def play(env, act, stochastic, video_path):
num_episodes = 0
video_recorder = None
video_recorder = VideoRecorder(env,
video_path,
enabled=video_path is not None)
obs = env.reset()
if args.visual:
action_names = distdeepq.actions_from_env(env)
plot_machine = distdeepq.PlotMachine(dist_params, env.action_space.n,
action_names)
while True:
env.unwrapped.render()
video_recorder.capture_frame()
action = act(np.array(obs)[None], stochastic=stochastic)[0]
obs, rew, done, info = env.step(action)
if args.visual:
plot_machine.plot_distribution(np.array(obs)[None])
if done:
obs = env.reset()
if len(info["rewards"]) > num_episodes:
if len(info["rewards"]) == 1 and video_recorder.enabled:
# save video of first episode
print("Saved video.")
video_recorder.close()
video_recorder.enabled = False
print(info["rewards"][-1])
num_episodes = len(info["rewards"])
def play(env, act, stochastic, video_path):
num_episodes = 0
video_recorder = None
video_recorder = VideoRecorder(env,
video_path,
enabled=video_path is not None)
obs = env.reset()
X = []
while True:
env.unwrapped.render()
video_recorder.capture_frame()
action = act(np.array(obs)[None], stochastic=stochastic)[0]
obs, rew, done, info = env.step(action)
if done:
obs = env.reset()
if len(info["rewards"]) > num_episodes:
if len(info["rewards"]) == 1 and video_recorder.enabled:
# save video of first episode
print("Saved video.")
video_recorder.close()
video_recorder.enabled = False
print("Score in this episode: " + str(info["rewards"][-1]))
X.append(info["rewards"][-1])
num_episodes = len(info["rewards"])
print("Average Score so far: " + str(sum(X) / float(num_episodes)))
def play(env, act, stochastic, video_path):
num_episodes = 0
video_recorder = None
video_recorder = VideoRecorder(env,
video_path,
enabled=video_path is not None)
obs = env.reset()
if args.show_observation:
fig = plt.figure()
im = plt.imshow(obs._frames[-1].reshape((84, 84)), cmap='Greys')
plt.show(False)
while True:
env.unwrapped.render()
if args.show_observation:
im.set_data(obs._frames[-1].reshape((84, 84)))
fig.canvas.draw()
video_recorder.capture_frame()
action = act(np.array(obs)[None], stochastic=stochastic)[0]
obs, rew, done, info = env.step(action)
if done:
obs = env.reset()
if len(info["rewards"]) > num_episodes:
if len(info["rewards"]) == 1 and video_recorder.enabled:
# save video of first episode
print("Saved video.")
video_recorder.close()
video_recorder.enabled = False
print(info["rewards"][-1])
num_episodes = len(info["rewards"])
def play(env, act, stochastic, video_path, clipped, num_trials=10):
num_episodes = 0
video_recorder = VideoRecorder(env,
video_path,
enabled=video_path is not None)
obs = env.reset()
reward = 0
num_played = 0
rewardArray = []
while num_played < num_trials:
env.unwrapped.render()
video_recorder.capture_frame()
action = act(np.array(obs)[None], stochastic=stochastic)[0]
obs, rew, done, info = env.step(action)
if clipped:
rew = clip_score(rew)
reward += rew
if done:
obs = env.reset()
if len(info["rewards"]) > num_episodes:
if len(info["rewards"]) == 1 and video_recorder.enabled:
# save video of first episode
print("Saved video.")
video_recorder.close()
video_recorder.enabled = False
print(info["rewards"][-1])
rewardArray.append(reward)
reward = 0
num_played += 1
num_episodes = len(info["rewards"])
return {"Nonclipped": info["rewards"], "Clipped": rewardArray}
def play(
env,
act,
craft_adv_obs,
craft_adv_obs2,
stochastic,
video_path,
attack,
m_target,
m_adv,
):
num_episodes = 0
num_moves = 0
num_transfer = 0
video_recorder = None
video_recorder = VideoRecorder(env,
video_path,
enabled=video_path is not None)
obs = env.reset()
while True:
env.unwrapped.render()
video_recorder.capture_frame()
# V: Attack #
if attack is not None:
# Craft adv. examples
with m_adv.get_session().as_default():
adv_obs = craft_adv_obs(np.array(obs)[None],
stochastic_adv=stochastic)[0]
with m_target.get_session().as_default():
action = act(np.array(adv_obs)[None], stochastic=stochastic)[0]
action2 = act(np.array(obs)[None], stochastic=stochastic)[0]
num_moves += 1
if action != action2:
num_transfer += 1
else:
# Normal
action = act(np.array(obs)[None], stochastic=stochastic)[0]
obs, rew, done, info = env.step(action)
if done:
obs = env.reset()
if len(info["rewards"]) > num_episodes:
if len(info["rewards"]) == 1 and video_recorder.enabled:
# save video of first episode
print("Saved video.")
