)
args = parser.parse_args()
samples = {
"state": [],
"state_img": [],
"next_state": [],
"next_state_img": [],
"reward": [],
"action": [],
"terminal": [],
}
env = LunarLander()
env.render()
env.viewer.window.on_key_press = key_press
env.viewer.window.on_key_release = key_release
a = np.array([0])
episode_rewards = []
steps = 0
while True:
episode_reward = 0
state = env.reset()
state_img = env.render(
mode="rgb_array")[::4, ::4, :] # downsampling (every 4th pixel).
while True:
if key == KEY.ENTER: _human_saliency = not _human_saliency
if key == KEY.F: _human_fast_forward = not _human_fast_forward
if key == KEY.A: _human_agent_action = ACTIONS.index('LEFT')
if key == KEY.D: _human_agent_action = ACTIONS.index('RIGHT')
if key == KEY.W: _human_agent_action = ACTIONS.index('FIRE')
if key == KEY.Q: _human_agent_action = ACTIONS.index('RIGHTFIRE')
if key == KEY.E: _human_agent_action = ACTIONS.index('LEFTFIRE')
def key_release(key, mod):
global _human_agent_action
_human_agent_action = ACTIONS.index('NOOP')
env.render()
env.unwrapped.viewer.window.on_key_press = key_press
env.unwrapped.viewer.window.on_key_release = key_release
# create models
if RAM:
a = random.randrange(env.action_space.n)
s, r, done, info = env.step(a)
N_STATE = len(s)
MODEL = LanderDQN if 'lunar' in opt.env else RamDQN
policy_net = MODEL(N_STATE, N_ACTIONS).to(device)
target_net = MODEL(N_STATE, N_ACTIONS).to(device)
else:
MODEL = DDQN if opt.dueling else DQN
policy_net = MODEL(n_actions=N_ACTIONS).to(device)
target_net = MODEL(n_actions=N_ACTIONS).to(device)