-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathscripts.py
485 lines (459 loc) · 23 KB
/
scripts.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
import os
import argparse
import numpy
import tokenizers
import transformers
import huggingface_hub
import qarac.corpora.BNCorpus
import qarac.models.QaracTrainerModel
import qarac.corpora.CombinedCorpus
import torch
import spacy
import pandas
import qarac.utils.CoreferenceResolver
import nltk.corpus
import difflib
import scipy.stats
import scipy.spatial
import seaborn
import tqdm
import gradio
import boto3
from google.colab import userdata
class SequenceCrossEntropyLoss(torch.nn.Module):
def __init__(self):
super(SequenceCrossEntropyLoss,self).__init__()
self.crossentropy = torch.nn.CrossEntropyLoss()
def forward(self,y_pred,y_true):
(batch_size,sequence_length,n_classes) = y_pred.shape
predictions = y_pred.view(-1,n_classes)
labels = y_true.view(-1)
return self.crossentropy(predictions,labels)
class CombinedLoss(torch.nn.Module):
def __init__(self):
super(CombinedLoss,self).__init__()
self.component_losses = (SequenceCrossEntropyLoss(),
torch.nn.MSELoss(),
SequenceCrossEntropyLoss(),
torch.nn.MSELoss())
def forward(self,y_pred,y_true):
return sum((fn(pred,obs)
for (fn,pred,obs) in zip(self.component_losses,
y_pred,
y_true)))
def capitalise(token,i):
return token.text_with_ws.title() if i==0 or token.tag_.startswith('NNP') else token.text_with_ws.lower()
def clean_question(doc):
words = [capitalise(token,i) for (i,token) in enumerate(doc)]
if words[-1]!='?':
words.append('?')
return ''.join(words)
def download_training_data():
if not os.path.exists('corpora'):
os.makedirs('corpora')
s3 = boto3.client('s3',
aws_access_key_id=userdata.get('AWS_KEY'),
aws_secret_access_key=userdata.get('AWS_SECRET'))
for obj in s3.list_objects(Bucket='qarac')['Contents']:
filename = obj['Key']
s3.download_file('qarac',filename,'corpora/{}'.format(filename))
def prepare_wiki_qa(filename,outfilename):
data = pandas.read_csv(filename,sep='\t')
data['QNum']=data['QuestionID'].apply(lambda x: int(x[1:]))
nlp = spacy.load('en_core_web_trf')
predictor = qarac.utils.CoreferenceResolver.CoreferenceResolver()
data['Resolved_answer'] = data.groupby('QNum')['Sentence'].transform(predictor)
unique_questions = data.groupby('QNum')['Question'].first()
cleaned_questions = pandas.Series([clean_question(doc)
for doc in nlp.pipe(unique_questions)],
index = unique_questions.index)
for (i,question) in cleaned_questions.items():
data.loc[data['QNum']==i,'Cleaned_question']=question
data[['Cleaned_question','Resolved_answer','Label']].to_csv(outfilename)
def prepare_training_datasets():
wikiqa = pandas.read_csv('corpora/WikiQA.csv')
avicenna = pandas.read_csv('corpora/Avicenna_Train.csv',encoding='iso-8859-1')
snli = pandas.read_csv('corpora/snli_1.0_train.csv')
question_answering = wikiqa.loc[wikiqa['Label']==1,
['Cleaned_question',
'Resolved_answer']].rename(columns={'Cleaned_question':'question',
'Resolved_answer':'answer'})
reasoning = avicenna.loc[avicenna['Syllogistic relation']=='yes',
['Premise 1',
'Premise 2',
'Conclusion']].rename(columns={'Premise 1':'proposition0',
'Premise 2':'proposition1',
'Conclusion':'conclusion'})
consistency = snli.loc[snli['gold_label']!='-',
['sentence1',
'sentence2']].rename(columns={'sentence1':'statement0',
'sentence2':'statement1'})
mapping = {'entailment':1.0,
'neutral':0.0,
'contradiction':-1.0}
consistency['consistency'] = snli.loc[snli['gold_label']!='-',
'gold_label'].apply(lambda x:mapping[x])
all_text = pandas.concat([wikiqa['Resolved_answer'],
