-
Notifications
You must be signed in to change notification settings - Fork 0
/
main_lippa_cnn_from_mmr_paper.py
135 lines (109 loc) · 5.36 KB
/
main_lippa_cnn_from_mmr_paper.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
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, transforms
import argparse
from gurobipy import *
import matplotlib.pyplot as plt
from temp_utils import *
from fgsm_and_lp_utils import *
import time
from datetime import datetime
import csv
from cleverhans.future.torch.attacks.projected_gradient_descent import projected_gradient_descent as pgd
from cnn_utils import *
import os
os.environ["CUDA_VISIBLE_DEVICES"]="-1" # trying to force the code to use cpu to see if we can gain time
import sys
model_pth_path = sys.argv[1]
csv_file_path = sys.argv[2]
EPS_H = float(sys.argv[3])
# python3 main_lippa_mlp_c.py 'model_pth_path' 'csv_file_path' EPS_H
######################## loading the data
import sys; sys.argv=['']; del sys # added to solve a problem with argparse and ipython
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument('--batch-size', type=int, default=64, metavar='N',
help='input batch size for training (default: 64)')
parser.add_argument('--test-batch-size', type=int, default=1024, metavar='N',
help='input batch size for testing (default: 1024)')
parser.add_argument('--epochs', type=int, default=100, metavar='N',
help='number of epochs to train (default: 100)')
parser.add_argument('--lr', type=float, default=0.01, metavar='LR',
help='learning rate (default: 0.01)')
parser.add_argument('--momentum', type=float, default=0.5, metavar='M',
help='SGD momentum (default: 0.5)')
parser.add_argument('--no-cuda', action='store_true', default=False,
help='disables CUDA training')
parser.add_argument('--seed', type=int, default=1, metavar='S',
help='random seed (default: 1)')
parser.add_argument('--log-interval', type=int, default=10, metavar='N',
help='how many batches to wait before logging training status')
parser.add_argument('--threshold', type=float, default=0.01, metavar='N',
help='minimum training error to stop epochs')
parser.add_argument('--save-model', action='store_true', default=False,
help='For Saving the current Model')
args = parser.parse_args()
use_cuda = not args.no_cuda and torch.cuda.is_available()
device = torch.device('cpu') #torch.device("cuda" if use_cuda else "cpu")
print(device)
kwargs = {'pin_memory': True} if use_cuda else {}
train_loader = torch.utils.data.DataLoader(
datasets.MNIST('../data', train=True, download=True,
transform=transforms.Compose([
transforms.ToTensor()
])),
batch_size=args.batch_size, shuffle=True, **kwargs)
test_loader = torch.utils.data.DataLoader(
datasets.MNIST('../data', train=False, transform=transforms.Compose([
transforms.ToTensor()
])),
batch_size=args.test_batch_size, shuffle=False, **kwargs)
cnn = CNN_from_mat()
cnn.load_state_dict(torch.load(model_pth_path,map_location=device)) # load the net
net = cnn
##################################################################
column_name = ["sn_vector" , "eps", "true_label", "y_2nd", "adv_pred","runtime" ]
with open(csv_file_path, mode = 'w') as csv_file:
csv_writer = csv.writer(csv_file, delimiter=';')
line_count = 0
csv_writer.writerow(column_name)
x_adv_dict = {}
# construction of the root problem
net_params = net.state_dict()
net_dict = { key : value.cpu().numpy() for key,value in zip(net_params.keys(),net_params.values()) }
net_dict['conv1.stride'] = 2 #net.conv1.stride[0]
net_dict['conv1.padding'] = 1 #net.conv2.padding[0]
net_dict['conv2.stride'] = 2
net_dict['conv2.padding']=1
root_advex_finder = lp_constructor_cnn(net_dict)
for sn in range(1000):
sample = test_loader.dataset[sn][0].to(device)#.cpu().numpy()
true_label = torch.tensor(test_loader.dataset[sn][1]).to(device)#.cpu().numpy() # add torch.tensor to prevent an error occuring on another torch version
net.to(device)
sample = sample.view(1,1,28,28)
prob, y_10, z_hat = net(sample)
y_torch, y_2nd_torch = torch.argsort(y_10[0],descending=True)[:2]
y_np = np.int(y_torch.cpu().numpy())
y_2nd_np = np.int(y_2nd_torch.cpu().numpy())
#print("iteration ", sn )
if (sn+1)%100 == 0 :
print("checkpoint at sample", sn )
csv_file.flush()
if (true_label!=y_torch) : # we are not interesting in examples that are already misclassified by the classifier
continue
pred_dict = {'y_np' : y_np, 'y_2nd_np':y_2nd_np}
tic = time.time()
eps, x_adv = lippa(sample, root_advex_finder,lp_cnn, net,pred_dict, eps_fgsm=0.5/256 ,tolerance=0.01, optim_time = 3.0, device = 'cpu',verbose = False, eps_h = EPS_H)
runtime = time.time()-tic
adv_pred = None
if eps is not None :
x_adv = x_adv.view(1,1,28,28)
_, adv_logit,_ = net(x_adv)
adv_pred = torch.argmax(adv_logit[0]).cpu().numpy()
x_adv_dict[sn] = x_adv
line_to_write = [sn, eps, y_np, y_2nd_np,adv_pred, runtime]
csv_writer.writerow(line_to_write)
line_count += 1
np.save(csv_file_path,x_adv_dict)