def _thunk():
env = make_atari(env_id)
env.seed(seed + rank)
if record_video:
video_path = os.path.join(output_dir, 'video/env-%d' % rank)
ensure_dir(video_path)
env = Monitor(env, video_path, video_callable=lambda episode_id: episode_id % record_video_freq == 0, force=True)
return wrap_deepmind(env, episode_life=True, clip_rewards=True, frame_stack=False)
def run(env_id, model_path):
env = make_atari(env_id)
env = wrap_deepmind(env,
episode_life=True,
clip_rewards=True,
frame_stack=False)
num_env = 1
valid_actions = valid_atari_actions(env, env_id)
num_actions = len(valid_actions)
ob_space = env.observation_space
obs = np.zeros((num_env, 84, 84, 4), dtype=np.uint8)
next_obs = env.reset()
obs = update_obs(obs, next_obs)
steps = 0
total_reward = 0
with tf.Session() as sess:
print('Loading Model %s' % model_path)
policy = Policy(sess, num_actions, num_env, num_steps=1)
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(model_path)
saver.restore(sess, ckpt.model_checkpoint_path)
while True:
env.render()
actions, values = policy.step(obs)
value = values[0]
steps += 1
next_obs, rewards, dones, info = env.step(
valid_actions[actions[0]])
total_reward += rewards
print('%d: reward=%f value=%f' % (steps, total_reward, value))
obs = update_obs(obs, next_obs)
if dones:
print('DONE')
steps = 0
total_reward = 0
next_obs = env.reset()
obs = np.zeros((num_env, 84, 84, 4), dtype=np.uint8)
obs = update_obs(obs, next_obs)
time.sleep(2)
def main():
env_id = 'PongNoFrameskip-v4'
seed = randint(0, 1000000)
env = make_atari(env_id, frame_skip=1)
env.seed(seed)
env = wrap_deepmind(env)
env.reset()
t = 1
while True:
env.render()
next_obs, reward, done, info = env.step(env.action_space.sample())
print('%d: reward=%f done=%s info=%s' % (t, reward, done, info))
if done:
print('DONE')
next_obs = env.reset()
t = 1
time.sleep(2)
def run(env_id, model_path, record_video, video_path=None):
env = make_atari(env_id)
env = wrap_deepmind(env,
episode_life=True,
clip_rewards=True,
frame_stack=False)
num_env = 1
ob_space = env.observation_space
ac_space = env.action_space
obs = np.zeros((num_env, 84, 84, 4), dtype=np.uint8)
next_obs = env.reset()
obs = update_obs(obs, next_obs)
ep = 1
steps = 0
total_reward = 0
with tf.Session() as sess:
print('Loading Model %s' % model_path)
policy = CnnPolicy(sess, ob_space, ac_space, nbatch=1, nsteps=1)
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(model_path)
saver.restore(sess, ckpt.model_checkpoint_path)
ts = ts_rand()
if record_video:
ensure_dir(video_path)
video_recorder = VideoRecorder(env,
path=ep_video_path(
video_path, ts, env_id, ep))
while True:
env.render()
if record_video:
video_recorder.capture_frame()
actions, values, _ = policy.step(obs)
value = values[0]
steps += 1
next_obs, rewards, dones, info = env.step(actions)
total_reward += rewards
print('%d: reward=%f value=%f' % (steps, total_reward, value))
obs = update_obs(obs, next_obs)
if dones:
print('DONE')
ep += 1
steps = 0
total_reward = 0
next_obs = env.reset()
obs = np.zeros((num_env, 84, 84, 4), dtype=np.uint8)
obs = update_obs(obs, next_obs)
if record_video:
video_recorder.close()
video_recorder = VideoRecorder(env,
path=ep_video_path(
video_path, ts, env_id,
ep),
enabled=record_video)
'`'
time.sleep(2)