os.environ["CUDA_VISIBLE_DEVICES"] = gpu_device
if discretize:
codes = np.load(codes_path)
if __name__ == '__main__':
mnist = input_data.read_data_sets('MNIST_data', one_hot=False)
with tf.Session() as sess:
model = Model()
saver = tf.train.Saver()
checkpoint = tf.train.latest_checkpoint(model_dir)
saver.restore(sess, checkpoint)
if discretize:
attack = CWAttack(model, attack_steps, step_size, epsilon, codes,
eval_batch_size, alpha)
else:
attack = LinfPGDAttack(model, epsilon, attack_steps, step_size,
random_start, loss_func)
# Iterate over the samples batch-by-batch
num_batches = int(math.ceil(num_eval_examples / eval_batch_size))
total_corr_nat = 0
total_corr_adv = 0
rate = []
for ibatch in range(num_batches):
bstart = ibatch * eval_batch_size
bend = min(bstart + eval_batch_size, num_eval_examples)
x_batch = mnist.test.images[bstart:bend, :]
with tf.Session(config=tf_config) as sess:
# Restore the checkpoint
saver.restore(sess, model_path)
# Iterate over the samples batch-by-batch
num_batches = int(math.ceil(num_eval_examples / eval_batch_size))
total_xent_nat = 0.
total_xent_adv = 0.
total_corr_nat = 0
total_corr_adv = 0
if discretize:
attack = CWAttack(model,
num_steps=attack_steps,
step_size=step_size,
epsilon=epsilon,
codes=codes,
batch_size=eval_batch_size,
alpha=alpha)
else:
attack = LinfPGDAttack(model,
epsilon=epsilon,
num_steps=attack_steps,
step_size=step_size,
random_start=random_start,
loss_func=loss_func)
for ibatch in range(num_batches):
bstart = ibatch * eval_batch_size
bend = min(bstart + eval_batch_size, num_eval_examples)