-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathGAuGE-GPT3.py
215 lines (188 loc) · 8.25 KB
/
GAuGE-GPT3.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
import pandas as pd
import pyterrier as pt
pt.init()
from time import sleep
import json
from tqdm import tqdm
import ir_datasets
import openai
import os
from openai.error import RateLimitError
import pyterrier_dr
import pyterrier_pisa
import heapq
import random
from pyterrier_t5 import MonoT5ReRanker
from rouge_score import rouge_scorer
model, prefix = 'text-davinci-edit-001', 'davinci'
LAMBDA = 10 # scaling parameter
ROUGETYPE = 'rouge2'
monoT5 = MonoT5ReRanker() # loads castorini/monot5-base-msmarco by default
scorer = rouge_scorer.RougeScorer([ROUGETYPE], use_stemmer=True)
openai.api_key = ''
bm25 = pyterrier_pisa.PisaIndex.from_dataset('msmarco_passage').bm25(num_results=1000)
electra = pyterrier_dr.ElectraScorer(verbose=False)
pipeline = bm25 >> pt.text.get_text(pt.get_dataset('irds:msmarco-passage'), 'text') >> electra
def query_rewrite(docid, passage, query, count=1):
while True:
try:
result = openai.Edit.create(
engine='text-davinci-edit-001',
input=passage,
instruction='Re-write the passage to better answer the question: ' + query,
api_key=os.getenv('OPENAI'),
temperature=0.7,
top_p=1,
n=count,
)
break
except openai.error.RateLimitError:
print('RATE LIMIT, sleeping for 10 seconds')
sleep(10)
for idx, g in enumerate(result['choices']):
if 'text' not in g.keys():
text = ''
print('empty')
else:
text = g['text'].strip().replace('\n', ' ').strip()
yield {'docno': f'{docid}-qr{idx}', 'text': text}
def rewrite(docid, passage, count=1):
while True:
try:
result = openai.Edit.create(
engine='text-davinci-edit-001',
input=passage,
instruction='Re-write the passage',
api_key=os.getenv('OPENAI'),
temperature=0.7,
top_p=1,
n=count,
)
break
except openai.error.RateLimitError:
print('RATE LIMIT, sleeping for 10 seconds')
sleep(10)
for idx, g in enumerate(result['choices']):
if 'text' not in g.keys():
text = ''
print('empty')
else:
text = g['text'].strip().replace('\n', ' ').strip()
yield {'docno': f'{docid}-r{idx}', 'text': text}
def combine(docno1, docno2, passage1, passage2, query, count=1):
while True:
try:
result = openai.Edit.create(
engine='text-davinci-edit-001',
input=f'Passage1: {passage1}\n\nPassge2: {passage2}\n\nAnswer:',
instruction='Combine ideas from both Passage1 and Passage2 to answer the question: ' + query,
api_key=os.getenv('OPENAI'),
temperature=0.7,
top_p=1,
n=count,
)
break
except openai.error.RateLimitError:
print('RATE LIMIT, sleeping for 10 seconds')
sleep(10)
for idx, g in enumerate(result['choices']):
text = ''
if 'text' in g.keys():
if '\n\nAnswer:' in g['text']:
text = g['text'].split('\n\nAnswer:')[1]
else:
print('empty')
yield {'docno': f'({docno1}+{docno2})-{idx}', 'text': text.replace('\n', ' ').strip()}
for DEPTH, DOC_DEPTH, no_of_mutations_per_iteration in [(2, 10, 8), (2, 2, 12)]: # not tuned vs tuned
# parameters recommended in Gen2IR baseline
print(DEPTH)
print(DOC_DEPTH)
print(no_of_mutations_per_iteration)
