-
Notifications
You must be signed in to change notification settings - Fork 13
/
prune_neuron_cifar.py
173 lines (148 loc) · 8.15 KB
/
prune_neuron_cifar.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
import os
import argparse
import numpy as np
import pandas as pd
import torch
from torch.utils.data import DataLoader
from torchvision.datasets import CIFAR10
import torchvision.transforms as transforms
import models
import data.poison_cifar as poison
parser = argparse.ArgumentParser(description='Train poisoned networks')
# Basic model parameters.
parser.add_argument('--arch', type=str, default='resnet18',
choices=['resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152', 'MobileNetV2', 'vgg19_bn'])
parser.add_argument('--checkpoint', type=str, required=True, help='The checkpoint to be pruned')
parser.add_argument('--widen-factor', type=int, default=1, help='widen_factor for WideResNet')
parser.add_argument('--batch-size', type=int, default=128, help='the batch size for dataloader')
parser.add_argument('--data-dir', type=str, default='../data', help='dir to the dataset')
parser.add_argument('--output-dir', type=str, default='logs/models/')
parser.add_argument('--trigger-info', type=str, default='', help='The information of backdoor trigger')
parser.add_argument('--poison-type', type=str, default='benign', choices=['badnets', 'blend', 'clean-label', 'benign'],
help='type of backdoor attacks for evaluation')
parser.add_argument('--poison-target', type=int, default=0, help='target class of backdoor attack')
parser.add_argument('--trigger-alpha', type=float, default=1.0, help='the transparency of the trigger pattern.')
parser.add_argument('--mask-file', type=str, required=True, help='The text file containing the mask values')
parser.add_argument('--pruning-by', type=str, default='threshold', choices=['number', 'threshold'])
parser.add_argument('--pruning-max', type=float, default=0.90, help='the maximum number/threshold for pruning')
parser.add_argument('--pruning-step', type=float, default=0.05, help='the step size for evaluating the pruning')
args = parser.parse_args()
args_dict = vars(args)
print(args_dict)
os.makedirs(args.output_dir, exist_ok=True)
device = 'cuda' if torch.cuda.is_available() else 'cpu'
def main():
MEAN_CIFAR10 = (0.4914, 0.4822, 0.4465)
STD_CIFAR10 = (0.2023, 0.1994, 0.2010)
transform_test = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(MEAN_CIFAR10, STD_CIFAR10)
])
# Step 1: create poisoned / clean test set
if args.trigger_info:
trigger_info = torch.load(args.trigger_info, map_location=device)
else:
if args.poison_type == 'benign':
trigger_info = None
else:
triggers = {'badnets': 'checkerboard_1corner',
'clean-label': 'checkerboard_4corner',
'blend': 'gaussian_noise'}
trigger_type = triggers[args.poison_type]
pattern, mask = poison.generate_trigger(trigger_type=trigger_type)
trigger_info = {'trigger_pattern': pattern[np.newaxis, :, :, :], 'trigger_mask': mask[np.newaxis, :, :, :],
'trigger_alpha': args.trigger_alpha, 'poison_target': np.array([args.poison_target])}
clean_test = CIFAR10(root=args.data_dir, train=False, download=True, transform=transform_test)
poison_test = poison.add_predefined_trigger_cifar(data_set=clean_test, trigger_info=trigger_info)
poison_test_loader = DataLoader(poison_test, batch_size=args.batch_size, num_workers=0)
clean_test_loader = DataLoader(clean_test, batch_size=args.batch_size, num_workers=0)
# Step 2: load model checkpoints and trigger info
net = getattr(models, args.arch)(num_classes=10)
net.load_state_dict(torch.load(args.checkpoint, map_location=device))
net = net.to(device)
criterion = torch.nn.CrossEntropyLoss().to(device)
# Step 3: pruning
mask_values = read_data(args.mask_file)
mask_values = sorted(mask_values, key=lambda x: float(x[2]))
