-
Notifications
You must be signed in to change notification settings - Fork 2
/
lambada_score.py
311 lines (270 loc) · 10.6 KB
/
lambada_score.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
# coding=utf-8
# Copyright (c) 2021 Jeffrey M. Binder. All rights reserved.
import json
import math
import numpy as np
import nltk
import os
import re
import sys
import torch
import transformers
from transformers import AutoTokenizer, AutoModelForCausalLM
from program import Program
from generator import PromptArrayGenerator
model_type = 'gpt2'
model_name_or_path = 'gpt2'
device = 'cuda'
test_mode = 'word'
repetition_penalty = None
suppress_punctuation = True
batch_size = 20
prompting_mode = 'sentence' # One of 'default', 'blank', 'fixed', 'word', 'phrase', 'sentence', 'sentence|blank', 'sentence|word', 'sentence|phrase', 'sentence|word|phrase'
prefix = '[...]'
fixed_negative_prompt = '[...] and'
finetune_sentence_tokenizer = False
regularize_text = False
overlap_factor = 0.0
re_phrase_boundary = re.compile('[,.:;?!"“”]')
# Initialize the model and tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
model = AutoModelForCausalLM.from_pretrained(model_name_or_path)
model.to(device)
model.eval()
generator = PromptArrayGenerator(
model,
tokenizer
)
if model_type == 'xlm':
re_word = re.compile(r"^ ?[A-Za-z']+(</w>)?$")
re_final_punct = re.compile(r"^.*?([^A-Za-z' ]+(</w>)?)$")
re_first_punct_on = re.compile(r"^.*?([^@A-Za-z' ]+.*)$")
elif model_type == 'ctrl':
re_word = re.compile(r"^ ?[@A-Za-z']+$")
re_final_punct = re.compile(r"^.*s?([^@A-Za-z' ]+)$")
re_first_punct_on = re.compile(r"^.*?([^@A-Za-z' ]+.*)$")
elif model_type == 'xlnet':
re_word = re.compile(r"^ ?[A-Za-z'▁]+$")
else:
re_word = re.compile(r"^ ?[A-Za-z']+$")
if model_type == 'xlm':
def is_word_piece(idx):
tok = tokenizer.convert_ids_to_tokens([idx])[0]
return re_word.match(tok) and not tok.endswith('</w>')
elif model_type == 'ctrl':
def is_word_piece(idx):
tok = tokenizer.convert_ids_to_tokens([idx])[0]
return tok.endswith('@@')
elif model_type == 'xlnet':
def is_word_piece(idx):
tok = tokenizer.convert_ids_to_tokens([idx])[0]
return re_word.match(tok) and not tok.startswith('▁')
else:
def is_word_piece(idx):
tok = tokenizer.convert_ids_to_tokens([idx])[0]
string = tokenizer.convert_tokens_to_string([tok])
return re_word.match(string) and not string.startswith(' ')
def is_punctuation(idx):
tok = tokenizer.convert_ids_to_tokens([idx])[0]
string = tokenizer.convert_tokens_to_string([tok])
return not re_word.match(string)
punctuation = []
word_pieces = []
vocab = tokenizer.get_vocab()
vocab_size = len(vocab)
for tok in vocab:
idx = vocab[tok]
tok = tokenizer.convert_tokens_to_string([tok])
if not re_word.match(tok):
punctuation.append([idx])
if model_type in ('xlm', 'ctrl') and test_mode == 'token' and is_word_piece(idx):
word_pieces.append([idx])
bos_token = tokenizer.bos_token or tokenizer.cls_token or ''
if model_type == 'ctrl':
bos_token = 'Books '
# The models have word pieces at the beginning of the word, so we must add in an offset when
# locating word boundaries
if model_type in ('xlm', 'ctrl'):
word_piece_offset = 1
else:
word_piece_offset = 0
if model_type in ('xlm', 'ctrl') and test_mode == 'token':
# Do not allow the prediction of word pieces in token mode because they cannot come at the
# end of sentence in these models
bad_words_ids = punctuation.copy() if suppress_punctuation else []
bad_words_ids += word_pieces
elif model_type in ('openai-gpt', 'gpt2', 'xlnet') and test_mode == 'word':
# Conversely, with these models, the word pieces come at the end, so they must be suppressed
# at the beginning when we are trying to predict a word.
bad_words_ids = punctuation.copy() if suppress_punctuation else []
bad_words_ids += word_pieces
else:
bad_words_ids = punctuation if suppress_punctuation else None
fixed_negative_prompt = Program.escape(fixed_negative_prompt)
def run_model(prompt):
output_sequences = generator(
prompt=prompt,
overlap_factor=overlap_factor,
num_return_sequences=1,
max_length=1,
do_sample=False,
repetition_penalty=repetition_penalty,
bad_words_ids=bad_words_ids,
output_token_ids=True,
)
if test_mode == 'word':
# Punctuation is not suppressed after the first token, since it provides one of the ways
# by which models can decide that the word has ended. The only straightforward way to implement
# this given how generate() is implemented is to call it twice.
guess_1 = output_sequences[0, -1]
tok_1 = tokenizer.decode([guess_1])
prompt_2 = '{' + prompt + '}' + tok_1
output_sequences_2 = generator(
prompt=prompt_2,
overlap_factor=overlap_factor,
num_return_sequences=1,
max_length=5,
do_sample=False,
repetition_penalty=repetition_penalty,
output_token_ids=True,
)
output_sequences = torch.cat([output_sequences, output_sequences_2], dim=1)
if test_mode == 'token':
guess = output_sequences[0, -1]
return guess
else:
n = output_sequences.shape[1]
j = 1 - word_piece_offset
while j < n - word_piece_offset and is_word_piece(output_sequences[0, j]):
j += 1
end = j + word_piece_offset
guess = output_sequences[0, :end].to('cpu')
return guess
sent_tokenizer = nltk.tokenize.punkt.PunktSentenceTokenizer()
if finetune_sentence_tokenizer:
f = open('../../data/gpt-2/data/lambada_development.jsonl')
text = []
text = text.replace('\n', ' ').replace(' ', ' ').replace('“', '"').replace('”', '"').replace('’', '\'').replace('‘', '\'')
for line in f.readlines():
text.append(json.loads(line)['text'] + ".")
