-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
245 lines (199 loc) · 10.6 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
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
import os
from absl import app
import numpy as np
import torch
import torch.backends.cudnn as cudnn
from torchvision import transforms
from tqdm import tqdm
from easydict import EasyDict
from cleverhans.torch.attacks.fast_gradient_method import fast_gradient_method
from cleverhans.torch.attacks.projected_gradient_descent import projected_gradient_descent
import simplejson
from datetime import datetime
from utils.data_loader import load_dataset
from utils.consts import Consts, ClassEncoder
from utils.early_stopping import EarlyStopping
from networks.vgg import VGG
from networks.resnet import ResNet18
from networks.resnext import ResNeXt29_2x64d
from networks.googlenet import GoogLeNet
from networks.efficientnet import EfficientNetB0
from networks.mobilenetv2 import MobileNetV2
from networks.shufflenetv2 import ShuffleNetV2
def train(net, optimizer, train_loader, epoch, consts: Consts, grayscale_converter):
""" Train model """
net.train()
train_loss = 0.0
num_correct, num_total = 0, 0
for x, y in tqdm(train_loader):
x, y = x.to(consts.DEVICE), y.to(consts.DEVICE)
if consts.ADV_TRAIN:
net.inference_mode()
# Replace clean example with adversarial example for adversarial training
# x = projected_gradient_descent(net, x, consts.EPS, 0.01, 40, np.inf)
x = fast_gradient_method(net, x, consts.EPS, np.inf)
net.non_inference_mode()
optimizer.zero_grad()
output_cls, output_aux = net(x)
if consts.AUXILIARY_TYPE == "Recon":
loss = net.loss(output_cls, y, output_aux, x, consts.AUXILIARY_TYPE, consts.AUXILIARY_WEIGHT)
elif consts.AUXILIARY_TYPE == "Fourier":
x_gray = grayscale_converter(x)
x_fourier_complex = torch.fft.fftshift(torch.fft.fft2(x_gray))
x_fourier_mag = torch.abs(x_fourier_complex)
x_fourier_ph = torch.angle(x_fourier_complex)
x_fourier = torch.cat((x_fourier_mag, x_fourier_ph), dim=1)
loss = net.loss(output_cls, y, output_aux, x_fourier, consts.AUXILIARY_TYPE, consts.AUXILIARY_WEIGHT)
else:
loss = net.loss(output_cls, y, None, None, consts.AUXILIARY_TYPE, None)
loss.backward()
optimizer.step()
train_loss += loss.item()
_, y_pred = output_cls.max(1) # model prediction on clean examples
num_correct += y_pred.eq(y).sum().item()
num_total += y.size(0)
train_acc = num_correct / num_total
train_loss_norm = train_loss / num_total
tqdm.write(
"epoch: {}/{}, train loss: {:.3f}, train acc ({}): {:.2f}%".format(
epoch, consts.NUM_EPOCHS, train_loss_norm, \
"adversarial data" if consts.ADV_TRAIN else "clean data", \
train_acc * 100.0
)
)
def validate(net, validation_loader, epoch, consts: Consts, grayscale_converter):
""" Validate model """
net.eval()
# Cannot simply use `with torch.no_grad():` here because adversarial generation requires gradient?
validation_loss = 0.0
num_correct, num_total = 0, 0
for x, y in tqdm(validation_loader):
x, y = x.to(consts.DEVICE), y.to(consts.DEVICE)
if consts.ADV_TRAIN:
net.inference_mode()
x = fast_gradient_method(net, x, consts.EPS, np.inf)
net.non_inference_mode()
output_cls, output_aux = net(x)
if consts.AUXILIARY_TYPE == "Recon":
loss = net.loss(output_cls, y, output_aux, x, consts.AUXILIARY_TYPE, consts.AUXILIARY_WEIGHT)
elif consts.AUXILIARY_TYPE == "Fourier":
x_gray = grayscale_converter(x)
x_fourier_complex = torch.fft.fftshift(torch.fft.fft2(x_gray))
x_fourier_mag = torch.abs(x_fourier_complex)
x_fourier_ph = torch.angle(x_fourier_complex)
x_fourier = torch.cat((x_fourier_mag, x_fourier_ph), dim=1)
loss = net.loss(output_cls, y, output_aux, x_fourier, consts.AUXILIARY_TYPE, consts.AUXILIARY_WEIGHT)
else:
loss = net.loss(output_cls, y, None, None, consts.AUXILIARY_TYPE, None)
validation_loss += loss.item()
_, y_pred = output_cls.max(1) # model prediction on clean examples
num_correct += y_pred.eq(y).sum().item()
num_total += y.size(0)
validation_acc = num_correct / num_total
validation_loss_norm = validation_loss / num_total
tqdm.write(
"epoch: {}/{}, validation loss: {:.3f}, validation acc ({}): {:.2f}%".format(
epoch, consts.NUM_EPOCHS, validation_loss_norm, \
"adversarial data" if consts.ADV_TRAIN else "clean data", \
validation_acc * 100.0
)
)
return validation_acc
def test(net, test_loader, consts: Consts):
""" Evaluate on clean and adversarial data """
print("\nEvaluating model on clean and adversarial data...")