print('No. \t Layer Name \t Neuron Idx \t Mask \t PoisonLoss \t PoisonACC \t CleanLoss \t CleanACC')
cl_loss, cl_acc = test(model=net, criterion=criterion, data_loader=clean_test_loader)
po_loss, po_acc = test(model=net, criterion=criterion, data_loader=poison_test_loader)
print('0 \t None \t None \t {:.4f} \t {:.4f} \t {:.4f} \t {:.4f}'.format(po_loss, po_acc, cl_loss, cl_acc))
if args.pruning_by == 'threshold':
results = evaluate_by_threshold(
net, mask_values, pruning_max=args.pruning_max, pruning_step=args.pruning_step,
criterion=criterion, clean_loader=clean_test_loader, poison_loader=poison_test_loader
)
else:
results = evaluate_by_number(
net, mask_values, pruning_max=args.pruning_max, pruning_step=args.pruning_step,
criterion=criterion, clean_loader=clean_test_loader, poison_loader=poison_test_loader
)
file_name = os.path.join(args.output_dir, 'pruning_by_{}.txt'.format(args.pruning_by))
with open(file_name, "w") as f:
f.write('No \t Layer Name \t Neuron Idx \t Mask \t PoisonLoss \t PoisonACC \t CleanLoss \t CleanACC\n')
f.writelines(results)
def read_data(file_name):
tempt = pd.read_csv(file_name, sep='\s+', skiprows=1, header=None)
layer = tempt.iloc[:, 1]
idx = tempt.iloc[:, 2]
value = tempt.iloc[:, 3]
mask_values = list(zip(layer, idx, value))
return mask_values
def pruning(net, neuron):
state_dict = net.state_dict()
weight_name = '{}.{}'.format(neuron[0], 'weight')
state_dict[weight_name][int(neuron[1])] = 0.0
net.load_state_dict(state_dict)
def evaluate_by_number(model, mask_values, pruning_max, pruning_step, criterion, clean_loader, poison_loader):
results = []
nb_max = int(np.ceil(pruning_max))
nb_step = int(np.ceil(pruning_step))
for start in range(0, nb_max + 1, nb_step):
i = start
for i in range(start, start + nb_step):
pruning(model, mask_values[i])
layer_name, neuron_idx, value = mask_values[i][0], mask_values[i][1], mask_values[i][2]
cl_loss, cl_acc = test(model=model, criterion=criterion, data_loader=clean_loader)
po_loss, po_acc = test(model=model, criterion=criterion, data_loader=poison_loader)
print('{} \t {} \t {} \t {} \t {:.4f} \t {:.4f} \t {:.4f} \t {:.4f}'.format(
i+1, layer_name, neuron_idx, value, po_loss, po_acc, cl_loss, cl_acc))
results.append('{} \t {} \t {} \t {} \t {:.4f} \t {:.4f} \t {:.4f} \t {:.4f}'.format(
i+1, layer_name, neuron_idx, value, po_loss, po_acc, cl_loss, cl_acc))
return results
def evaluate_by_threshold(model, mask_values, pruning_max, pruning_step, criterion, clean_loader, poison_loader):
results = []
thresholds = np.arange(0, pruning_max + pruning_step, pruning_step)
start = 0
for threshold in thresholds:
idx = start
for idx in range(start, len(mask_values)):
if float(mask_values[idx][2]) <= threshold:
pruning(model, mask_values[idx])
start += 1
else:
break
layer_name, neuron_idx, value = mask_values[idx][0], mask_values[idx][1], mask_values[idx][2]
cl_loss, cl_acc = test(model=model, criterion=criterion, data_loader=clean_loader)
po_loss, po_acc = test(model=model, criterion=criterion, data_loader=poison_loader)
print('{:.2f} \t {} \t {} \t {} \t {:.4f} \t {:.4f} \t {:.4f} \t {:.4f}'.format(
start, layer_name, neuron_idx, threshold, po_loss, po_acc, cl_loss, cl_acc))
results.append('{:.2f} \t {} \t {} \t {} \t {:.4f} \t {:.4f} \t {:.4f} \t {:.4f}\n'.format(
start, layer_name, neuron_idx, threshold, po_loss, po_acc, cl_loss, cl_acc))
return results
def test(model, criterion, data_loader):
model.eval()
total_correct = 0
total_loss = 0.0
with torch.no_grad():
for i, (images, labels) in enumerate(data_loader):
images, labels = images.to(device), labels.to(device)
output = model(images)
total_loss += criterion(output, labels).item()
pred = output.data.max(1)[1]
total_correct += pred.eq(labels.data.view_as(pred)).sum()
loss = total_loss / len(data_loader)
acc = float(total_correct) / len(data_loader.dataset)
return loss, acc
if __name__ == '__main__':
main()