avicenna['Premise 1'],
avicenna['Premise 1'],
reasoning['conclusion'],
snli['sentence1'],
snli['sentence2']]).to_frame(name='all_text').reset_index(drop=True)
all_text.to_csv('corpora/all_text.csv')
question_answering.to_csv('corpora/question_answering.csv')
reasoning.to_csv('corpora/reasoning_train.csv')
consistency.to_csv('corpora/consistency.csv')
def train_models(path,progress=gradio.Progress(track_tqdm=True)):
device = torch.device('cuda:0')
torch.cuda.empty_cache()
tokenizer = tokenizers.Tokenizer.from_pretrained('roberta-base').to(device)
trainer = qarac.models.QaracTrainerModel.QaracTrainerModel('roberta-base',
tokenizer)
trainer.to(device)
loss_fn = CombinedLoss()
loss_fn.cuda()
optimizer = torch.optim.NAdam(trainer.parameters(),lr=5.0e-5)
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer,gamma=0.9)
training_data = qarac.corpora.CombinedCorpus.CombinedCorpus(tokenizer,
all_text='corpora/all_text.csv',
question_answering='corpora/question_answering.csv',
reasoning='corpora/reasoning_train.csv',
consistency='corpora/consistency.csv',
device=device)
n_batches = len(training_data)
history = {}
for epoch in range(25):
print("Epoch",epoch)
epoch_label = 'Epoch {}'.format(epoch)
epoch_data = {}
for (batch,(X,Y)) in enumerate(tqdm.tqdm(training_data)):
prediction = trainer(X['all_text'],
X['offset_text'],
X['question'],
X['answer'],
X['proposition0'],
X['proposition1'],
X['conclusion_offset'],
X['statement0'],
X['statement1'])
loss = loss_fn(prediction,Y)
loss.backward()
optimizer.step()
optimizer.zero_grad()
if batch % 1024 == 0 or batch == n_batches-1:
epoch_data[batch] = loss.item()
history[epoch_label] = epoch_data
scheduler.step()
huggingface_hub.login(token=userdata.get('HUGGINGFACE_TOKEN'))
trainer.question_encoder.push_to_hub('{}/qarac-roberta-question-encoder'.format(path))
trainer.answer_encoder.push_to_hub('{}/qarac-roberta-answer-encoder'.format(path))
trainer.decoder.push_to_hub('{}/qarac-roberta-decoder'.format(path))
return history
def test_encode_decode(path):
encoder = transformers.Transformer.from_pretrained('{}/qarac-roberta-answer-encoder'.format(path))
decoder = transformers.Transformer.from_pretrained('{}/qarac-robeerta-decoder'.format(path))
tokenizer=tokenizers.Tokenizer.from_pretrained('roberta-base')
exclude = tokenizer.encode('<s> </s> <pad>').ids
analyser = difflib.SequenceMatcher(lambda x: x in exclude)
bnc = nltk.corpus.reader.bnc.BNCCorpusReader('/'.join([os.environ['HOME'],
'BNC',
'Texts']),
fileids=r'[A-K]/\w*/\w*\.xml')
matches = []
batch = []
pad_token = tokenizer.token_to_id('<pad>')
for sent in bnc.sents(strip_space=False):
batch.append(tokenizer.encode(''.join(sent)))
if len(batch)==32:
maxlen = max((len(sentence) for sentence in batch))
for sample in batch:
sample.pad(maxlen,pad_id=pad_token)
input_ids = torch.tensor([sample.ids for sample in batch])
attention_mask = torch.not_equal(input_ids,pad_token)
vectors = encoder(input_ids,
attention_mask)
decoded = decoder.generate(vector=vectors)
for (s1,s2) in zip(batch,decoded):
analyser.set_seqs(s1.ids, s2)
matches.append(analyser.ratio())
batch = []
if len(batch)!=0:
maxlen = max((len(sentence) for sentence in batch))
for sample in batch:
sample.pad(maxlen,pad_id=pad_token)
input_ids = torch.tensor([sample.ids for sample in batch])
attention_mask = torch.not_equal(input_ids, pad_token)
vectors = encoder(input_ids,
attention_mask)
decoded = decoder.generate(vector=vectors)
for (s1,s2) in zip(batch,decoded):
analyser.set_seqs(s1.ids, s2)
matches.append(analyser.ratio())
matches = numpy.array(matches)
