []
for i,pr in enumerate(tok_preds):
sent = hldict[i]['sent'][setting]
hldict[i]['toppreds'] = pr
hldict[i]['tgtprob'] = tok_probs[i]
hldict[i]['topprobs'] = top_probs[i]
score = 0
if sent in used: continue
used.append(sent)
for candidate in pr:
if candidate.strip() in hldict[i]['exp']:
score = 1
if score == 1:
ctup = (hldict[i]['sent'][setting],hldict[i]['exp'],pr,hldict[i]['constraint'])
if ctup not in correct:
correct.append(ctup)
tot_score.append(score)
by_constraint_score[hldict[i]['constraint']].append(score)
conddict = make_conddict(hldict)
n4report = sim_fk_N400(conddict,fklog,setting,k=k,bert=bert)
tot_acc = get_acc(tot_score)
report = '\nPrediction accuracies:\n'
report += 'EXP TGT in TOP %s preds: %s (%s/%s)\n'%(k,tot_acc,sum(tot_score),len(tot_score))
report += 'in TOP %s for H: %s\n'%(k,get_acc(by_constraint_score['H']))
report += 'in TOP %s for L: %s\n'%(k,get_acc(by_constraint_score['L']))
return report,n4report,correct,tot_acc
def sim_fk_N400(conddict,logfile,setting,k=5,bert=True):
thresh = 0.01
exp_top = {'H':[],'L':[]}
exp_top_thresh = {'H':[],'L':[]}
wc_boost = {'H':[],'L':[]}
allprobs = {'H':[],'L':[]}
for it in conddict:
exp_prob,wc_prob,bc_prob = [conddict[it][cont]['tgtprob'] for cont in ['exp','wc','bc']]
logfile.write(conddict[it]['exp']['sent'][setting])
logfile.write(' ' + '/'.join([conddict[it][cont]['tgt'] for cont in ['exp','wc','bc']]) + '\n')
logfile.write('TGT probs: %s\n'%[exp_prob,wc_prob,bc_prob])
logfile.write('PREDICTED: %s\n'%conddict[it]['exp']['toppreds'])
logfile.write(str(conddict[it]['exp']['topprobs']) + '\n')
cons = conddict[it]['exp']['constraint']
logfile.write(cons + '\n')
allprobs[cons].append((exp_prob,wc_prob,bc_prob))
if (exp_prob > wc_prob) and (exp_prob > bc_prob):
exp_top[cons].append(1)
logfile.write('EXP TOP\n')
if abs(exp_prob - wc_prob) > thresh and abs(exp_prob - bc_prob) > thresh:
exp_top_thresh[cons].append(1)
else:
exp_top_thresh[cons].append(0)
else:
exp_top[cons].append(0)
exp_top_thresh[cons].append(0)
if (wc_prob > bc_prob) and abs(wc_prob - bc_prob) > thresh:
wc_boost[cons].append(1)
logfile.write('WC BOOST\n')
else:
wc_boost[cons].append(0)
logfile.write('----\n\n\n')
report = '\nTGT probability patterning:\n'
report += 'AVG PROB HIGH EXP: %s\n'%np.average([e[0] for e in allprobs['H']])
report += 'AVG PROB HIGH WC: %s\n'%np.average([e[1] for e in allprobs['H']])
report += 'AVG PROB HIGH BC: %s\n'%np.average([e[2] for e in allprobs['H']])
report += 'AVG PROB LOW EXP: %s\n'%np.average([e[0] for e in allprobs['L']])
report += 'AVG PROB LOW WC: %s\n'%np.average([e[1] for e in allprobs['L']])
report += 'AVG PROB LOW BC: %s\n'%np.average([e[2] for e in allprobs['L']])
report += 'EXP TOP: %s\n'%get_acc(exp_top['H']+exp_top['L'])
report += 'EXP TOP HIGH: %s\n'%get_acc(exp_top['H'])
report += 'EXP TOP LOW: %s\n'%get_acc(exp_top['L'])
report += 'EXP TOP w/ THRESH %s: %s\n'%(thresh,get_acc(exp_top_thresh['H']+exp_top_thresh['L']))
report += 'EXP TOP HIGH w/ THRESH %s: %s\n'%(thresh,get_acc(exp_top_thresh['H']))
