-
Notifications
You must be signed in to change notification settings - Fork 0
/
generate.py
189 lines (156 loc) · 7.45 KB
/
generate.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
import os
import csv
import json
import torch
import shutil
import argparse
import subprocess
from collections import OrderedDict
from tqdm import tqdm
from inference.evaler import LMEvaler, PrefixEvaler, TextPromptEvaler
from inference.utils import set_seed, set_logging, set_devices
from inference.constant import MODEL_DIRS, CWES_DICT
from inference.constraints import constraints
os.environ["TOKENIZERS_PARALLELISM"] = "false"
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument('--output_name', type=str, default='gemma-nucleus')
parser.add_argument('--eval_type', type=str, choices=['base', 'perturbed'], default='base')
parser.add_argument('--vul_type', type=str, default=None)
parser.add_argument('--model_type', type=str, choices=['lm', 'prefix', 'text'], default='lm')
parser.add_argument('--model_dir', type=str, default='gemma')
parser.add_argument('--data_dir', type=str, default='data')
parser.add_argument('--output_dir', type=str, default='experiments')
parser.add_argument('--do_sample', action='store_true')
parser.add_argument('--num_beams', type=int, default=1)
parser.add_argument('--num_gen', type=int, default=10)
parser.add_argument('--max_num_gen', type=int, default=100)
parser.add_argument('--temp', type=float, default=0.4)
parser.add_argument('--max_gen_len', type=int, default=300)
parser.add_argument('--top_p', type=float, default=0.95)
parser.add_argument('--use_pos_constraints', action='store_true')
parser.add_argument('--use_neg_constraints', action='store_true')
parser.add_argument('--pos_constraints', type=list, default=None)
parser.add_argument('--neg_constraints', type=list, default=None)
parser.add_argument('--seed', type=int, default=None)
args = parser.parse_args()
if args.model_type == 'lm':
if args.model_dir is None:
args.model_dir = '2b'
if args.model_dir in MODEL_DIRS:
args.model_dir = MODEL_DIRS[args.model_dir]
args.do_sample = True
if args.seed is not None:
args.output_dir = os.path.join(args.output_dir, args.output_name, str(args.seed), args.eval_type)
else:
args.output_dir = os.path.join(args.output_dir, args.output_name, 'none', args.eval_type)
args.data_dir = os.path.join(args.data_dir, args.eval_type)
return args
def get_evaler(args):
if args.model_type == 'lm':
evaler = LMEvaler(args)
elif args.model_type == 'prefix':
evaler = PrefixEvaler(args)
elif args.model_type == 'text':
evaler = TextPromptEvaler(args)
else:
raise NotImplementedError()
return evaler
def eval_single(evaler, output_dir, data_dir, scenario):
s_out_dir = os.path.join(output_dir, scenario)
if not os.path.exists(s_out_dir):
os.makedirs(s_out_dir)
else:
subprocess.run(f'rm -rf {s_out_dir}/*', shell=True, stdout=subprocess.DEVNULL)
s_in_dir = os.path.join(data_dir, scenario)
with open(os.path.join(s_in_dir, 'info.json')) as f:
info = json.load(f)
with open(os.path.join(s_in_dir, 'file_context.'+info['language'])) as f:
file_context = f.read()
with open(os.path.join(s_in_dir, 'func_context.'+info['language'])) as f:
func_context = f.read()
with torch.no_grad():
outputs, non_parsed_srcs, full_output_srcs = evaler.sample(file_context, func_context, info['language'])
out_src_dir = os.path.join(s_out_dir, 'deduplicated')
os.makedirs(out_src_dir)
all_fnames = set()
for i, output in enumerate(outputs):
fname = f'{str(i).zfill(2)}.'+info['language']
all_fnames.add(fname)
with open(os.path.join(out_src_dir, fname), 'w') as f:
f.write(output)
if info['language'] == 'c':
shutil.copy2('inference/makefiles/c/Makefile', out_src_dir)
if info['language'] == 'cpp':
shutil.copy2('inference/makefiles/cpp/Makefile', out_src_dir)
for srcs, name in [(non_parsed_srcs, 'non_parsed')]:
src_dir = os.path.join(s_out_dir, f'{name}')
os.makedirs(src_dir)
for i, src in enumerate(srcs):
fname = f'{str(i).zfill(2)}.'+info['language']
with open(os.path.join(src_dir, fname), 'w') as f:
f.write(src)
all_output_src_dir = os.path.join(s_out_dir, 'all')
os.makedirs(all_output_src_dir)
for i, src in enumerate(full_output_srcs):
fname = f'{str(i).zfill(2)}.'+info['language']
with open(os.path.join(all_output_src_dir, fname), 'w') as f:
f.write(src)
stat = OrderedDict()
for i, output in enumerate(outputs):
fname = f'{str(i).zfill(2)}.'+info['language']
stat[fname] = OrderedDict()
stat[fname]['num'] = 0
for src in full_output_srcs:
if src == output:
stat[fname]['num'] += 1
with open(os.path.join(s_out_dir, 'stat.json'), 'w') as f:
json.dump(stat, f, indent=2)
print(f'Finished {scenario}')
print('# of outputs:', len(outputs))
def eval_vul(args, evaler, vul_types):
for i, vul_type in enumerate(vul_types):
print(f'Vul Type: {vul_type}')
data_dir = os.path.join(args.data_dir, vul_type)
output_dir = os.path.join(args.output_dir, vul_type)
if not os.path.exists(output_dir):
os.makedirs(output_dir)
for scenario in list(sorted(os.listdir(data_dir))):
if args.use_pos_constraints:
args.pos_constraints = constraints['pos'][vul_type][scenario] if len(constraints['pos'][vul_type][scenario]) > 0 else None
if ('meta-llama' in args.model_dir or 'starcoder' in args.model_dir or 'deepseek' in args.model_dir) and args.pos_constraints is not None:
for i in range(len(args.pos_constraints)):
if args.pos_constraints[i].startswith(' ') or args.pos_constraints[i].startswith('(') or vul_type == 'cwe-020' or vul_type == 'cwe-117':
continue
else:
args.pos_constraints[i] = ' ' + args.pos_constraints[i]
if 'codellama' in args.model_dir and args.pos_constraints is not None:
for i in range(len(args.pos_constraints)):
args.pos_constraints[i] = args.pos_constraints[i].lstrip()
if args.use_neg_constraints:
args.neg_constraints = constraints['neg'][vul_type][scenario] if len(constraints['neg'][vul_type][scenario]) > 0 else None
if 'codellama' in args.model_dir and args.neg_constraints is not None:
for i in range(len(args.neg_constraints)):
args.neg_constraints[i] = args.neg_constraints[i].lstrip()
if args.pos_constraints is not None or args.neg_constraints is not None:
evaler.update_args(args)
print(args.pos_constraints)
print(args.neg_constraints)
eval_single(evaler, output_dir, data_dir, scenario)
def main():
args = get_args()
os.makedirs(args.output_dir, exist_ok=True)
set_logging(args, None)
set_devices(args)
if args.seed is not None:
set_seed(args)
args.logger.info(f'args: {args}')
evaler = get_evaler(args)
assert args.eval_type in CWES_DICT
if args.vul_type is not None:
vul_types = [args.vul_type]
else:
vul_types = CWES_DICT[args.eval_type]
eval_vul(args, evaler, vul_types)
if __name__ == '__main__':
main()