print("Accuracy: mean = {0}, sd = {1}".format(matches.mean(),
matches.sd()))
(alpha,beta,loc,scale)=scipy.stats.beta.fit(matches,floc=0.0,fscale=1.0)
print("Beta distribution parameters alpha = {0}, beta = {1}".format(alpha,beta))
(hist,bins) = numpy.histogram(matches,bins='fd')
with pandas.option_context('plotting.backend','matploblib.backends.backend_svg') as options:
axes = pandas.Series(hist,index=(bins[1:]+bins[:-1]/2)).plot.bar()
axes.get_figure().savefig('encode_decode_histogram.svg')
percent = numpy.linspace(0.0,1.0,101)
percentiles = numpy.quantile(matches,percent)
with pandas.option_context('plotting.backend','matplotlib.backends.backend_svg') as options:
axes = pandas.Series(percentiles, index=percent).plot.bar()
axes.get_figure().savefig('encode_decode_percentile.svg')
def test_question_answering(path):
question_encoder = transformers.Transformer.from_pretrained('{}/qarac-roberta-question-encoder'.format(path))
answer_encoder = transformers.Transformer.from_pretrained('{}/qarac-roberta-answer-encoder'.format(path))
tokenizer = tokenizers.Tokenizer.from_pretrained('roberta-base')
data = pandas.read_csv('WikiQA.tsv',sep='\t')
data['QNum']=data['QuestionID'].apply(lambda x: int(x[1:]))
nlp = spacy.load('en_core_web_trf')
predictor = qarac.utils.CoreferenceResolver.CoreferenceResolver()
data['Resolved_answer'] = data.groupby('QNum')['Sentence'].transform(predictor)
unique_questions = data.groupby('QNum')['Question'].first()
cleaned_questions = pandas.Series([clean_question(doc)
for doc in nlp.pipe(unique_questions)],
index = unique_questions.index)
def tokenize(column):
return tokenizer.encode_batch(column.apply(lambda x:tokenizers.TextInputSequence(x)),
add_special_tokens=False)
questions = tokenize(cleaned_questions)
maxlen=max((len(question) for question in questions))
pad_token = tokenizer.token_to_id('<pad>')
for question in questions:
question.pad(maxlen,pad_id=pad_token)
question_ids = torch.tensor([question.ids
for question in questions])
attention_mask = torch.not_equal(question_ids,
pad_token)
q_vectors = question_encoder(question_ids,
attention_mask=attention_mask).numpy()
answers = tokenize(data['Resolved_answer'])
maxlen = max((len(answer) for answer in answers))
for answer in answers:
answer.pad(maxlen,pad_id=pad_token)
answer_ids = torch.tensor([answer.ids
for answer in answers])
attention_mask = torch.not_equal(answer_ids,
pad_token)
answer_lookup = scipy.spatial.KDTree(answer_encoder(answer_ids,
attention_mask=attention_mask).numpy())
n_correct = 0
all_distances = 0.0
correct_distances = 0.0
wrong_distances = 0.0
all_sq = 0.0
correct_sq = 0.0
wrong_sq = 0.0
for (i,qv) in enumerate(q_vectors):
(d,row) = answer_lookup.query(qv)
dsq=d**2.0
correct = (row['QNum']==i and row['Label']==1)
all_distances+=d
all_sq+=dsq
if correct:
n_correct+=1
correct_distances+=d
correct_sq+=dsq
else:
wrong_distances+=d
wrong_sq+=dsq
N = cleaned_questions.shape[0]
print("{0} questions, {1} possible answers, {2} correct answers".format(N,
data.shape[0],
n_correct))
accuracy = n_correct/N
baseline = N/data.shape[0]
kappa = 1.0 - ((1.0-accuracy)/(1.0-baseline))
print(("Accuracy: {0}, Baseline {1}, kappa{2} ".format(accuracy,baseline,kappa)))
mean_dist =all_distances/N
mean_sq = all_sq/N
all_sd = numpy.sqrt(mean_sq-(mean_dist**2.0))
print("Question-answer distances")
print("All: mean {0}, sd {1}".format(mean_dist,all_sd))
correct_mean = correct_distances/n_correct
correct_meansq = correct_sq/n_correct
correct_sd = numpy.sqrt(correct_meansq - (correct_mean**2.0))
print("Correct: mean {0}, sd {1}".format(correct_mean,correct_sd))