report += 'EXP TOP LOW w/ THRESH %s: %s\n'%(thresh,get_acc(exp_top_thresh['L']))
report += 'WC BOOST HIGH: %s\n'%get_acc(wc_boost['H'])
report += 'WC BOOST LOW: %s\n'%get_acc(wc_boost['L'])
return report
def test_rr_acc(clozedict,inputlist,tgtlist,clozelist,model,tokenizer,rrlog,k=5,bert=True,scat=None):
tok_preds,top_probs = tp.get_predictions(inputlist,model,tokenizer,k=k,bert=bert)
tok_probs = tp.get_probabilities(inputlist,tgtlist,model,tokenizer,bert=bert)
tot_score = []
correct = []
correct_by_quartile = {'q1_corr':[],'q2_corr':[],'q3_corr':[],'q4_corr':[]}
avgcloze = np.average(clozelist)
q1 = np.percentile(clozelist,25)
q3 = np.percentile(clozelist,75)
q2 = np.percentile(clozelist,50)
q4 = max(clozelist)
predcounts = Counter()
for i,pred in enumerate(tok_preds):
clozedict[i]['toppreds'] = pred
clozedict[i]['tgtprob'] = tok_probs[i]
clozedict[i]['topprobs'] = top_probs[i]
score = 0
itm = clozedict[i]['item']
cond = clozedict[i]['cond']
for candidate in pred:
if candidate.strip() in [e.split()[0] for e in clozedict[i]['exp']]:
score = 1
if score == 1:
ctup = (clozedict[i]['sent'],clozedict[i]['exp'],pred,clozedict[i]['maxcloze'])
if ctup not in correct:
correct.append(ctup)
tot_score.append(score)
if clozedict[i]['maxcloze'] <= q1:
correct_by_quartile['q1_corr'].append(score)
elif clozedict[i]['maxcloze'] <= q2:
correct_by_quartile['q2_corr'].append(score)
elif clozedict[i]['maxcloze'] <= q3:
correct_by_quartile['q3_corr'].append(score)
else:
correct_by_quartile['q4_corr'].append(score)
conddict = make_conddict(clozedict)
n4report = sim_rr_N400(conddict,rrlog,scat=scat,k=k,bert=bert)
report = '\nPrediction accuracies:\n'
report += 'TGT in top %s preds: %s (%s/%s)\n'%(k,get_acc(tot_score),sum(tot_score),len(tot_score))
report += 'TGT in top %s for Q1: %s (%s upper, %s items)\n'%(k,get_acc(correct_by_quartile['q1_corr']),q1,len(correct_by_quartile['q1_corr']))
report += 'TGT in top %s for Q2: %s (%s upper, %s items)\n'%(k,get_acc(correct_by_quartile['q2_corr']),q2,len(correct_by_quartile['q2_corr']))
report += 'TGT in top %s for Q3: %s (%s upper, %s items)\n'%(k,get_acc(correct_by_quartile['q3_corr']),q3,len(correct_by_quartile['q3_corr']))
report += 'TGT in top %s for Q4: %s (%s upper, %s items)\n'%(k,get_acc(correct_by_quartile['q4_corr']),q4,len(correct_by_quartile['q4_corr']))
report += 'AVG CLOZE: %s\n'%avgcloze
report += 'MED CLOZE: %s\n'%q2
return report,n4report,correct,predcounts
def sim_rr_N400(conddict,logfile,scat=None,k=5,bert=True):
thresh = .01
pattern = []
pattern_thresh = []
same = []
probpairs = []
clozepairs = []
for it in conddict:
a_prob,b_prob = (conddict[it]['a']['tgtprob'],conddict[it]['b']['tgtprob'])
logfile.write(conddict[it]['a']['sent'] + ' ' + conddict[it]['a']['tgt'] + '\n')
logfile.write(conddict[it]['b']['sent'] + ' ' + conddict[it]['b']['tgt'] + '\n')
logfile.write('TGT probs: %s\n'%[a_prob,b_prob])
logfile.write('TGT cloze: %s\n'%[conddict[it]['a']['tgtcloze'],conddict[it]['b']['tgtcloze']])
