-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain_sbert_search.py
223 lines (166 loc) · 7.5 KB
/
train_sbert_search.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
import json
import os
import random
import string
import nltk
from nltk.corpus import stopwords
from pytorch_pretrained_bert import BertTokenizer
from sentence_transformers import SentenceTransformer, InputExample, losses, util
from torch.utils.data import DataLoader
import pandas as pd
nltk.download('stopwords')
root = "/Users/sahiravi/Documents/Research/VL project/scratch/data/coco"
root = "/ubc/cs/research/nlp/sahiravi/vlc_transformer/scratch/data/coco"
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
s = set(stopwords.words('english'))
semantic_search_model = "msmarco-roberta-base-ance-firstp"
def _load_json(path):
with open(path, 'r') as f:
return json.load(f)
def _intersection(lst1, lst2):
return len(set(lst1).intersection(lst2)) > 0
def prepare_data_ok(exp_name, set_name, base_dir):
# Load questions:
annotations = _load_json(f'{root}/okvqa/mscoco_{set_name}2014_annotations.json')["annotations"]
questions_path = f'{root}/okvqa/OpenEnded_mscoco_{set_name}2014_questions.json'
question_csv = questions_path.split("json")[0] + 'csv'
questions_df = pd.read_csv(question_csv)
questions_df["question_id"] = questions_df["question_id"].astype(str)
questions = dict(
zip(list(questions_df["question_id"].values), list(questions_df["question"].values)))
# # Load expansions:
# expansions = _load_json(
# f'{root}/okvqa/commonsense/expansions/' + exp_name + '_okvqa_' + set_name + '.json')
# Load raw expansions:
raw_expansions = _load_json(f'{root}/okvqa/commonsense/expansions/question_expansion_sentences_{set_name}_okvqa_{exp_name}.json')
good_data_list = []
bad_data_list = []
data_list = []
for q in annotations:
q_id = q['question_id']
im_id = q['image_id']
raw_ques = questions[str(q_id)]
ans = q["answers"]
ans_text = ' '.join([a['answer'] for a in ans])
ans_text = ans_text.translate(str.maketrans('', '', string.punctuation))
ans_text = ans_text.lower()
ans_tokens = tokenizer.tokenize(ans_text)
ans_tokens = [t for t in ans_tokens if t not in s]
#exps = expansions['{:012d}.jpg'.format(im_id)][str(q_id)][0]
#exps = exps.split('.')
exps = raw_expansions[str(q_id)]
exps = [e.strip() + '.' for e in exps]
for exp_text in exps:
raw_text = exp_text
exp_text = exp_text.translate(str.maketrans('', '', string.punctuation))
exp_text = exp_text.lower()
exp_tokens = tokenizer.tokenize(exp_text)
exp_tokens = [t for t in exp_tokens if t not in s]
if _intersection(ans_tokens, exp_tokens):
good_data_list.append({'sents': [raw_ques, raw_text], 'label': 0.8})
else:
bad_data_list.append({'sents': [raw_ques, raw_text], 'label': 0.2})
random.shuffle(bad_data_list)
bad_data_list = bad_data_list[:len(good_data_list)]
data_list = good_data_list + bad_data_list
random.shuffle(data_list)
print('good:', len(good_data_list))
print('bad:', len(bad_data_list))
print('total:', len(data_list))
# Save data:
savepath = os.path.join(base_dir, exp_name + '_okvqa_' + set_name + '.json')
with open(savepath, 'w') as f:
json.dump(data_list, f, indent=4)
return savepath
def prepare_data(exp_name, set_name, base_dir):
# Load questions:
annotations = _load_json(
f'{root}/aokvqa/aokvqa_v1p0_' + set_name + '.json')
# Load expansions:
# expansions = _load_json(
# f'{root}/aokvqa/commonsense/expansions/' + exp_name + '_aokvqa_' + set_name + '.json')
raw_expansions = _load_json(
f'{root}/aokvqa/commonsense/expansions/question_expansion_sentences_{set_name}_aokvqa_{exp_name}.json')
good_data_list = []
bad_data_list = []
data_list = []
for q in annotations:
q_id = q['question_id']
im_id = q['image_id']
raw_ques = q['question']
ans_text = ' '.join(q['direct_answers'])
ans_text = ans_text.translate(str.maketrans('', '', string.punctuation))
ans_text = ans_text.lower()
ans_tokens = tokenizer.tokenize(ans_text)
ans_tokens = [t for t in ans_tokens if t not in s]
#exps = expansions['{:012d}.jpg'.format(im_id)][str(q_id)][0]
#exps = exps.split('.')
exps = raw_expansions[str(q_id)]
exps = [e.strip() + '.' for e in exps]
for exp_text in exps:
raw_text = exp_text
exp_text = exp_text.translate(str.maketrans('', '', string.punctuation))
exp_text = exp_text.lower()
exp_tokens = tokenizer.tokenize(exp_text)
exp_tokens = [t for t in exp_tokens if t not in s]
if _intersection(ans_tokens, exp_tokens):
good_data_list.append({'sents': [raw_ques, raw_text], 'label': 0.8})
else:
bad_data_list.append({'sents': [raw_ques, raw_text], 'label': 0.2})
random.shuffle(bad_data_list)
bad_data_list = bad_data_list[:len(good_data_list)]
data_list = good_data_list + bad_data_list
random.shuffle(data_list)
print('good:', len(good_data_list))
print('bad:', len(bad_data_list))
print('total:', len(data_list))
# Save data:
savepath = os.path.join(base_dir, exp_name + '_aokvqa_' + set_name + '.json')
with open(savepath, 'w') as f:
json.dump(data_list, f, indent=4)
return savepath
def train(train_file, base_dir):
# Define the model. Either from scratch of by loading a pre-trained model
model = SentenceTransformer(semantic_search_model)
# Define your train examples. You need more than just two examples...
train_egs = _load_json(train_file)
train_examples = [InputExample(texts=e['sents'], label=e['label']) for e in train_egs]
# Define your train dataset, the dataloader and the train loss
train_dataloader = DataLoader(train_examples, shuffle=True, batch_size=64)
train_loss = losses.CosineSimilarityLoss(model)
# Tune the model
model.fit(train_objectives=[(train_dataloader, train_loss)],optimizer_params ={'lr': 1e-5},
epochs=5, warmup_steps=1000, output_path=base_dir)
def test(base_dir):
model = SentenceTransformer(base_dir)
# Two lists of sentences
sentences1 = ['What is the mood of the cow?',
'What is on the table?',
'What is missing from the dinner?']
sentences2 = ['the cow is happy',
'a woman watches TV',
'man is capable of eating dinner']
# Compute embedding for both lists
embeddings1 = model.encode(sentences1, convert_to_tensor=True)
embeddings2 = model.encode(sentences2, convert_to_tensor=True)
# Compute cosine-similarities
cosine_scores = util.cos_sim(embeddings1, embeddings2)
# Output the pairs with their score
for i in range(len(sentences1)):
print("{} \t\t {} \t\t Score: {:.4f}".format(sentences1[i], sentences2[i], cosine_scores[i][i]))
if __name__ == '__main__':
aok = True
if aok:
base_path = 'sbert-aok-qo/'
if not os.path.exists(base_path):
os.makedirs(base_path)
savepath = prepare_data('semq.o', 'train', base_path)
train(savepath, base_path)
test(base_path)
else:
base_path = 'sbert-ok-qo/'
if not os.path.exists(base_path):
os.makedirs(base_path)
savepath = prepare_data_ok('semq.o', 'train', base_path)
train(savepath, base_path)
test(base_path)