net.eval()
net.inference_mode()
accuracies = {'clean': [], 'fgm': [], 'pgd': []}
for rep in range(consts.TEST_REPS):
report = EasyDict(nb_test=0, correct=0, correct_fgm=0, correct_pgd=0)
tqdm.write("repetition: {}/{}".format(rep + 1, consts.TEST_REPS))
for x, y in tqdm(test_loader):
x, y = x.to(consts.DEVICE), y.to(consts.DEVICE)
x_fgm = fast_gradient_method(net, x, consts.EPS, np.inf)
x_pgd = projected_gradient_descent(net, x, consts.EPS, 0.01, 40, np.inf)
_, y_pred = net(x).max(1) # model prediction on clean examples
_, y_pred_fgm = net(x_fgm).max(1) # model prediction on FGM adversarial examples
_, y_pred_pgd = net(x_pgd).max(1) # model prediction on PGD adversarial examples
report.nb_test += y.size(0) # This should be identical over repetitions
report.correct += y_pred.eq(y).sum().item() # This should be identical over repetitions
report.correct_fgm += y_pred_fgm.eq(y).sum().item() # This shall change over repetitions
report.correct_pgd += y_pred_pgd.eq(y).sum().item() # This shall change over repetitions
accuracies['clean'] += [report.correct / report.nb_test * 100.0]
accuracies['fgm'] += [report.correct_fgm / report.nb_test * 100.0]
accuracies['pgd'] += [report.correct_pgd / report.nb_test * 100.0]
acc_text = []
acc_text.append("test acc on clean examples: {:.3f}\u00B1{:.3f}%".format(
np.mean(accuracies['clean']), np.std(accuracies['clean'])
))
acc_text.append("test acc on FGM adversarial examples: {:.3f}\u00B1{:.3f}%".format(
np.mean(accuracies['fgm']), np.std(accuracies['fgm'])
))
acc_text.append("test acc on PGD adversarial examples: {:.3f}\u00B1{:.3f}%".format(
np.mean(accuracies['pgd']), np.std(accuracies['pgd'])
))
for acc in acc_text:
print(acc)
return acc_text
def main(_):
consts = Consts()
# 1 ################################ Book keeping jobs and logistics ################################
torch.manual_seed(consts.RANDOM_SEED)
data = load_dataset(consts.DATASET, consts.BATCH_SIZE, consts.NUM_WORKERS)
timestamp = datetime.now().strftime("%Y-%m-%d-%H-%M-%S")
weight_path = './saved/saved_weights/' + consts.DATASET + '_' + consts.MODEL + '_' + str(timestamp) + '_'
consts_path = './saved/saved_settings/' + consts.DATASET + '_' + consts.MODEL + '_' + str(timestamp) + '_consts.txt'
os.makedirs('/'.join(os.path.abspath(weight_path).split('/')[:-1]), exist_ok = True)
os.makedirs('/'.join(os.path.abspath(consts_path).split('/')[:-1]), exist_ok = True)
with open(consts_path, 'w') as file:
simplejson.dump(ClassEncoder().encode(consts), file)
# 2 ################# Instantiate model, loss, and optimizer for training ############################
if consts.AUXILIARY_TYPE == "Fourier":
decoded_channel = 2
else:
decoded_channel = 3
if consts.MODEL == "VGG11":
net = VGG(vgg_name='VGG11', img_dim=data.img_dim, decoded_channel=decoded_channel, num_classes=consts.NUM_CLASSES)
elif consts.MODEL == "ResNet18":
net = ResNet18(img_dim=data.img_dim, decoded_channel=decoded_channel, num_classes=consts.NUM_CLASSES)
elif consts.MODEL == "ResNeXt29_2x64d":
net = ResNeXt29_2x64d(img_dim=data.img_dim, decoded_channel=decoded_channel, num_classes=consts.NUM_CLASSES)
elif consts.MODEL == "GoogLeNet":
net = GoogLeNet(img_dim=data.img_dim, decoded_channel=decoded_channel, num_classes=consts.NUM_CLASSES)
elif consts.MODEL == "EfficientNetB0":
net = EfficientNetB0(img_dim=data.img_dim, decoded_channel=decoded_channel, num_classes=consts.NUM_CLASSES)
elif consts.MODEL == "MobileNetV2":
net = MobileNetV2(img_dim=data.img_dim, decoded_channel=decoded_channel, num_classes=consts.NUM_CLASSES)
elif consts.MODEL == "ShuffleNetV2":
net = ShuffleNetV2(net_size=0.5, img_dim=data.img_dim, decoded_channel=decoded_channel, num_classes=consts.NUM_CLASSES)
net = net.to(consts.DEVICE)
if consts.DEVICE == "cuda":
cudnn.benchmark = True
if consts.OPTIMIZER == "SGD":
optimizer = torch.optim.SGD(net.parameters(), lr=consts.LEARNING_RATE, momentum=consts.MOMENTUM, weight_decay=consts.WEIGHT_DECAY)
elif consts.OPTIMIZER == "Adam":
optimizer = torch.optim.Adam(net.parameters(), lr=consts.LEARNING_RATE)
# scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=200)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='max', factor=0.2, verbose=True, patience=5)
early_stopping = EarlyStopping(criterion_name='Validation Accuracy', mode='max',
patience=10, verbose=True,
path=weight_path+'best_model.pt')
# This transform is a grayscale conversion. Only used before Fourier Transform.
grayscale_converter = transforms.Grayscale()
# 3 ################################### Train, validate and Test #####################################
# torch.autograd.set_detect_anomaly(True)
for epoch in range(1, consts.NUM_EPOCHS + 1):
train(net, optimizer, data.train, epoch, consts, grayscale_converter)
validation_acc = validate(net, data.validation, epoch, consts, grayscale_converter)
# scheduler.step()
scheduler.step(validation_acc)
early_stopping(validation_acc, net)
if early_stopping.early_stop:
print("Early stopping!")
break
accuracies = test(net, data.test, consts)
with open(consts_path, 'a') as file:
for acc in accuracies:
file.write('\n' + acc)
if __name__ == "__main__":
app.run(main)