def main():
if os.path.exists(directory):
shutil.rmtree(directory)
if not os.path.exists(directory):
os.makedirs(directory)
# device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
cnn = CNN()
# cnn.to(device)
cnn.eval()
cnn.load_state_dict(torch.load('model.pkl'))
print("Model Loaded")
criterion = nn.MultiLabelSoftMarginLoss()
data_loader = setting.get_test_data_loader()
correct = 0
total = 0
count = 1
for i, (images, labels) in enumerate(data_loader):
image = images
variable = Variable(image)
# variable = Variable(image).to(device)
variable.requires_grad = True
predict_label = cnn(variable)
# predict_label = cnn(variable).to(device)
loss = criterion(predict_label, labels.float())
cnn.zero_grad()
loss.backward()
data_grad = variable.grad.data
perturbed_image = fgsm_attack(variable, epsilon, data_grad)
predict_label = cnn(perturbed_image)
numpy_image = perturbed_image.squeeze().detach().cpu().numpy()
plt.imshow(numpy_image)
fig = plt.gcf()
fig.savefig(directory + '/' + str(count) + '.png')
count += 1
c0 = setting.character_set[np.argmax(predict_label[0, 0:setting.character_set_length].data.numpy())]
c1 = setting.character_set[np.argmax(predict_label[0, setting.character_set_length:2 * setting.character_set_length].data.numpy())]
c2 = setting.character_set[np.argmax(predict_label[0, 2 * setting.character_set_length:3 * setting.character_set_length].data.numpy())]
c3 = setting.character_set[np.argmax(predict_label[0, 3 * setting.character_set_length:4 * setting.character_set_length].data.numpy())]
predict_label = '%s%s%s%s' % (c0, c1, c2, c3)
true_label = setting.decode(labels.numpy()[0])
total += labels.size(0)
if predict_label == true_label:
correct += 1
if total % 20 == 0:
print('Test Accuracy of the model on the %d test images: %f %%' % (total, 100 * correct / total))
print('Test Accuracy of the model on the %d test images: %f %%' % (total, 100 * correct / total))
def main():
cnn = CNN()
cnn.eval()
cnn.load_state_dict(torch.load('model1.pkl'))
print("load cnn net.")
test_dataloader = my_dataset.get_test_data_loader()
correct = 0
total = 0
epsilon = float(input("please input epsilon = "))
criterion = nn.MultiLabelSoftMarginLoss()
for i, (images, labels) in enumerate(test_dataloader):
vimage = Variable(images)
labels = Variable(labels.float())
vimage.requires_grad = True
predict_labels = cnn(vimage)
loss = criterion(predict_labels, labels)
cnn.zero_grad()
loss.backward()
data_grad = vimage.grad.data
perturbed_data = fgsm_attack(vimage, epsilon, data_grad)
perturbed_label = cnn(perturbed_data)
c0 = captcha_setting.ALL_CHAR_SET[np.argmax(
perturbed_label[0,
0:captcha_setting.ALL_CHAR_SET_LEN].data.numpy())]
c1 = captcha_setting.ALL_CHAR_SET[np.argmax(
perturbed_label[0, captcha_setting.ALL_CHAR_SET_LEN:2 *
captcha_setting.ALL_CHAR_SET_LEN].data.numpy())]
c2 = captcha_setting.ALL_CHAR_SET[np.argmax(
perturbed_label[0, 2 * captcha_setting.ALL_CHAR_SET_LEN:3 *
captcha_setting.ALL_CHAR_SET_LEN].data.numpy())]
c3 = captcha_setting.ALL_CHAR_SET[np.argmax(
perturbed_label[0, 3 * captcha_setting.ALL_CHAR_SET_LEN:4 *
captcha_setting.ALL_CHAR_SET_LEN].data.numpy())]
perturbed_label = '%s%s%s%s' % (c0, c1, c2, c3)
true_label = one_hot_encoding.decode(labels.numpy()[0])
total += labels.size(0)
if (perturbed_label == true_label):
correct += 1
if (total % 100 == 0):
print('Test Accuracy of the model on the %d test images: %f %%' %
(total, 100 * correct / total))
print('Test Accuracy of the model on the %d test images: %f %%' %
(total, 100 * correct / total))
