def test_attack():
agent = Agent(args.img_stack, device)
agent.load_param()
env = Env(args.seed, args.img_stack, args.action_repeat)
# load adv input, by default general attack perturbation
delta_s = np.load('param/adv_general.npy')
if args.attack_type != 'general':
file_path = 'param/adv_' + args.attack_type
if args.attack_type == 'patch':
file_path += '_' + args.patch_type
file_path += '.npy'
delta_s = np.load(file_path)
# show adv
fig = plt.figure(figsize=(8, 8))
plt.title('Stack of ' + str(args.img_stack) +
' adversarial signals seen by Agent')
plt.axis('off')
columns, rows = args.img_stack // 2, args.img_stack // 2
for i in range(1, columns * rows + 1):
# denormalize while showing the image
img = (delta_s[i - 1] + 1) * 128
fig.add_subplot(rows, columns, i)
plt.imshow(img, cmap='gray')
plt.show()
for i_ep in range(10):
score = 0
state = env.reset()
for t in range(1000):
# steps range to render attack in 1000
attack_render = [30, 40]
if t in np.arange(attack_render[0], attack_render[1] + 1):
if t in attack_render:
s_with_ds = (state + delta_s)
# clip the image limits and denormalize for displaying
s_with_ds = np.clip(s_with_ds, -1, 0.9921875)
s_with_ds = (s_with_ds + 1) * 128
title = 'Attack Started' if t == attack_render[
0] else 'Attack ended'
title += ' (showing first frame of 4 frames visible to policy)'
plt.imshow(s_with_ds[0], cmap='gray')
plt.axis('off')
plt.title(title)
plt.show()
state += delta_s
action = agent.select_action(state)
state_, reward, done, die = env.step(action *
np.array([2., 1., 1.]) +
np.array([-1., 0., 0.]))
if args.render:
env.render()
score += reward
state = state_
if done:
break
print('Ep {}\tScore: {:.2f}\t'.format(i_ep, score))
def run_agent():
agent = Agent(args.img_stack, device)
agent.load_param()
env = Env(args.seed, args.img_stack, args.action_repeat)
state = env.reset()
# Prepare attack
attack = AdvAttack(args.attack_type)
attack.initialize_perturbation(state.shape)
attack.load_networks()
for i_ep in range(50):
score = 0
state = env.reset()
for t in range(1000):
action = agent.select_action(state)
# update buffer for training the attack
attack.update_buffer(state)
# write to tensorboard
input_imgs_to_net = torch.tensor(
(attack.buffer['s'] + attack.buffer['d_s']))
input_imgs_grid = make_grid(input_imgs_to_net[0].reshape(
4, 1, 96, 96))
writer.add_image('Four stack of input state with adversarial',
input_imgs_grid)
writer.add_graph(attack.net, input_imgs_to_net)
writer.close()
# train attack
attack.train()
state_, reward, done, die = env.step(action *
np.array([2., 1., 1.]) +
np.array([-1., 0., 0.]))
if args.render:
env.render()
score += reward
state = state_
if done or die:
break
print('Ep {}\tScore: {:.2f}\t'.format(i_ep, score))