wrong_mean = wrong_distances/(N-n_correct)
wrong_meansq = wrong_sq/(N-n_correct)
wrong_sd = numpy.sqrt(wrong_meansq - (wrong_mean**2.0))
print("Wrong: mean {0}, sd {1}".format(wrong_mean,wrong_sd))
def test_reasoning(path):
encoder = transformers.Transformer.from_pretrained('{}/qarac-roberta-answer-encoder'.format(path))
decoder = transformers.Transformer.from_pretrained('{}/qarac-robeerta-decoder'.format(path))
tokenizer=tokenizers.Tokenizer.from_pretrained('roberta-base')
exclude = tokenizer.encode('<s> </s> <pad>').ids
analyser = difflib.SequenceMatcher(lambda x: x in exclude)
data = pandas.read_csv('corpora/Avicenna_Test.csv',encoding='iso-8859-1')
data = data.loc[data['Syllogistic relation']=='yes']
def tokenize(column):
return tokenizer.encode_batch(column.apply(lambda x:tokenizers.TextInputSequence(x)),
add_special_tokens=False)
p0 = tokenize(data['Premise 1'])
p1 = tokenize(data['Premise 2'])
c = tokenize(data['Conclusion'])
p0_batch = []
p1_batch = []
c_batch = []
n=0
pad_token = tokenizer.token_to_id('<pad>')
matches=[]
for (p0_sample,p1_sample,c_sample) in zip(p0,p1,c):
p0_batch.append(p0_sample)
p1_batch.append(p1_sample)
c_batch.append(c_sample)
n+=1
if n==32:
maxlen=max((len(sample for sample in p0_batch)))
for sample in p0_batch:
sample.pad(maxlen,pad_token)
p0_in = torch.tensor([sample.ids for sample in p0.batch])
p0_attn = torch.not_equal(p0_in,
pad_token)
maxlen=max((len(sample for sample in p1_batch)))
for sample in p1_batch:
sample.pad(maxlen,pad_token)
p1_in = torch.tensor([sample.ids for sample in p1.batch])
p1_attn = torch.not_equal(p0_in,
pad_token)
predictions = decoder.generate(vector=(encoder(p0_in,
attention_mask=p0_attn)
+encoder(p1_in,
attention_mask=p1_attn)))
for (s1,s2) in zip(c_batch,predictions):
analyser.set_seqs(s1.ids, s2)
matches.append(analyser.ratio())
n=0
p0_batch=[]
p1_batch=[]
c_batch=[]
if n!=0:
maxlen=max((len(sample for sample in p0_batch)))
for sample in p0_batch:
sample.pad(maxlen,pad_token)
p0_in = torch.tensor([sample.ids for sample in p0.batch])
p0_attn = torch.not_equal(p0_in,
pad_token)
maxlen=max((len(sample for sample in p1_batch)))
for sample in p1_batch:
sample.pad(maxlen,pad_token)
p1_in = torch.tensor([sample.ids for sample in p1.batch])
p1_attn = torch.not_equal(p0_in,
pad_token)
predictions = decoder.generate(vector=(encoder(p0_in,
attention_mask=p0_attn)
+encoder(p1_in,
attention_mask=p1_attn)))
for (s1,s2) in zip(c_batch,predictions):
analyser.set_seqs(s1.ids, s2)
matches.append(analyser.ratio())
matches = numpy.array(matches)
print("Accuracy: mean = {0}, sd = {1}".format(matches.mean(),
matches.sd()))
(alpha,beta,loc,scale)=scipy.stats.beta.fit(matches,floc=0.0,fscale=1.0)
print("Beta distribution parameters alpha = {0}, beta = {1}".format(alpha,beta))
(hist,bins) = numpy.histogram(matches,bins='fd')
with pandas.option_context('plotting.backend','matploblib.backends.backend_svg') as options:
axes = pandas.Series(hist,index=(bins[1:]+bins[:-1]/2)).plot.bar()
axes.get_figure().savefig('reasoning_histogram.svg')
percent = numpy.linspace(0.0,1.0,101)
percentiles = numpy.quantile(matches,percent)
with pandas.option_context('plotting.backend','matplotlib.backends.backend_svg') as options:
axes = pandas.Series(percentiles, index=percent).plot.bar()
axes.get_figure().savefig('reasoning_percentile.svg')
def test_consistency(path):
encoder = transformers.Transformer.from_pretrained('{}/qarac-roberta-answer-encoder'.format(path))
tokenizer = tokenizer=tokenizers.Tokenizer.from_pretrained('roberta-base')
data = pandas.read_csv('corpora/snli_1.0_test.csv')
