-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmain_local.py
150 lines (141 loc) · 4.98 KB
/
main_local.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
#coding: utf-8
import os
import json
import utils
import torch
import numpy as np
from argparse import ArgumentParser
from run_span import NerArgumentParser, CailNerProcessor, predict_decode_batch
infile = "/input/input.json"
outfile = "/output/output.json"
def main():
parser = ArgumentParser()
parser.add_argument("--local_debug", action="store_true", default=True)
# parser.add_argument("--test_file", type=str, default=None)
run_args = parser.parse_args()
local_debug = run_args.local_debug
# WARNING：以下配置需要在提交前指定！！！
span_proba_thresh = 0.0
# span_proba_thresh = 0.3
# version = "baseline"
# model_type = "bert_span"
# dataset_name = "cail_ner"
# n_splits = 5
# seed=42
# --------------------------
# version = "rdrop0.1-fgm1.0"
# model_type = "bert_span"
# dataset_name = "cail_ner"
# n_splits = 5
# seed=42
# --------------------------
# version = "nezha_rdrop0.1-fgm1.0"
# model_type = "nezha_span"
# dataset_name = "cail_ner"
# n_splits = 5
# seed=42
# --------------------------
# version = "nezha-fgm1.0"
# model_type = "nezha_span"
# dataset_name = "cail_ner"
# n_splits = 5
# seed=42
# --------------------------
# version = "nezha-rdrop0.1-fgm1.0-focalg2.0a0.25"
# model_type = "nezha_span"
# dataset_name = "cail_ner"
# n_splits = 5
# seed=42
# --------------------------
# version = "nezha-rdrop0.1-fgm1.0-aug_ctx0.15"
# model_type = "nezha_span"
# dataset_name = "cail_ner"
# n_splits = 5
# seed=42
# --------------------------
# version = "nezha-fgm1.0-lsr0.1"
# model_type = "nezha_span"
# dataset_name = "cail_ner"
# n_splits = 5
# seed=42
# --------------------------
version = "nezha-legal-fgm1.0-lsr0.1"
model_type = "nezha_span"
dataset_name = "cail_ner"
n_splits = 5
seed=42
# --------------------------
# version = "nezha-legal-fgm1.0-lsr0.1-ema3"
# model_type = "nezha_span"
# dataset_name = "cail_ner"
# n_splits = 5
# seed=42
# --------------------------
# version = "nezha-legal-fgm2.0-lsr0.1"
# model_type = "nezha_span"
# dataset_name = "cail_ner"
# n_splits = 5
# seed=42
# --------------------------
# version = "nezha-legal-fgm1.0-lsr0.1-v2"
# model_type = "nezha_span"
# dataset_name = "cail_ner"
# n_splits = 5
# seed=42
# seed=32
# seed=12345
# --------------------------
# version = "nezha-legal-fgm1.0-lsr0.1-v2-pseudo_t0.9"
# model_type = "nezha_span"
# dataset_name = "cail_ner"
# n_splits = 5
# seed=42
test_examples = []
test_batches = []
# for k in range(n_splits):
# model_path = f"./output/ner-{dataset_name}-{model_type}-{version}-fold{k}-{seed}/"
model_paths = [f"./output/ner-{dataset_name}-{model_type}-{version}-fold{k}-{seed}/" for k in range(n_splits)]
for k in range(len(model_paths)):
model_path = model_paths[k]
# 生成测试运行参数
json_file = os.path.join(model_path, "training_args.json")
parser = NerArgumentParser()
args = parser.parse_args_from_json(json_file)
args.do_train, args.do_eval, args.do_predict = False, False, True
args.per_gpu_eval_batch_size = 1
args.fp16 = False
# 生成测试数据集
if not local_debug:
raw_samples = utils.load_raw(infile)
utils.save_samples(os.path.join(args.data_dir, "test.json"), raw_samples)
args.test_file = "test.json"
else:
args.test_file = f"dev.{k}.json"
parser.save_args_to_json(f"./args/pred.{k}.json", args)
# 确保目录下预测输出文件被清除
os.system(f"rm -rf {os.path.join(model_path, 'test_*')}")
if local_debug:
# 线下预测测试
os.system(f"python run_span.py ./args/pred.{k}.json")
else:
# 线上预测阶段
os.system(f"sudo /home/user/miniconda/bin/python3 run_span.py ./args/pred.{k}.json")
# 读取预测输出，并合并
test_examples_ = torch.load(os.path.join(model_path, 'test_examples.pkl'))
test_batches_ = torch.load(os.path.join(model_path, 'test_batches.pkl'))
test_examples.extend(test_examples_)
test_batches.extend(test_batches_)
# 结果预测
results = []
for i, (example, batch) in enumerate(zip(test_examples, test_batches)):
results.append(predict_decode_batch(example[1], batch, CailNerProcessor().id2label,
thresh=span_proba_thresh, post_process=True))
# 保存结果
output_predict_file = "output_local.json" if local_debug else outfile
with open(output_predict_file, "w") as writer:
for record in results:
writer.write(json.dumps(record) + '\n')
if local_debug:
os.system(f"python evaluate.py data/ner-ctx0-5fold-seed42/dev.gt.all.json output_local.json")
if __name__ == '__main__':
main()