-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathgen_end2end.py
160 lines (130 loc) · 6.52 KB
/
gen_end2end.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
import argparse
import copy
import numpy as np
import random
import re
from tqdm import tqdm
import functools
from contextlib import closing
# import multiprocessing as mp
from multiprocessing import Pool
from utils import *
def worker_process(data_in):
pid, records, cuda_id = data_in
import os
os.environ["CUDA_VISIBLE_DEVICES"]=str(cuda_id)
import torch
from load_model import gen_zhixi_model
from triple_level_search import init_toolkit, gen_step, generate_given_triples2
print(f'Start process {pid} on {cuda_id}')
head_mode = 'head-first'
head_model, tokenizer = gen_zhixi_model(
base_model='model_hub/zhixi-13b-with-lora-merged',
lora_weights='')
head_self = head_model._orig_mod
device = head_self.device
tools = init_toolkit()
end2end_result = []
for record in records:
try:
idx = record['id'] % 1000000
input = record['input']
instruction = record['instruction']
raw_relation_set = record['instruction'].split(']')[0][12:-1].split("', '")
print(idx, input)
if raw_relation_set == ['']:
print(f'WARNING: Empty relation set for idx {idx}')
print('~~~~~~\n')
continue
# 中文括号替换英文括号，在 main.py 中会替换回来
has_bracket = '(' in input or ')' in input
if has_bracket:
input = copy.deepcopy(input).replace('(', '（').replace(')', '）').replace('[', '<').replace(']', '>')
# 替换部分 relation 以达到更好效果，在 main.py 中会替换回来
if record['cate'] == '医学':
change_flag = False
for rel_idx, rel in enumerate(raw_relation_set):
if rel == '包含':
raw_relation_set[rel_idx] = '属于'
change_flag = True
if change_flag:
print(f'WARNING: change relation_set to {raw_relation_set}')
if record['cate'] == '自然科学':
change_flag = False
for rel_idx, rel in enumerate(raw_relation_set):
if rel == '组成':
raw_relation_set[rel_idx] = '组成成分'
change_flag = True
if rel == '性质':
raw_relation_set[rel_idx] = '特性'
change_flag = True
if rel == '用途':
raw_relation_set[rel_idx] = '应用场景'
change_flag = True
if change_flag:
print(f'WARNING: change relation_set to {raw_relation_set}')
if idx == 545 and raw_relation_set == ['2020年8月7日']:
raw_relation_set = ['类型', '上映时间']
for rel_idx in range(len(raw_relation_set)):
try:
if idx == 539 and rel_idx == 0: continue # 这一条会不断生成相同的三元组结果直到达到最大生成长度，可以手动直接跳过
# 对于有 n 个 relation 的句子，我们会将 relation_set 打乱顺序 n 次
# 对于第 i 次，将原始 raw_relation_set 中的第 i 个 relation 置于最前，其他的打乱顺序排列
relation_set = copy.deepcopy(raw_relation_set)
relation_set.remove(raw_relation_set[rel_idx])
random.shuffle(relation_set)
relation_set = [raw_relation_set[rel_idx]] + relation_set
print(relation_set)
prefix = ''
input_ids = gen_inputs(tools['prompt'], tokenizer, input, relation_set, mode=head_mode, prefix=prefix)
output_ids, prob_list = gen_step(
head_self, tokenizer, input_ids.to(device),
input, relation_set,
generation_config=tools['gen_config'],
triple_type=head_mode
)
output_text = prefix + tokenizer.decode(output_ids[0, input_ids.size(1):], skip_special_tokens=True)
output_pairs = split_triple_probs(tokenizer, input, relation_set, input_ids[0].cpu().numpy(), output_ids[0].cpu().numpy(), np.array(prob_list))
end2end_result.append((idx, relation_set, output_text, np.array(prob_list), output_pairs))
except Exception as e:
print(f'Error: when generate end2end results for {idx} with relation set {relation_set}', e)
print('~~~~~~\n')
except Exception as e:
print('Error', idx, e)
return end2end_result, pid
def get_args(args=None):
parser = argparse.ArgumentParser()
parser.add_argument('--data_path', type=str, default='data/test.json')
parser.add_argument('--new_temp_path', type=str, default='new_temp.pt')
parser.add_argument('--num_process', type=int, default=1)
parser.add_argument('--gpu', nargs='+', type=int, default=[0])
return parser.parse_args(args)
if __name__ == '__main__':
args = get_args()
if args.num_process == 1 and len(args.gpu) > 0:
print(f'Will launch on {len(args.gpu)} GPU(s) {args.gpu}')
elif args.num_process > 1 and args.num_process == len(args.gpu):
print(f'Will launch {args.num_process} tasks on {args.num_process} GPUs {args.gpu}')
else:
print(
'''Running arguments error. For example, you can use:
python gen_end2end.py --num_process 4 --gpu 0,1,2,3
This command launches 4 process, each on one gpu. It requires more than 50 GB memory for each GPU.
python gen_end2end.py --num_process 1 --gpu 0,1,2,3
This command launches a single process on 4 gpus. It requires more than 15 GB memory for each GPU.
''')
exit()
records = load_data(args.data_path)
if args.num_process == 1:
end2end_result, _ = worker_process((
0, records, ','.join([str(ele) for ele in args.gpu])))
else:
tasks = [(i, records[i::args.num_process], args.gpu[i]) for i in range(args.num_process)]
end2end_result = []
with closing(Pool(processes=args.num_process)) as pool:
for ret, ret_id in pool.imap_unordered(functools.partial(worker_process), tasks):
print(ret_id, 'finished')
end2end_result += ret
import torch
h_rel_result, t_rel_result, mainitem_result = {}, {}, {}
torch.save((end2end_result, h_rel_result, t_rel_result, mainitem_result), args.new_temp_path)