data = data.loc[data['gold_label']!='-']
pad_token=tokenizer.token_to_id('<pad>')
def tokenize(column):
return tokenizer.encode_batch(column.apply(lambda x:tokenizers.TextInputSequence(x)),
add_special_tokens=False)
s0 =tokenize(data['sentence1'])
s1 = tokenize(data['sentence1'])
maxlen = max((len(sentence for sentence in s0)))
for sentence in s0:
sentence.pad(maxlen,pad_id=pad_token)
s0_in = torch.tensor([sentence.ids for sentence in s0])
s0_attn = torch.not_equal(s0_in,
pad_token)
maxlen = max((len(sentence for sentence in s1)))
for sentence in s1:
sentence.pad(maxlen,pad_id=pad_token)
s1_in = torch.tensor([sentence.ids for sentence in s1])
s1_attn = torch.not_equal(s1_in,
pad_token)
s0_vec = encoder(s0_in,attention_mask=s0_attn)
s1_vec = encoder(s1_in,attention_mask=s1_attn)
cosine = torch.nn.CosineSimilarity(dim=2,eps=1.0e-12)
consistency = cosine(s0_vec,s1_vec).numpy()
results = pandas.DataFrame({'label':data['gold_label'],
'score':consistency})
third = 1.0/3.0
def predicted_labels(x):
return 'entailment' if x>third else 'contradiction' if x<-third else 'neutral'
results['prediction'] = results['score'].apply(predicted_labels)
confusion=results.groupby('label')['prediction'].value_counts().fillna(0)
seaborn.heatmap(confusion).save('consistency_confusion_matrix.svg')
correct = pandas.Series({label:confusion[label,label]
for label in confusion.index})
print("Accuracy: {}".format(correct.sum()/data.shape[0]))
print("Precision")
print(correct/confusion.sum(axis='columns'))
print("Recall")
print(correct/confusion.sum(axis='rows'))
def stats(group):
(alpha,beta,loc,scale) = scipy.stats.beta.fit(group)
mean = group.mean()
sd = group.std()
return pandas.Series({'mean':mean,
'sd':sd,
'min':loc,
'max':loc+scale,
'alpha':alpha,
'beta':beta})
print(results.groupby('label')['score'].apply(stats))
quartiles = numpy.quantile(consistency,[0.0,0.25,0.5,0.75,1.0])
IQR = quartiles[3]-quartiles[1]
bin_width = 2.0*IQR/(data.shape[0]**1.5)
n_bins = int((quartiles[4]-quartiles[0])/bin_width)
bins = numpy.linspace(quartiles[0],quartiles[4],n_bins)
def hist(col):
(result,_) = numpy.histogram(col,bins)
return result
histograms = results.groupby('label')['score'].apply(hist)
histograms.coluumns = (bins[1:]+bins[:-1])/2
with pandas.option_context('plotting.backend','matploblib.backends.backend_svg') as options:
axes=histograms.T.plot.bar(stacked=True)
axes.get_figure().savefig('consistency_histograms.svg')
percent = numpy.linspace(0.0,1.0,101)
percentiles = results.groupby('label')['score'].apply(lambda x: numpy.percentile(x,percent))
with pandas.option_context('plotting.backend','matploblib.backends.backend_svg') as options:
axes=percentiles.T.plot.line()
axes.get_figure().savefig('consistency_percentiles.svg')
if __name__ == '__main__':
parser = argparse.ArgumentParser(prog='QARAC',
description='Experimental NLP system, aimed at improving factual accuracy')
parser.add_argument('task')
parser.add_argument('-f','--filename')
parser.add_argument('-t','--training-task')
parser.add_argument('-o','--outputfile')
args = parser.parse_args()
if args.task == 'prepare_wiki_qa':
prepare_wiki_qa(args.filename,args.outputfile)
elif args.task == 'prepare_training_datasets':
prepare_training_datasets()
elif args.task == 'train_models':
train_models(args.filename)
elif args.task == 'test_encode_decode':
test_encode_decode(args.filename)
elif args.task== 'test_question_answering':
test_question_answering(args.filename)
elif args.task=="test_reasoning":
test_reasoning(args.filename)
elif args.task=='test_consistency':
test_consistency(args.filename)