code_level_0 = 0.25 * np.asarray(code[0]).flatten()
code_level_1 = 0.25 * np.asarray(code[1:5]).flatten()
code_level_2 = 0.5 * np.asarray(code[5:]).flatten()
return np.concatenate((code_level_0, code_level_1, code_level_2))
VOC_SIZE = 100
PYRAMID_LEVEL = 1
DSIFT_STEP_SIZE = 4
# DSIFT_STEP_SIZE is related to the function
# extract_DenseSift_descriptors in utils.py
# and build_spatial_pyramid in spm.py
if __name__ == '__main__':
x_train, y_train = load_cifar10_data(dataset='train')
x_test, y_test = load_cifar10_data(dataset='test')
print("Dense SIFT feature extraction")
x_train_feature = [extract_DenseSift_descriptors(img) for img in x_train]
x_test_feature = [extract_DenseSift_descriptors(img) for img in x_test]
x_train_kp, x_train_des = zip(*x_train_feature)
x_test_kp, x_test_des = zip(*x_test_feature)
print("Train/Test split: {:d}/{:d}".format(len(y_train), len(y_test)))
print("Codebook Size: {:d}".format(VOC_SIZE))
print("Pyramid level: {:d}".format(PYRAMID_LEVEL))
print("Building the codebook, it will take some time")
codebook = build_codebook(x_train_des, VOC_SIZE)
# import cPickle
#
from mlp_hn_clf import HighwayClassifier
import numpy as np
import tensorflow as tf
import utils
if __name__ == '__main__':
(X_train, y_train), (X_test, y_test) = utils.load_cifar10_data()
X_train = (X_train / 255.0).mean(axis=3).reshape(-1, 32 * 32)
X_test = (X_test / 255.0).mean(axis=3).reshape(-1, 32 * 32)
y_train = y_train.ravel()
y_test = y_test.ravel()
clf = HighwayClassifier(32 * 32, 10, 20)
log = clf.fit(X_train,
y_train,
n_epoch=30,
en_exp_decay=False,
val_data=(X_test, y_test))
pred = clf.predict(X_test)
final_acc = (pred == y_test).mean()
print("final testing accuracy: %.4f" % final_acc)
def attack_bayes():
"""
Perform BayesOpt attack
"""
# get dataset and list of indices of images to attack
if args.dset == 'mnist':
test_dataset = load_mnist_data()
samples = np.arange(args.start, args.start + args.num_attacks)
elif args.dset == 'cifar10':
test_dataset = load_cifar10_data()
samples = np.arange(args.start, args.start + args.num_attacks)
elif args.dset == 'imagenet':
if args.arch == 'inception_v3':
test_dataset = load_imagenet_data(299, 299)
else:
test_dataset = load_imagenet_data()
# see readme
samples = np.load('random_indices_imagenet.npy')
samples = samples[args.start:args.start + args.num_attacks]
print("Length of sample_set: ", len(samples))
results_dict = {}
# loop over images, attacking each one if it is initially correctly classified
for idx in samples[:args.num_attacks]:
image, label = test_dataset[idx]
image = image.unsqueeze(0).to(device)
print(f"Image {idx:d} Original label: {label:d}")
predicted_label = torch.argmax(cnn_model.predict_scores(image))
print("Predicted label: ", predicted_label.item())
# ignore incorrectly classified images
if label == predicted_label:
# itr, success = bayes_opt(image, label)
retry = 0
while retry < 5:
try:
itr, success = bayes_opt(image, label)
break
except:
print('Retry_{}'.format(retry))
retry += 1
print(itr, success)
if success:
results_dict[idx] = itr
else:
results_dict[idx] = 0
sys.stdout.flush()
# results saved as dictionary, with entries of the form
# dataset idx : 0 if unsuccessfully attacked, # of queries if successfully attacked
print('RESULTS', results_dict)
if args.save:
pickle.dump(
results_dict,
open(
f"{args.dset:s}{args.arch:s}_{args.start:d}_{args.iter:d}_{args.dim:d}_{args.eps:.2f}_{args.num_attacks:d}.pkl",
'wb'))
from utils import extract_sift_descriptors
from utils import build_codebook
from utils import input_vector_encoder
from classifier import svm_classifier
import numpy as np
VOC_SIZE = 100
if __name__ == '__main__':
# Training
x_train, y_train = load_cifar10_data(dataset='train')
x_test, y_test = load_cifar10_data(dataset='test')
print "SIFT feature extraction"
x_train = [extract_sift_descriptors(img) for img in x_train]
x_test = [extract_sift_descriptors(img) for img in x_test]
# Remove None in SIFT extraction
x_train = [each for each in zip(x_train, y_train) if each[0] != None]
x_train, y_train = zip(*x_train)
x_test = [each for each in zip(x_test, y_test) if each[0] != None]
x_test, y_test = zip(*x_test)
print "Train/Test split: {:d}/{:d}".format(len(y_train), len(y_test))
print "Codebook Size: {:d}".format(VOC_SIZE)