# Load 5K samples
with open("../common/5k_samples.pkl","r") as f:
samples_5k = pickle.load(f)
images = samples_5k["images"]
labels = samples_5k["labels"]
# Aggregate
xs, y_trues, y_preds, y_preds_adversarial, noises = [], [], [], [], []
# Attack each example
attacker = Attack(weights="../common/weights.pkl")
for x, y_true in tqdm(zip(images, labels)):
y_target = random.choice( list(set([0,1,2,3,4,5,6,7,8,9]) - set([y_true])) )
noise, y_pred, y_pred_adversarial = attacker.attack(x, y_true, y_target, regularization="l2")
if y_pred == y_true:
# store
xs.append(x)
y_trues.append(y_true)
y_preds.append(y_pred)
y_preds_adversarial.append(y_pred_adversarial)
noises.append(noise.squeeze())
else:
print "y_pred != y_true, not storing to disk"
with open("bulk_mnist_adversarial_examples.pkl","w") as f:
save_dict = {"xs":xs,
"y_trues":y_trues,
"y_preds":y_preds,
