def Create_X(pathTo_ILSVRC2012, mode):
validate_arguments(mode)
un_zip(os.path.join(pathTo_ILSVRC2012, 'ILSVRC2012_img_train'))
datafile = os.path.join('data', 'imagenet_data_' + str(mode) + '_mode.npy')
print('>> Creating pre-processed imagenet data...')
X = create_imagenet_npy(Pos_unzip, mode=mode)
print('>> Saving the pre-processed imagenet data')
# Save the pre-processed images
# Caution: This can take take a lot of space. Comment this part to discard saving.
np.save(datafile, X)
return X
print('>> Computing gradient function...')
def grad_fs(image_inp, indices):
return persisted_sess.run(dydx,
feed_dict={
persisted_input: image_inp,
inds: indices
}).squeeze(axis=1)
if os.path.isfile(file_perturbation) == 0:
# Load/Create data
datafile = os.path.join('data', npy_data)
if os.path.isfile(datafile) == 0:
print('>> Creating pre-processed imagenet data...')
X = create_imagenet_npy(path_train_imagenet)
print('>> Saving the pre-processed imagenet data')
if not os.path.exists('data'):
os.makedirs('data')
# Save the pre-processed images
# Caution: This can take take a lot of space. Comment this part to discard saving.
np.save(os.path.join('data', npy_data), X)
else:
print('>> Pre-processed imagenet data detected')
X = np.load(datafile)
# Running universal perturbation
for x in range(995):
v = targeted_perturbation(X,
def f(image_inp): return persisted_sess.run(persisted_output,
feed_dict={persisted_input: np.reshape(image_inp, (-1, 224, 224, 3))})
print('>> Compiling the gradient tensorflow functions. This might take some time...')
y_flat = tf.reshape(persisted_output, (-1,))
inds = tf.placeholder(tf.int32, shape=(2,))
dydx = jacobian(y_flat, persisted_input, inds)
print('>> Computing gradient function...')
def grad_fs(image_inp, indices): return persisted_sess.run(dydx, feed_dict={persisted_input: image_inp,
inds: indices}).squeeze(axis=1)
for i in range(100):
npy_img = 'data/npy_img/10classes1000imgs/' + str(i) +'.npy'
npy_per = 'data/npy_per/10classes1000imgs/' + str(i) + '.npy'
if os.path.isfile(npy_img) == 0:
X = create_imagenet_npy(path_train_imagenet, len_batch=1100, num_class=10, p=0, q=0, r=i*10)
np.save(npy_img, X)
else:
X = np.load(npy_img)
if os.path.isfile(npy_per) == 0:
v = targeted_perturbation(X, f, grad_fs, delta=0.25, max_iter_uni=10, target=target)
np.save((npy_per), v)
print(">> Computing gradient function...")
def grad_fs(image_inp, inds):
return [
persisted_sess.run(dydx[i],
feed_dict={persisted_input: image_inp})
for i in inds
]
# Load/Create data
datafile = os.path.join('data', 'imagenet_data.npy')
if os.path.isfile(datafile) == 0:
print(">> Creating pre-processed imagenet data...")
#预处理一个batch的图片并返回
X, dirs, sub_dirs = create_imagenet_npy(PATH_TRAIN_IMAGENET, 1000)
print(X.shape)
# print(">> Saving the pre-processed imagenet data")
# if not os.path.exists('data'):
# os.makedirs('data')
# Save the pre-processed images
# Caution: This can take take a lot of space. Comment this part to discard saving.
# np.save(os.path.join('data', 'imagenet_data.npy'), X)
else:
print(">> Pre-processed imagenet data detected")
X = np.load(datafile)
#核心部分！！计算通用干扰