for ds, dsid in [('dl19', 'msmarco-passage/trec-dl-2019/judged'), ('dl20', 'msmarco-passage/trec-dl-2020/judged'),
('dev', 'msmarco-passage/dev/small')]:
savedf = pd.DataFrame(
columns=['Query', 'Result_RR', 'Top2', 'MonoT5_RR', 'Result_Gen', 'MonoT5_Gen', 'Iter_Gen', 'Judge'])
query_count = 0
for i, query in enumerate(tqdm(ir_datasets.load(dsid).queries)):
query_count += 1
print(query)
res = pipeline.search(query.text)
# print(res['text'].iloc[2])
original = res['text'].iloc[0]
#print('original: ' + original)
top1 = original
top2 = res['text'].iloc[1]
print('top1')
print(top1)
print('top2')
print(top2)
# Add re-ranking retrieval results to heap
heap = [(float('-inf'), '', '')]
for item in res.itertuples(index=False):
scores = scorer.score(top1 + ' ' + top2, item.text)
# print(item.score)
# print(scores[ROUGETYPE][2])
# print(item.score + scores[ROUGETYPE][2])
heap.append((item.score, item.docno, item.text))
heap = sorted(heap)
iter = 0
while True:
last_heap_depth_score = heap[-1 * DEPTH][0]
# Mutations
res = []
for n in range(no_of_mutations_per_iteration):
iter += 1
case = random.random()
try:
if case <= 0.33:
docid = int(random.random() * 100) % DOC_DEPTH + 1
res.extend(rewrite(heap[-1 * docid][1], heap[-1 * docid][2]))
elif case <= 0.66:
docid = int(random.random() * 100) % DOC_DEPTH + 1
res.extend(query_rewrite(heap[-1 * docid][1], heap[-1 * docid][2], query.text))
else:
docid1 = int(random.random() * 100) % DOC_DEPTH + 1
docid2 = int(random.random() * 100) % DOC_DEPTH + 1
if docid1 == docid2:
if docid1 == DOC_DEPTH:
docid2 -= 1
else:
docid2 += 1
res.extend(combine(heap[-1 * docid1][1], heap[-1 * docid2][1], heap[-1 * docid1][2],
heap[-1 * docid2][2], query.text))
except:
print('err_er')
continue
# Evaluate new documents
res = pd.DataFrame({'qid': query.query_id, 'query': query.text, 'docno': [x['docno'] for x in res],
'text': [x['text'] for x in res]})
res = electra(res)
# print(res)
# Add new documents to heap
for item in res.itertuples(index=False):
scores = scorer.score(top1 + ' ' + top2, item.text)
#print(scores[ROUGETYPE][2])
#print(item.score)
#lambda predetermined for normalization
heap.append((item.score+ LAMBDA*scores[ROUGETYPE][2], item.docno, item.text))
heap = sorted(heap)
print(' '.join([str(x) for x in heap[-1]]))
# Termination criteria
if heap[-1 * DEPTH][0] <= last_heap_depth_score:
break
print('final')
print(heap[-1][2])
inp1 = pd.DataFrame([['q1', query, 'd1', original]], columns=['qid', 'query', 'docno', 'text'])
mono_score_rr = monoT5.transform(inp1).loc[0].at["score"]
inp2 = pd.DataFrame([['q1', query, 'd1', heap[-1][2]]], columns=['qid', 'query', 'docno', 'text'])
mono_score_gen = monoT5.transform(inp2).loc[0].at["score"]
j = mono_score_gen - mono_score_rr
if j > 0:
j = 1
elif j < 0:
j = -1
savedf = pd.concat([savedf, pd.DataFrame([{'Query': query, 'Result_RR': original, 'Top2': top2, 'MonoT5_RR': mono_score_rr, 'Result_Gen': heap[-1][2],
'MonoT5_Gen': mono_score_gen, 'Iter_Gen': iter, 'Judge': j}])], ignore_index=True)
print(savedf)
savedf.to_csv('gpt3/' + ROUGETYPE + ds + '_' + str(DEPTH) + '_' + str(DOC_DEPTH) + '_' + str(
no_of_mutations_per_iteration) + '.csv', index=False)