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}
class Runner:
def __init__(self, args):
with open(args.src_filepath, "rb") as f:
model = pickle.load(f)
env = gym.make(args.env_name)
self.model = model
self.env = env
self.n_iter = args.n_iter
self.rec_flag = args.rec
self.recorder = VideoRecorder(env, base_path=args.src_filepath)
def __call__(self):
for i in range(self.n_iter):
score = self.get_score()
print(score)
def close(self):
self.recorder.close()
self.env.close()
def get_score(self):
env = self.env
obs = env.reset()
acc = 0
while True:
y = self.model(obs)
action = np.random.choice(len(y), p=F.softmax(y))
obs, reward, done, info = env.step(action)
acc += reward
if self.rec_flag:
self.recorder.capture_frame()
if done:
break
return acc
def test_no_frames():
env = BrokenRecordableEnv()
rec = VideoRecorder(env)
rec.close()
assert rec.empty
assert rec.functional
assert not os.path.exists(rec.path)
def play(env, act, stochastic, video_path, nb_atoms):
num_episodes = 0
video_recorder = VideoRecorder(
env, video_path, enabled=video_path is not None)
obs = env.reset()
if args.visual:
action_names = dqn_core.actions_from_env(env)
plot_machine = PlotMachine(nb_atoms, env.action_space.n, action_names)
while True:
env.unwrapped.render()
video_recorder.capture_frame()
action = act(np.array(obs)[None], args.alpha, 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 main():
args = parse_args()
env = make_atari(args.env)
env = wrap_deepmind(env)
# setup the model to process actions for one environment and one step at a time
model = acktr_disc.Model(policies.CnnPolicy, env.observation_space, env.action_space, 1, 1)
# load the trainable parameters from our trained file
model.load(args.model_path)
# keep track of the last 4 frames of observations
env_width = env.observation_space.shape[0]
env_height = env.observation_space.shape[1]
obs_history = np.zeros((1, env_width, env_height, 4), dtype=np.uint8)
# if we're supposed to show how the model sees the game
if args.show_observation:
obs = env.reset()
import pygame
from pygame import surfarray
# the default size is too small, scale it up
scale_factor = args.scale_factor
screen = pygame.display.set_mode((env_width*scale_factor, env_height*scale_factor), 0, 8)
# setup a gray palette
pygame.display.set_palette(tuple([(i, i, i) for i in range(256)]))
# if we're supposed to record video
video_path = args.video_path
if video_path is not None:
video_recorder = VideoRecorder(
env, base_path=video_path, enabled=video_path is not None)
while True:
obs, done = env.reset(), False
episode_rew = 0
while not done:
env.render()
if args.show_observation:
# use the Kronecker product to scale the array up for display, and also transpose x/y axes because pygame
# displays as column/row instead of gym's row/column
transposed = obs_history[0,:,:,-1].transpose((1,0))
scaled_array = np.uint8(np.kron(transposed, np.ones((scale_factor, scale_factor))))
surfarray.blit_array(screen, scaled_array)
pygame.display.flip()
if video_path is not None:
video_recorder.capture_frame()
# add the current observation onto our history list
obs_history = np.roll(obs_history, shift=-1, axis=3)
obs_history[:, :, :, -1] = obs[None][:, :, :, 0]
# get the suggested action for the current observation history
action, v, _ = model.step(obs_history)
obs, rew, done, info = env.step(action)
episode_rew += rew
print("Episode reward", episode_rew)
# if we're taking video, stop it now and clear video path so no more frames are added if we're out of lives or there are no lives in this game
if video_path is not None and ('ale.lives' not in info or info['ale.lives'] == 0):
video_path = None
video_recorder.close()
def test_record_breaking_render_method():
env = BrokenRecordableEnv()
rec = VideoRecorder(env)
rec.capture_frame()
rec.close()
assert rec.empty
assert rec.broken
assert not os.path.exists(rec.path)
def test_record_breaking_render_method():
env = BrokenRecordableEnv()
rec = VideoRecorder(env)
rec.capture_frame()
rec.close()
assert rec.empty
assert rec.broken
assert not os.path.exists(rec.path)
def test_text_envs():
env = gym.make('FrozenLake-v0')
video = VideoRecorder(env)
try:
env.reset()
video.capture_frame()
video.close()
finally:
os.remove(video.path)
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 test_text_envs():
env = gym.make('FrozenLake-v0')
video = VideoRecorder(env)
try:
env.reset()
video.capture_frame()
video.close()
finally:
os.remove(video.path)
def test_record_simple():
env = gym.make("CartPole-v1")
rec = VideoRecorder(env)
env.reset()
rec.capture_frame()
rec.close()
assert not rec.empty
assert not rec.broken
assert os.path.exists(rec.path)
f = open(rec.path)
assert os.fstat(f.fileno()).st_size > 100