video_recorder.close()
video_recorder.enabled = False
print("Reward: " + str(info["rewards"][-1]))
num_episodes = len(info["rewards"])
print("Episode: " + str(num_episodes))
success = float(num_transfer / num_moves) * 100.0
print("Percentage of successful attacks: " + str(success))
num_moves = 0
num_transfer = 0
def play(env, act, craft_adv_obs, stochastic, video_path, game_name, attack,
defense):
if defense == 'foresight':
vf, game_screen_mean = load_visual_foresight(game_name)
pred_obs = deque(maxlen=4)
num_episodes = 0
video_recorder = None
video_recorder = VideoRecorder(env,
video_path,
enabled=video_path is not None)
t = 0
obs = env.reset()
while True:
#env.unwrapped.render()
video_recorder.capture_frame()
# Attack
if craft_adv_obs != None:
# Craft adv. examples
adv_obs = craft_adv_obs(np.array(obs)[None],
stochastic=stochastic)[0]
action = act(np.array(adv_obs)[None], stochastic=stochastic)[0]
else:
# Normal
action = act(np.array(obs)[None], stochastic=stochastic)[0]
# Defense
if t > 4 and defense == 'foresight':
pred_obs.append(
foresee(U.get_session(), old_obs, old_action, np.array(obs),
game_screen_mean, vf, env.action_space.n, t))
if len(pred_obs) == 4:
action = act(np.stack(pred_obs, axis=2)[None],
stochastic=stochastic)[0]
old_obs = obs
old_action = action
# RL loop
obs, rew, done, info = env.step(action)
t += 1
if done:
t = 0
obs = env.reset()
if len(info["rewards"]) > num_episodes:
if len(info["rewards"]) == 1 and video_recorder.enabled:
# save video of first episode
print("Saved video.")
video_recorder.close()
video_recorder.enabled = False
print(info["rewards"][-1])
num_episodes = len(info["rewards"])
def play(env, act, craft_adv_obs, craft_adv_obs2, stochastic, video_path,
attack, m_target, m_adv):
num_episodes = 0
num_moves = 0
num_transfer = 0
video_recorder = None
video_recorder = VideoRecorder(
env, video_path, enabled=video_path is not None)
obs = env.reset()
while True:
env.unwrapped.render()
video_recorder.capture_frame()
# V: Attack #
if attack is not None:
# Craft adv. examples
with m_adv.get_session().as_default():
adv_obs = \
craft_adv_obs(np.array(obs)[None],
stochastic_adv=stochastic)[0]
with m_target.get_session().as_default():
action = act(np.array(adv_obs)[None],
stochastic=stochastic)[0]
action2 = act(np.array(obs)[None], stochastic=stochastic)[0]
num_moves += 1
if action != action2:
num_transfer += 1
else:
# Normal
action = act(np.array(obs)[None], stochastic=stochastic)[0]
obs, rew, done, info = env.step(action)
if done:
obs = env.reset()
if len(info["rewards"]) > num_episodes:
if len(info["rewards"]) == 1 and video_recorder.enabled:
# save video of first episode
print("Saved video.")
video_recorder.close()
video_recorder.enabled = False
print('Reward: ' + str(info["rewards"][-1]))
num_episodes = len(info["rewards"])
print('Episode: ' + str(num_episodes))
success = float(num_transfer / num_moves) * 100.0
print("Percentage of successful attacks: " + str(success))
num_moves = 0
num_transfer = 0
def play(env, act, stochastic, video_path):
num_episodes = 0
video_recorder = None
video_recorder = VideoRecorder(
env, video_path, enabled=video_path is not None)
obs = env.reset()
while True:
env.unwrapped.render()
video_recorder.capture_frame()
action = act(np.array(obs)[None], stochastic=stochastic)[0]
obs, rew, done, info = env.step(action)
if done:
obs = env.reset()
if len(info["rewards"]) > num_episodes:
if len(info["rewards"]) == 1 and video_recorder.enabled:
# save video of first episode
print("Saved video.")
video_recorder.close()
video_recorder.enabled = False
print(info["rewards"][-1])
num_episodes = len(info["rewards"])
batch_size=BATCH_SIZE,
shuffle=True)
match_env(env_real, env_sim)
video_recorder = VideoRecorder(env_real, 'real.mp4', enabled=True)
video_recorder2 = VideoRecorder(env_sim, 'sim.mp4', enabled=True)
for i, data in enumerate(dataloader_train):
for j in range(50):
env_sim.render()
env_real.render()
action = data["actions"][0, j].numpy()
video_recorder.capture_frame()
video_recorder2.capture_frame()
obs_real, _, _, _ = env_real.step(action.copy())
obs_simp, _, _, _ = env_sim.step(action.copy())
env_real.reset()
env_sim.reset()
match_env(env_real, env_sim)
if i == 10:
break
video_recorder.close()
video_recorder.enabled = False
video_recorder2.close()
video_recorder2.enabled = False
