-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
209 lines (188 loc) · 8.84 KB
/
main.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
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
"""main file to run the lda approximation"""
# -*- coding: utf-8 -*-
# !/usr/bin/env python3
import time
import socket
import logging
import datetime
import argparse
import os
import torch
import numpy as np
from utils.words import save_train_data
from utils.lda import train_lda
from utils.train import train
from utils.eval import evaluate, topic_stacking_attack
torch.backends.cudnn.benchmark = True
def main(gpu: int, num_workers: int, num_topics: int, from_scratch: bool, learning_rate: float,
epochs: int, batch_size: int, verbose: bool, attack_id: int, random_test: bool,
advs_eps: float, l2_attack: bool, max_iteration: int, prob_attack: bool,
full_attack: bool, topic_stacking: bool) -> None:
"""main function"""
start = time.perf_counter()
if verbose:
# logging for gensim output
logging.basicConfig(format='%(asctime)s : %(levelname)s : %(message)s', level=logging.INFO)
# set devices properly
if gpu == 0:
device = 'cuda:0' if torch.cuda.is_available() else 'cpu'
if gpu == 1:
device = 'cuda:1' if torch.cuda.is_available() else 'cpu'
# set model paths
# lda_path = "./models/freq_lda_model" if freq_id else "./models/lda_model"
# data_path = "./data/wiki_data_freq.tar" if freq_id else "./data/wiki_data.tar"
# dnn_path = "./models/dnn_model_freq" if freq_id else "./models/dnn_model"
lda_path = "./models/lda_model"
data_path = "./data/wiki_data.tar"
dnn_path = "./models/dnn_model"
# print a summary of the chosen arguments
print("\n\n\n"+"#"*50)
print("## " + str(datetime.datetime.now().strftime("%A, %d. %B %Y %I:%M%p")))
print(f"## System: {torch.get_num_threads()} CPU cores with "
f"{os.cpu_count()} threads and "
f"{torch.cuda.device_count()} GPUs on {socket.gethostname()}")
if device == 'cpu':
print("## Using: CPU with ID {}".format(device))
else:
print("## Using: {} with ID {}".format(torch.cuda.get_device_name(device=device), device))
print("## Using {} workers for LDA computation".format(num_workers))
print("## Num_topics: {}".format(num_topics))
print("## Learning_rate: {}".format(learning_rate))
print("## Batch_size: {}".format(batch_size))
print("## Epochs: {}".format(epochs))
if bool(attack_id) or full_attack:
if not prob_attack:
print("## Target Word ID: {}".format(attack_id if not full_attack else "Full Attack"))
else:
print("## Target: Whole Distribution")
print("## Advs. Epsilon: {}".format(advs_eps))
if l2_attack:
print("## Attack mode: L2 (rounded floats)")
else:
print("## Attack mode: LINF (integers)")
print("## Random Test: {}".format(random_test))
print("#"*50)
print("\n\n")
if not os.path.isfile(lda_path):
# obtain a preprocessed list of words
train_lda(num_workers, num_topics, None)
elif from_scratch:
# obtain a preprocessed list of words
train_lda(num_workers, num_topics, None)
elif not bool(attack_id) and not full_attack and not prob_attack and not topic_stacking:
print("[ a trained LDA model already exists. Train again? [y/n] ]")
if from_scratch or input() == "y":
# obtain a preprocessed list of words
train_lda(num_workers, num_topics, None)
if not os.path.isfile(data_path):
# save the lda model data as training data with labels
save_train_data(freq_id=None)
elif from_scratch:
# save the lda model data as training data with labels
save_train_data(freq_id=None)
elif not bool(attack_id) and not full_attack and not prob_attack and not topic_stacking:
print("[ training data/labels already exists. Save them again? [y/n] ]")
if from_scratch or input() == "y":
# save the lda model data as training data with labels
save_train_data(freq_id=None)
if not os.path.isfile(dnn_path):
# train the DNN model on the lda dataset
train(epochs=epochs,
learning_rate=learning_rate,
batch_size=batch_size,
num_topics=num_topics,
device_name=device,
model_path=dnn_path,
freq_id=None,
verbose=verbose)
elif from_scratch:
# train the DNN model on the lda dataset
train(epochs=epochs,
learning_rate=learning_rate,
batch_size=batch_size,
num_topics=num_topics,
device_name=device,
model_path=dnn_path,
freq_id=None,
verbose=verbose)
elif not bool(attack_id) and not full_attack and not prob_attack and not topic_stacking:
print("[ a trained DNN model already exists. Train again? [y/n] ]")
if from_scratch or input() == "y":
# train the DNN model on the lda dataset
train(epochs=epochs,
learning_rate=learning_rate,
batch_size=batch_size,
num_topics=num_topics,
device_name=device,
model_path=dnn_path,
freq_id=None,
verbose=verbose)
# evaluate both the lda and the dnn model and print their top topics
if full_attack:
total_success = 0
successful_topics = []
unsuccessful_topics = []
for topic_target in range(num_topics):
success_flag = evaluate(num_topics,
topic_target,
random_test,
advs_eps,
device,
l2_attack,
max_iteration,
prob_attack)
if success_flag:
total_success += 1
successful_topics.append(topic_target)
else:
unsuccessful_topics.append(topic_target)
print("\n-> {} / {} attacks successful!".format(total_success, num_topics))
print("successful topics: {}".format(successful_topics))
print("unsuccessful topics: {}".format(unsuccessful_topics))
elif topic_stacking:
topic_stacking_attack(device,
advs_eps,
num_topics,
max_iteration)
else:
success_flag = evaluate(num_topics,
attack_id,
random_test,
advs_eps,
device,
l2_attack,
max_iteration,
prob_attack)
end = time.perf_counter()
duration = (np.round(end - start) / 60.) / 60.
print(f"\nComputation time: {duration:0.4f} hours")
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--gpu", "-g", help="GPU", type=int, default=0)
parser.add_argument("--attack_id", "-a", help="id of the target word", type=int, default=None)
parser.add_argument("--advs_eps", "-ae", help="epsilon for the adversarial attack",
type=float, default=100)
parser.add_argument("--batch_size", "-b", help="batch size", type=int, default=512)
parser.add_argument("--epochs", "-e", help="training epochs", type=int, default=100)
parser.add_argument("--max_iteration", "-mi", help="max. attack iters", type=int, default=200)
parser.add_argument("--learning_rate", "-l", help="learning rate", type=float, default=0.01)
parser.add_argument("--num_workers", "-w", help="number of workers for lda",
type=int, default=8)
parser.add_argument("--num_topics", "-t", help="number of topics for lda",
type=int, default=50)
parser.add_argument("--from_scratch", "-s", help="train lda from scratch",
action='store_true', default=False)
parser.add_argument("--random_test", "-r", help="enable random test documents",
action='store_true', default=False)
parser.add_argument("--verbose", "-v", help="set gensim to verbose mode",
action='store_true', default=False)
parser.add_argument("--prob_attack", "-pa", help="try to use a whole distribution as target",
action='store_true', default=False)
parser.add_argument("--l2_attack", "-l2", help="set attack to l2 mode",
action='store_true', default=False)
parser.add_argument("--full_attack", "-f", help="perform an attack on every topic",
action='store_true', default=False)
parser.add_argument("--topic_stacking", "-ts", help="performs topic stacking method",
action='store_true', default=False)
args = parser.parse_args()
main(**vars(args))