logfile.write('PREDICTED: %s'%conddict[it]['a']['toppreds'] + '\n')
logfile.write(str(conddict[it]['a']['topprobs']) + '\n')
logfile.write('PREDICTED: %s'%conddict[it]['b']['toppreds'] + '\n')
logfile.write(str(conddict[it]['b']['topprobs']) + '\n')
if (a_prob > b_prob):
pattern.append(1)
logfile.write('PATTERN\n')
else:
pattern.append(0)
if (a_prob > b_prob) and (abs(a_prob - b_prob) > thresh):
pattern_thresh.append(1)
logfile.write('PATTERN THRESH\n')
else:
pattern_thresh.append(0)
if (abs(a_prob - b_prob) < thresh):
same.append(1)
logfile.write('NO DIFF\n')
else:
same.append(0)
logfile.write('----\n\n\n')
probpairs.append((a_prob,b_prob))
clozepairs.append((conddict[it]['a']['tgtcloze'],conddict[it]['b']['tgtcloze']))
probdiffs = [e[0] - e[1] for e in probpairs]
clozediffs = [e[0] - e[1] for e in clozepairs]
report = '\nTarget probs vs cloze:\n'
if sum(probdiffs) > 0:
report += 'PEARSON: %s\n'%scipy.stats.pearsonr(probdiffs,clozediffs)[0]
report += 'SPEARMAN: %s\n'%scipy.stats.spearmanr(probdiffs,clozediffs)[0]
else:
report += 'PEARSON: ---\n'
report += 'SPEARMAN: ---\n'
report += 'GOOD TGT HIGHER: %s\n'%get_acc(pattern)
report += 'GOOD TGT HIGHER BY %s: %s\n'%(thresh,get_acc(pattern_thresh))
report += 'DIFF BELOW THRESH %s: %s\n'%(thresh,get_acc(same))
report += 'AVG PROB DIFF: %s\n'%np.average(probdiffs)
report += 'AVG CLOZE DIFF: %s\n'%np.average(clozediffs)
if scat:
plot_rr_prcl(clozediffs,probdiffs,scat)
return report
def plot_rr_prcl(clozediffs,probdiffs,scat):
plt.scatter(clozediffs,probdiffs)
plt.ylabel('Model probability differences',fontsize='x-large')
plt.yticks(fontsize='large')
plt.xlabel('Cloze differences',fontsize='x-large')
plt.xticks(fontsize='large')
plt.ylim(-.2,.7)
plt.savefig(scat)
plt.clf()
def test_nkf_acc(nkfdict,inputlist,tgtlist,model,tokenizer,nkflog,k=5,bert=True):
correct = []
tot_score = []
tok_preds,top_probs = tp.get_predictions(inputlist,model,tokenizer,k=k,bert=bert)
tok_probs = tp.get_probabilities(inputlist,tgtlist,model,tokenizer,bert=bert)
for i,pred in enumerate(tok_preds):
nkfdict[i]['toppreds'] = pred
nkfdict[i]['tgtprob'] = tok_probs[i]
nkfdict[i]['topprobs'] = top_probs[i]
score = 0
for candidate in pred:
if candidate.strip() in nkfdict[i]['exp']:
score = 1
if score == 1:
ctup = (nkfdict[i]['sent'],nkfdict[i]['exp'],pred,nkfdict[i]['cond'])
if ctup not in correct:
correct.append(ctup)
if nkfdict[i]['cond'] == 'TA':
tot_score.append(score)
conddict = make_conddict(nkfdict)
n4report = sim_nkf_N400(conddict,nkflog,k=k,bert=bert)
report = "\nPrediction 'accuracy':\n"
report += 'TRUE TGT in top %s preds: %s (%s/%s)\n'%(k,get_acc(tot_score),sum(tot_score),len(tot_score))
return report,n4report,correct
def sim_nkf_N400(conddict,logfile,k=5,bert=True):
pattern = []
same = []
preftrue = {'aff':[],'neg':[]}
preftrue_l = {'aff':[],'neg':[]}
preftrue_u = {'aff':[],'neg':[]}
preftrue_thresh = {'aff':[],'neg':[]}
lic = None
thresh = .01
for it in conddict:
if 'licensing' in conddict[it]['TA']:
lic = conddict[it]['TA']['licensing']