text = '\n'.join(text)
f.close()
sent_tokenizer.train(text)
def split_last_sentence(text):
# The following is necessary to get the sentence tokenizer to behave
regularized_text = text.replace('\n', ' ').replace(' ', ' ').replace('“', '"').replace('”', '"').replace('’', '\'').replace('‘', '\'')
sentences = sent_tokenizer.tokenize(regularized_text)
n = len(sentences[-1])
return text[:-(n+1)], text[-n:]
def interpret_line(line):
text = json.loads(line)['text']
if regularize_text:
text = text.replace('\n', ' ').replace(' ', ' ').replace('“', '"').replace('”', '"').replace('’', '\'').replace('‘', '\'')
# Separate the prompt from the desired output
ids = tokenizer.encode(text, add_special_tokens=False, return_tensors="pt")
if test_mode == 'token':
prompt = ids[0,:-1]
answer = ids[0,-1]
else:
n = ids.shape[1]
i = 1 + word_piece_offset
while i <= n:
if not is_word_piece(ids[0,-i]):
break
i += 1
i -= word_piece_offset
prompt = ids[0,:-i]
answer = ids[0,-i:]
prompt = tokenizer.decode(prompt)
prompt = Program.escape(prompt)
if prompting_mode == 'default':
pass
elif prompting_mode == 'blank':
prompt = f'{prompt}~'
elif prompting_mode == 'fixed':
prompt = f'{prompt}~{fixed_negative_prompt}'
elif prompting_mode == 'word':
toks = nltk.word_tokenize(prompt)
last_tok = Program.escape(toks[-1])
prompt = f'{prompt}~{prefix}{last_tok}'
elif prompting_mode == 'phrase':
phrases = re_phrase_boundary.split(prompt)
last_phrase = Program.escape(phrases[-1])
prompt = f'{prompt}~{prefix}{last_phrase}'
elif prompting_mode == 'sentence':
first_sentences, last_sentence = split_last_sentence(prompt)
last_sentence = Program.escape(last_sentence)
prompt = f'{prompt}~{prefix}{last_sentence}'
elif prompting_mode == 'sentence|blank':
first_sentences, last_sentence = split_last_sentence(prompt)
last_sentence = Program.escape(last_sentence)
prompt = f'{prompt}~{prefix}{{{last_sentence}|}}'
elif prompting_mode == 'sentence|word':
_, last_sentence = split_last_sentence(prompt)
last_sentence = Program.escape(last_sentence)
toks = nltk.word_tokenize(prompt)
last_tok = Program.escape(toks[-1])
prompt = f'{prompt}~{prefix}{{{last_sentence}|{last_tok}}}'
elif prompting_mode == 'sentence|phrase':
_, last_sentence = split_last_sentence(prompt)
last_sentence = Program.escape(last_sentence)
phrases = re_phrase_boundary.split(prompt)
last_phrase = Program.escape(phrases[-1])
prompt = f'{prompt}~{prefix}{{{last_sentence}|{last_phrase}}}'
elif prompting_mode == 'sentence|word|phrase':
_, last_sentence = split_last_sentence(prompt)
last_sentence = Program.escape(last_sentence)
toks = nltk.word_tokenize(prompt)
last_tok = Program.escape(toks[-1])
phrases = re_phrase_boundary.split(prompt)
last_phrase = Program.escape(phrases[-1])
prompt = f'{prompt}~{prefix}{{{last_sentence}|{last_tok}|{last_phrase}}}'
else:
raise ValueError("Unknown prompting mode!")
return text, prompt, answer
f = open('../../data/gpt-2/data/lambada_test.jsonl')
total_score = 0.0
texts = []
prompts = []
answers = []
for line in f.readlines():
text, prompt, answer = interpret_line(line)
texts.append(text)
prompts.append(prompt)
answers.append(answer)
n = 0
ncorrect = 0
for text, prompt, answer in zip(texts, prompts, answers):
guess = run_model(prompt)
n += 1
if model_type == 'ctrl' and test_mode == 'token':
guess = [guess]
answer = [answer]
if test_mode == 'token':
if model_type in ('xlm', 'ctrl'):
guess_text = tokenizer.decode(guess)
m = re_final_punct.match(guess_text)
if m:
guess_text = guess_text[:-len(m.group(1))]
answer_text = tokenizer.decode(answer)
correct = guess_text == answer_text
else:
correct = guess == answer
else:
if model_type in ('xlm', 'ctrl'):
guess_text = tokenizer.decode(guess)
m = re_first_punct_on.match(guess_text)
if m:
guess_text = guess_text[:-len(m.group(1))]
answer_text = tokenizer.decode(answer)
correct = guess_text == answer_text
else:
correct = guess.equal(answer)
if correct:
ncorrect += 1
if n % 100 == 0:
guess = tokenizer.decode(guess)
print('----------')
print(f'Text: {text}')
print(f'Guess: {guess} - {"correct" if correct else "wrong"} ({ncorrect}/{n} = {100*ncorrect/n})')
print(f'Final results: {ncorrect}/{n} = {100*ncorrect/n}')