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 test_record_simple():
env = gym.make("CartPole-v1")
rec = VideoRecorder(env)
env.reset()
rec.capture_frame()
rec.close()
assert not rec.empty
assert not rec.broken
assert os.path.exists(rec.path)
f = open(rec.path)
assert os.fstat(f.fileno()).st_size > 100
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 render(env, recorde=False):
if recorde:
rec = VideoRecorder(env)
else:
rec = None
mean_reward = 0.0
mean_traj_reward = 0.0
max_run_time = 0.0
min_run_time = 1e+10
mean_run_time = 0.0
for i in range(5):
total_reward = 0.0
traj_total_reward = 0.0
idx = 0
done = False
obs = env.reset()
while done == False:
env.render()
x = np.reshape(obs, [1, -1])
pred = rl_model.run(x, None)
action = np.argmax(pred)
obs, _, done, info = env.step(action)
total_reward += _
traj_total_reward += hc_model.predict(obs.reshape([1, -1]))
if rec != None:
rec.capture_frame()
idx += 1
if done or idx > 300:
if idx > max_run_time:
max_run_time = idx
elif idx < min_run_time:
min_run_time = idx
mean_run_time += idx
mean_reward += total_reward
mean_traj_reward += traj_total_reward
break
if rec != None:
rec.close()
print "[ RunLength =",5," MeanReward =",mean_reward / 5.0, "MeantrajReward =",mean_traj_reward/5.0,\
" MeanRunTime =",mean_run_time / 5.0, " MaxRunTime =",max_run_time," MinRunTime =",min_run_time,"]"
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"])
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"])
def test_record_unrecordable_method():
env = UnrecordableEnv()
rec = VideoRecorder(env)
assert not rec.enabled
rec.close()
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
utils.to_bin(pole_angle, pole_angle_bins),
utils.to_bin(angle_rate_of_change, angle_rate_bins)
])
if not (done):
#qlearn.learn(state, action, reward, nextState)
state = nextState
rewards_sum += reward
else:
# Q-learn stuff
reward = -200
rewards_sum += reward
break
print(rewards_sum)
test_reward.append(rewards_sum)
record.close()
print("Test Meaning_value: {}".format(np.mean(test_reward)))
print("Test std: {}".format(np.std(test_reward)))
# print("length: ")
print("Test finished...")
x = np.asarray(x, dtype=int)
reward_y = np.asarray(reward_y, dtype=float)
plt.figure()
plt.plot(x, reward_y)
plt.savefig('cartpole.png')
plt.show()
trajectory = last_time_steps.tolist()
class VideoWrapper(Wrapper):
"""Creates videos from wrapped environment by called render after each timestep."""
def __init__(self, env, directory, single_video=True):
"""
:param env: (gym.Env) the wrapped environment.
:param directory: the output directory.
:param single_video: (bool) if True, generates a single video file, with episodes
concatenated. If False, a new video file is created for each episode.
Usually a single video file is what is desired. However, if one is
searching for an interesting episode (perhaps by looking at the
metadata), saving to different files can be useful.
"""
super(VideoWrapper, self).__init__(env)
self.episode_id = 0
self.video_recorder = None
self.single_video = single_video
self.directory = osp.abspath(directory)
# Make sure to not put multiple different runs in the same directory,
# if the directory already exists
error_msg = "You're trying to use the same directory twice, " \
"this would result in files being overwritten"
assert not os.path.exists(self.directory), error_msg
os.makedirs(self.directory, exist_ok=True)
def _step(self, action):
obs, rew, done, info = self.env.step(action)
if done:
winners = [i for i, d in info.items() if 'winner' in d]
metadata = {'winners': winners}
self.video_recorder.metadata.update(metadata)
self.video_recorder.capture_frame()
return obs, rew, done, info
def _reset(self):
self._reset_video_recorder()
self.episode_id += 1
return self.env.reset()
def _reset_video_recorder(self):
"""Called at the start of each episode (by _reset). Always creates a video recorder
if one does not already exist. When a video recorder is already present, it will only
create a new one if `self.single_video == False`."""
if self.video_recorder is not None:
# Video recorder already started.
if not self.single_video:
# We want a new video for each episode, so destroy current recorder.
self.video_recorder.close()
self.video_recorder = None
if self.video_recorder is None:
# No video recorder -- start a new one.
self.video_recorder = VideoRecorder(
env=self.env,
base_path=osp.join(self.directory,
'video.{:06}'.format(self.episode_id)),
metadata={'episode_id': self.episode_id},
)
def _close(self):
if self.video_recorder is not None:
self.video_recorder.close()
self.video_recorder = None
super(VideoWrapper, self)._close()
def test_record_unrecordable_method():
env = UnrecordableEnv()
rec = VideoRecorder(env)
assert not rec.enabled
rec.close()