def _init_session_manager(self, session_manager=None):
if session_manager is None:
self._session_manager = SessionManager(
local_init_op=self._local_init_op,
ready_op=self._ready_op, graph=self._graph,
recovery_wait_secs=self._recovery_wait_secs)
else:
self._session_manager = session_manager
def foolbox_generate_adversarial_example(
label: int,
create_model,
input_fn: Callable[[], tf.Tensor],
attack_fn: Callable[..., Attack],
model_dir=None,
checkpoint_path=None,
preprocessing=(0, 1),
channel_axis=1,
bounds=(0, 1),
attack_params={},
**kwargs,
) -> Optional[np.ndarray]:
# Check that model has been trained.
if not checkpoint_path:
checkpoint_path = saver.latest_checkpoint(model_dir)
if not checkpoint_path:
raise ValueError(
"Could not find trained model in model_dir: {}.".format(model_dir))
with tf.Graph().as_default():
features = input_fn()
model = create_model()
image_tensor = tf.placeholder(features.dtype, features.shape)
logits = model(image_tensor)
sm = SessionManager()
with sm.prepare_session(
master="",
saver=tf.train.Saver(),
checkpoint_filename_with_path=checkpoint_path,
config=new_session_config(),
) as sess:
image = sess.run(features)[0]
attack_model = TensorFlowModel(
image_tensor,
logits,
bounds=bounds,
channel_axis=channel_axis,
preprocessing=preprocessing,
)
if attack_fn is None:
return image[np.newaxis]
attack = attack_fn(attack_model)
# adversarial_example = attack(image, label=label, **kwargs)
adversarial_example = attack(image, label=label, **attack_params)
# diff = np.abs(image - adversarial_example)
# print(diff.max()*255)
if adversarial_example is None:
return None
else:
return adversarial_example[np.newaxis]
def generate_adversarial_example(
label: int,
create_model,
input_fn: Callable[[], tf.Tensor],
attack_fn: Callable[..., Attack],
model_dir=None,
checkpoint_path=None,
preprocessing=(0, 1),
image_id=None,
**kwargs,
) -> Optional[np.ndarray]:
# Check that model has been trained.
if not checkpoint_path:
checkpoint_path = saver.latest_checkpoint(model_dir)
if not checkpoint_path:
raise ValueError(
"Could not find trained model in model_dir: {}.".format(model_dir))
with tf.Graph().as_default():
features = input_fn()
model = create_model()
image_tensor = tf.placeholder(features.dtype, features.shape)
logits = model(image_tensor)
torch_model = TorchLeNet()
path = abspath("lenet_model.pth")
torch_model.load_state_dict(torch.load(path))
torch_model.cuda()
device_id = torch.cuda.current_device()
torch_model.eval()
attack_model = foolbox.models.PyTorchModel(
torch_model,
bounds=(0, 1),
num_classes=10,
channel_axis=1,
preprocessing=preprocessing,
)
sm = SessionManager()
with sm.prepare_session(
master="",
saver=tf.train.Saver(),
checkpoint_filename_with_path=checkpoint_path,
config=new_session_config(),
) as sess:
image = sess.run(features)[0]
# attack_model = TensorFlowModel(
# image_tensor,
# logits,
# bounds=(0, 1),
# channel_axis=1,
# preprocessing=preprocessing)
# image, _ = torch_mnist.image(image_id, is_val=True)
# transform = transforms.Compose([transforms.ToTensor()])
# image = transform(image).cpu().numpy()
attack = attack_fn(attack_model)
with torch.cuda.device(device_id):
adversarial_example = attack(image, label=label, **kwargs)
# adversarial_example = attack(image[0], label=label, **kwargs)
if adversarial_example is None:
return None
else:
return adversarial_example[np.newaxis]