for true_cond,false_cond,pol in [('TA','FA','aff'),('TN','FN','neg')]:
true_prob,false_prob = (conddict[it][true_cond]['tgtprob'],conddict[it][false_cond]['tgtprob'])
logfile.write(str(conddict[it][true_cond]['sent'] + ' ' + conddict[it][true_cond]['tgt']) + '\n')
logfile.write(str(conddict[it][false_cond]['sent'] + ' ' + conddict[it][false_cond]['tgt']) + '\n')
logfile.write(u'TGT probs: %s\n'%[true_prob,false_prob])
logfile.write(u'PREDICTED: %s'%conddict[it][true_cond]['toppreds'] + '\n')
logfile.write(str(conddict[it][true_cond]['topprobs']) + '\n')
logfile.write(u'PREDICTED: %s'%conddict[it][false_cond]['toppreds'] + '\n')
logfile.write(str(conddict[it][false_cond]['topprobs']) + '\n')
logfile.write(u'---\n\n\n')
if true_prob > false_prob:
score = 1
else:
score = 0
preftrue[pol].append(score)
if lic:
if lic == 'Y':
preftrue_l[pol].append(score)
elif lic == 'N':
preftrue_u[pol].append(score)
else:
print('LICENSING ERROR')
if (true_prob > false_prob) and (abs(true_prob - false_prob) > thresh):
preftrue_thresh[pol].append(1)
else:
preftrue_thresh[pol].append(0)
report = '\nPreference for true vs false sentences:\n'
report += 'PREF TRUE: %s\n'%get_acc(preftrue['aff'] + preftrue['neg'])
report += 'AFF: %s\n'%get_acc(preftrue['aff'])
report += 'NEG: %s\n'%get_acc(preftrue['neg'])
report += 'PREF TRUE AFF THRESH %s: %s (%s/%s)\n'%(thresh,get_acc(preftrue_thresh['aff']),sum(preftrue_thresh['aff']),len(preftrue_thresh['aff']))
report += 'PREF TRUE NEG THRESH %s: %s (%s/%s)\n'%(thresh,get_acc(preftrue_thresh['neg']),sum(preftrue_thresh['neg']),len(preftrue_thresh['neg']))
if lic:
report += 'PREF TRUE LICENSED: %s\n'%get_acc(preftrue_l['aff'] + preftrue_l['neg'])
report += 'AFF: %s\n'%get_acc(preftrue_l['aff'])
report += 'NEG: %s\n'%get_acc(preftrue_l['neg'])
report += 'PREF TRUE UNLICENSED: %s\n'%get_acc(preftrue_u['aff'] + preftrue_u['neg'])
report += 'AFF: %s\n'%get_acc(preftrue_u['aff'])
report += 'NEG: %s\n'%get_acc(preftrue_u['neg'])
report += '\n\n'
return report
def run_fk_all(args,out,models,logcode,klist,hldict,inputlist,tgtlist,bert=True):
outstring = ''
outstring += '\n\n***\nSETTING: %s\n***\n\n'%logcode
print(logcode)
acclist = []
acclist_names = []
for modelname,model,tokenizer in models:
outstring += '\n\n***\nMODEL: %s\n***\n'%modelname
print(modelname)
reports = []
for k in klist:
with open(os.path.join(args.resultsdir,'FK-%s_predlog_%s-%s'%(logcode,modelname,k)),'w') as fklog:
report,n4report,corr,acc = test_fk_acc(hldict,inputlist,tgtlist,model,tokenizer,logcode,fklog,k=k,bert=bert)
acclist.append(acc)
acclist_names.append(modelname + '-%s'%k)
for crritem in corr:
fklog.write(str(crritem) + '\n')
reports.append((report,n4report,k))
for acc,n4,k in reports:
outstring +='\nFED/KUT k=%s acc\n'%k
outstring += acc
outstring +='\nFED/KUT N400\n'
outstring += n4
outstring +='\n----\n\n'
return acclist,acclist_names,outstring
def run_rr_all(args,out,models,logcode,klist,clozedict,inputlist,tgtlist,clozelist,bert=True):
