-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathgenerate_script.py
190 lines (168 loc) · 7.57 KB
/
generate_script.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
from __future__ import absolute_import, division, print_function
import glob
import importlib
import inspect
import os
import decorator
import numpy as np
import six
import tensorflow as tf
flags = tf.flags
flags.DEFINE_boolean("train", True, "train models")
flags.DEFINE_boolean("carlini", False, "test models using Carlini l2-attack")
FLAGS = tf.app.flags.FLAGS
@decorator.decorator
def concat_commands(f, *args, **kwargs):
commands = f(*args, **kwargs)
print("\n".join(commands))
def get_tmpl_str(f, script_name, add_args=None, exclude_args=None):
"""
script_name: python module where flags are defined
add_args: additional flags which are not included in the function definition.
exclude_args: flags to exclude from function signature
"""
if add_args is None:
add_args = []
if exclude_args is None:
exclude_args = []
argspec = inspect.getargspec(f)
arg_names = argspec.args
if len(add_args) > 0:
for arg in add_args:
if isinstance(arg, (tuple, list)) > 0:
arg_names.insert(arg[0], arg[1])
else:
arg_names.append(arg)
script_module = importlib.import_module(script_name)
defined_flags = script_module.FLAGS._flags().keys()
str_bfr = six.StringIO()
str_bfr.write("python %(script_name)s.py " % locals())
for arg_name in arg_names:
if arg_name not in exclude_args:
assert arg_name in defined_flags, arg_name
str_bfr.write("--%(arg_name)s=%%(%(arg_name)s)s " % locals())
tmpl_str = str_bfr.getvalue()[:-1]
return tmpl_str
@concat_commands
def generate_critic(name='critic',
model='mlp', layer_dims="1200-1200-1200", activation_fn='relu',
critic_model='mlp', critic_layer_dims="1200-1200",
pretrain_niter=0, niter=100,
lmbd=1.0, lmbd_rec_l1=0, lmbd_rec_l2=0, lmbd_grad=10, weight_decay=0.0,
lr=0.0005, lr_decay_factor=0.5, lr_decay_step=40,
critic_lr=0.001, critic_lr_decay_factor=0.5, critic_lr_decay_step=40,
attack_iter=50, attack_overshoot=0.02,
attack_confidence="same", val_attack_confidence=0.8,
attack_random=True,
attack_uniform=False,
attack_label_smoothing=0.0,
train_dir="runs_%(model)s",
seed=1,
gpu_memory=1.0,
runs=10):
np.random.seed(seed)
tmpl_str = get_tmpl_str(
generate_critic, 'train_critic', exclude_args=['runs'])
name = name % locals()
train_dir = train_dir % locals()
for i in range(runs):
seed = np.random.randint(1234)
yield tmpl_str % locals()
@concat_commands
def generate_test(root_dir,
working_dir=None,
attack_iter=50,
attack_clip=0.5,
attack_box_clip=False,
attack_overshoot=0.02,
hc_confidence=0.8,
dataset="test",
sort_labels=True,
filter_dirs=False):
tmpl_str = get_tmpl_str(
generate_test,
'test',
add_args=[(0, 'load_dir')],
exclude_args=['root_dir', 'filter_dirs'])
working_dirs = glob.glob(working_dir + '/*')
for load_dir in sorted(glob.glob(root_dir)):
if working_dir is None:
working_dir = os.path.abspath(os.path.join(load_dir, os.pardir))
working_dir = working_dir.replace("runs", "test")
else:
working_dir = working_dir % locals()
working_dirs = [
os.path.basename(working_path)
for working_path in glob.glob(os.path.join(working_dir, '*'))
]
if not filter_dirs or os.path.basename(load_dir) not in working_dirs:
yield tmpl_str % locals()
@concat_commands
def generate_test_carlini(root_dir,
working_dir=None,
num_examples=10000,
batch_size=100,
carlini_batch_size=100,
carlini_max_iter=10000,
carlini_confidence=0,
carlini_binary_steps=9,
generate_summary=True,
sort_labels=True,
dataset="test",
filter_dirs=False):
tmpl_str = get_tmpl_str(
generate_test_carlini,
'test_carlini',
add_args=[(0, 'load_dir')],
exclude_args=['root_dir', 'filter_dirs'])
working_dirs = glob.glob(working_dir + '/*')
for load_dir in sorted(glob.glob(root_dir)):
if working_dir is None:
working_dir = os.path.abspath(os.path.join(load_dir, os.pardir))
working_dir = working_dir.replace("runs", "test") + '_ca'
working_dirs = [
os.path.basename(working_path)
for working_path in glob.glob(os.path.join(working_dir, '*'))
]
if not filter_dirs or os.path.basename(load_dir) not in working_dirs:
yield tmpl_str % locals()
if __name__ == '__main__':
layer_dims = '1200-1200-1200'
critic_layer_dims = '1200-1200'
runs = 10
if FLAGS.train:
generate_critic(pretrain_niter=1, niter=100,
model='mlp', critic_model='critic_mlp', activation_fn='relu',
name='critic', layer_dims=layer_dims, critic_layer_dims=critic_layer_dims,
lmbd=0.5, lmbd_grad=10.0, lmbd_rec_l1=0.0, lmbd_rec_l2=0.01,
lr=0.0005, lr_decay_step=40, lr_decay_factor=0.5,
critic_lr=0.001, critic_lr_decay_step=40, critic_lr_decay_factor=0.5,
attack_confidence="class_running_mean",
train_dir="runs_mlp",
attack_random=False, attack_uniform=False, attack_label_smoothing=0.0,
attack_overshoot=0.05, runs=runs, attack_iter=5, gpu_memory=0.8)
generate_critic(pretrain_niter=1, niter=100,
model='lenet5', critic_model='critic_mlp', activation_fn='relu',
name='critic', layer_dims=layer_dims, critic_layer_dims=critic_layer_dims,
lmbd=0.1, lmbd_grad=10.0, lmbd_rec_l1=0.0, lmbd_rec_l2=0.01,
lr=0.0005, lr_decay_step=40, lr_decay_factor=0.5,
critic_lr=0.001, critic_lr_decay_step=40, critic_lr_decay_factor=0.5,
attack_confidence="class_running_mean",
train_dir="runs_lenet5",
attack_random=False, attack_uniform=False, attack_label_smoothing=0.0,
attack_overshoot=0.05, runs=runs, attack_iter=5, gpu_memory=0.8)
else:
if not FLAGS.carlini:
generate_test('./runs_mlp/*', attack_iter=500,
attack_clip=0.1, dataset="test",
working_dir="test_mlp", attack_box_clip=True,
hc_confidence="same")
generate_test('./runs_lenet5/*', attack_iter=500,
attack_clip=0.1, dataset="test",
working_dir="test_lenet5", attack_box_clip=True,
hc_confidence="same")
else:
generate_test_carlini('./runs_mlp/*', working_dir='test_mlp_ca',
carlini_max_iter=1000)
generate_test_carlini('./runs_lenet5/*', working_dir='test_lenet5_ca',
carlini_max_iter=1000)