-
Notifications
You must be signed in to change notification settings - Fork 0
/
inference.py
93 lines (72 loc) · 3.63 KB
/
inference.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
import jsonlines
import json
import torch
from tqdm import tqdm
from config_parser import create_config
from model.models import TFIDFSearcher, RobertaEncoder
from model.CrossCaseCL import CrossCaseCL
from transformers import BertTokenizerFast, AutoModel
from sklearn.metrics.pairwise import cosine_similarity
if __name__ == "__main__":
checkpoint = './output/model/harm_simcse_contra_attn_loss_roberta/3-10k.pkl'
infer_args = dict(tfidf=True, roberta=False, ours=False)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
test_path = './data/new_short/test_example'
if 'jsonl' in test_path:
test_data = list(jsonlines.open(test_path))
else:
test_data = json.load(open(test_path, encoding='utf-8'))
if infer_args['roberta']:
roberta = RobertaEncoder('hfl/chinese-roberta-wwm-ext').to(device)
if infer_args['tfidf']:
tfidf = TFIDFSearcher()
if infer_args['ours']:
config = create_config('previous_configs/config/train/train.config')
ours = CrossCaseCL(config).to(device)
checkpoint = torch.load(checkpoint, map_location=lambda storage, loc: storage)
ours.load_state_dict(checkpoint['model'], strict=False)
tokenizer = BertTokenizerFast.from_pretrained('hfl/chinese-roberta-wwm-ext')
outputs = dict(tfidf=[], roberta=[], ours=[])
for data in tqdm(test_data, desc='Testing...'):
doc_tmp = dict(tfidf=[], roberta=[], ours=[])
fact_list = data['fact']
evidence_list = data['evidence']
test_names = []
test_scores = []
if infer_args['tfidf']:
sim_score_tfidf = tfidf.search(fact_list, evidence_list).tolist()
test_names.append('tfidf')
test_scores.append(sim_score_tfidf)
if infer_args['roberta']:
roberta.eval()
with torch.no_grad():
sim_score_roberta = roberta.search(fact_list, evidence_list).tolist()
test_names.append('roberta')
test_scores.append(sim_score_roberta)
if infer_args['ours']:
ours.eval()
with torch.no_grad():
# ours
inputs_f = tokenizer(fact_list, padding='longest', truncation=True, max_length=128, return_tensors='pt')
inputs_e = tokenizer(evidence_list, padding='longest', truncation=True, max_length=128, return_tensors='pt')
for ipt_key in inputs_f.keys():
inputs_f[ipt_key] = inputs_f[ipt_key].to(device)
inputs_e[ipt_key] = inputs_e[ipt_key].to(device)
embedding_f = ours.encoder_f(**inputs_f)
embedding_e = ours.encoder_f(**inputs_e)
sim_score_ours = cosine_similarity(embedding_f.cpu(), embedding_e.cpu()).tolist()
test_names.append('ours')
test_scores.append(sim_score_ours)
for name, scores in list(zip(test_names, test_scores)):
for fid, score in enumerate(scores):
fact = fact_list[fid]
tmp = dict(query=fact, evidences=[])
for eid, s in enumerate(score):
evidence = evidence_list[eid]
tmp['evidences'].append({'evidence': evidence, 'pred_score': s})
tmp['evidences'] = sorted(tmp['evidences'], key=lambda x: x['pred_score'], reverse=True)
doc_tmp[name].append(tmp)
outputs[name].append(doc_tmp[name])
for name in outputs:
with open('./data/new_short/predictions/{}-pred.json'.format(name), 'w', encoding='utf-8') as f:
json.dump(outputs[name], f, indent=4, ensure_ascii=False)