out.write('\n\n***\nSETTING: %s\n***\n\n'%logcode)
for modelname,model,tokenizer in models:
out.write('\n\n***\nMODEL: %s\n***\n'%modelname)
reports = []
for k in klist:
with open(os.path.join(args.resultsdir,'RR-%s_predlog_%s-%s'%(logcode,modelname,k)),'w') as rrlog:
print('CHOW k=%s'%k)
report,n4report,corr,prcounts = test_rr_acc(clozedict,inputlist,tgtlist,clozelist,model,tokenizer,rrlog,k=k,bert=bert)
for crritem in corr:
rrlog.write(str(crritem) + '\n')
rrlog.write('\n'+ str(prcounts))
reports.append((report,n4report,k))
for acc,n4,k in reports:
out.write('\nCHOW k=%s acc\n'%k)
out.write(acc)
out.write('\nCHOW N400\n')
out.write(n4)
out.write('\n----\n\n')
return report,n4report
def run_neg_all(args,out,models,klist,inputlist,negdict,tgtlist,dataname,logcode,bert=True):
for modelname,model,tokenizer in models:
out.write('\n\n***\nMODEL: %s\n***\n'%modelname)
print(modelname)
reports = []
for k in klist:
with open(args.resultsdir+'/%s_predlog_%s-%s'%(logcode,modelname,k),'w') as nkflog:
report,n4report,corr = test_nkf_acc(negdict,inputlist,tgtlist,model,tokenizer,nkflog,k=k,bert=bert)
for crritem in corr:
nkflog.write(str(crritem) + '\n')
reports.append((report,n4report,k))
for acc,n4,k in reports:
out.write('\n%s k=%s acc\n'%(dataname,k))
out.write(acc)
out.write('\n%s N400\n'%dataname)
out.write(n4)
out.write('\n----\n\n')
#runs all three datasets WITH all additional perturbations tried in the paper
def run_aux_tests(args,models,klist,bert=True):
acclists_shuf = []
acclists_shufnw = []
acclists = []
with open(args.resultsdir+'/results-cprag.txt','w') as out:
hldict,inputlist,_,inputlist_nw,_,tgtlist = process_fk(args.cprag_stim)
_,_,outstring = run_fk_all(args,out,models,'orig',klist,hldict,inputlist,tgtlist,bert=bert)
out.write(outstring)
_,_,outstring = run_fk_all(args,out,models,'trunc',klist,hldict,inputlist_nw,tgtlist,bert=bert)
out.write(outstring)
for i in range(3):
_,_,inputlist_shuf,_,inputlist_shufnw,_ = process_fk(args.cprag_stim)
acclist,acclist_names_shuf,_ = run_fk_all(args,out,models,'shuf',klist,hldict,inputlist_shuf,tgtlist,bert=bert)
acclists_shuf.append(acclist)
acclist,acclist_names_shufnw,_ = run_fk_all(args,out,models,'shuftrunc',klist,hldict,inputlist_shufnw,tgtlist,bert=bert)
acclists_shufnw.append(acclist)
out.write('\n\nSHUF ACCLISTS')
out.write(str(acclist_names_shuf) + '\n')
out.write(str(acclists_shuf))
out.write('\n\nSHUF-TRUNC ACCLISTS')
out.write(str(acclist_names_shufnw) + '\n')
out.write(str(acclists_shufnw))
out.write('\n\nSHUF ACCURACIES\n')
accs_by_modk = list(zip(*acclists_shuf))
i = 0
for modelname,_,_ in models:
for k in klist:
this_accs = accs_by_modk[i]
out.write('%s k=%s: %s pm %s\n'%(modelname,k,np.average(this_accs),np.std(this_accs)))
i += 1
out.write('\n\nSHUF-TRUNC ACCURACIES\n')
accs_by_modk = list(zip(*acclists_shufnw))
i = 0
for modelname,_,_ in models:
for k in klist:
this_accs = accs_by_modk[i]
out.write('%s k=%s: %s pm %s\n'%(modelname,k,np.average(this_accs),np.std(this_accs)))
i += 1
with open(args.resultsdir+'/results-role.txt','w') as out:
clozedict,inputlist,tgtlist,clozelist = process_rr(args.role_stim,gen_obj=False,gen_subj=False)
run_rr_all(args,out,models,'orig',klist,clozedict,inputlist,tgtlist,clozelist,bert=bert)
clozedict,inputlist,tgtlist,clozelist = process_rr(args.role_stim,gen_obj=True,gen_subj=False)
run_rr_all(args,out,models,'-obj',klist,clozedict,inputlist,tgtlist,clozelist,bert=bert)
clozedict,inputlist,tgtlist,clozelist = process_rr(args.role_stim,gen_obj=False,gen_subj=True)
run_rr_all(args,out,models,'-subj',klist,clozedict,inputlist,tgtlist,clozelist,bert=bert)
clozedict,inputlist,tgtlist,clozelist = process_rr(args.role_stim,gen_obj=True,gen_subj=True)
run_rr_all(args,out,models,'-obsub',klist,clozedict,inputlist,tgtlist,clozelist,bert=bert)
with open(args.resultsdir+'/results-neg.txt','w') as out:
inputlist,negdict,tgtlist = process_fischler(args.negsimp_stim)
run_neg_all(args,out,models,klist,inputlist,negdict,tgtlist,'FISCHLER','FS',bert=bert)
inputlist,negdict,tgtlist = process_nk(args.negnat_stim)
run_neg_all(args,out,models,klist,inputlist,negdict,tgtlist,'NIEUWLAND','NK',bert=bert)
#runs all three datasets without any perturbations from paper
def run_three_orig(args,models,klist,bert=True):
with open(args.resultsdir+'/results-neg.txt','w') as out:
inputlist,negdict,tgtlist = process_fischler(args.negsimp_stim)
run_neg_all(args,out,models,klist,inputlist,negdict,tgtlist,'FISCHLER','FS',bert=bert)
inputlist,negdict,tgtlist = process_nk(args.negnat_stim)
run_neg_all(args,out,models,klist,inputlist,negdict,tgtlist,'NIEUWLAND','NK',bert=bert)
with open(args.resultsdir+'/results-role.txt','w') as out:
clozedict,inputlist,tgtlist,clozelist = process_rr(args.role_stim,gen_obj=False,gen_subj=False)
run_rr_all(args,out,models,'orig',klist,clozedict,inputlist,tgtlist,clozelist,bert=bert)
with open(args.resultsdir+'/results-cprag.txt','w') as out:
hldict,inputlist,_,_,_,tgtlist = process_fk(args.cprag_stim)
_,_,outstring = run_fk_all(args,out,models,'orig',klist,hldict,inputlist,tgtlist,bert=bert)
out.write(outstring)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--cprag_stim", default=None, type=str)
parser.add_argument("--role_stim", default=None, type=str)
parser.add_argument("--negsimp_stim", default=None, type=str)
parser.add_argument("--negnat_stim", default=None, type=str)
parser.add_argument("--resultsdir",default=None, type=str)
parser.add_argument("--bertbase",default=None, type=str)
parser.add_argument("--bertlarge",default=None, type=str)
parser.add_argument("--incl_perturb", action="store_true")
args = parser.parse_args()
print('LOADING MODELS')
bert_base,tokenizer_base = tp.load_model(args.bertbase)
bert_large,tokenizer_large = tp.load_model(args.bertlarge)
klist = [1,5]
models = [('bert-base-uncased',bert_base,tokenizer_base),('bert-large-uncased',bert_large,tokenizer_large)]
print('RUNNING EXPERIMENTS')
if args.incl_perturb:
run_aux_tests(args,models,klist,bert=True)
else:
run_three_orig(args,models,klist,bert=True)