def _enjoy():
# Launch the env with our helper function
env = launch_env()
print("Initialized environment")
# Wrappers
env = ResizeWrapper(env)
env = NormalizeWrapper(env)
env = ImgWrapper(env) # to make the images from 160x120x3 into 3x160x120
env = ActionWrapper(env)
env = DtRewardWrapper(env)
print("Initialized Wrappers")
state_dim = env.observation_space.shape
action_dim = env.action_space.shape[0]
max_action = float(env.action_space.high[0])
# Initialize policy
policy = DDPG(state_dim, action_dim, max_action, net_type="cnn")
policy.load(filename='ddpg', directory='reinforcement/pytorch/models/')
obs = env.reset()
done = False
while True:
while not done:
action = policy.predict(np.array(obs))
# Perform action
obs, reward, done, _ = env.step(action)
env.render()
done = False
obs = env.reset()
def _enjoyWindow():
model = WindowModel(action_dim=2, max_action=1.)
try:
state_dict = torch.load('./models/windowimitate.pt')
model.load_state_dict(state_dict)
except:
print('failed to load model')
exit()
model.eval().to(device)
env = launch_env1()
env = ResizeWrapper(env)
env = NormalizeWrapper(env)
env = ImgWrapper(env)
env = ActionWrapper(env)
env = DtRewardWrapper(env)
obs = env.reset()
obsWindow = np.zeros((12, 160, 120))
while True:
obsWindow[:9, :, :] = obsWindow[3:, :, :]
obsWindow[9:12, :, :] = obs
obs = torch.from_numpy(obsWindow).float().to(device).unsqueeze(0)
action = model(obs)
action = action.squeeze().data.cpu().numpy()
obs, reward, done, info = env.step(action)
env.render()
if done:
if reward < 0:
print('*** FAILED ***')
time.sleep(0.7)
obs = env.reset()
env.render()
def _enjoy():
model = Model(action_dim=2, max_action=1.)
try:
state_dict = torch.load('trained_models/imitate.pt',
map_location=device)
model.load_state_dict(state_dict)
except:
print('failed to load model')
exit()
model.eval().to(device)
env = launch_env()
env = ResizeWrapper(env)
env = NormalizeWrapper(env)
env = ImgWrapper(env)
env = ActionWrapper(env)
env = DtRewardWrapper(env)
obs = env.reset()
while True:
obs = torch.from_numpy(obs).float().to(device).unsqueeze(0)
action = model(obs)
action = action.squeeze().data.cpu().numpy()
obs, reward, done, info = env.step(action)
env.render()
if done:
if reward < 0:
print('*** FAILED ***')
time.sleep(0.7)
obs = env.reset()
env.render()
def _enjoy(args):
# Launch the env with our helper function
env = launch_env()
print("Initialized environment")
# Wrappers
env = ResizeWrapper(env)
env = GrayscaleWrapper(env)
env = NormalizeWrapper(env)
env = FrameStack(env, 4)
env = DtRewardWrapper(env)
env = ActionWrapper(env)
print("Initialized Wrappers")
state_dim = env.observation_space.shape
action_dim = env.action_space.shape[0]
max_action = float(env.action_space.high[0])
# Initialize policy
# policy = TD3(state_dim, action_dim, max_action, net_type="cnn")
# policy.load(filename=args.policy, directory='reinforcement/pytorch/models/')
policy = policies[args.policy](state_dim, action_dim, max_action)
policy.load("reinforcement/pytorch/models/", args.policy)
obs = env.reset()
done = False
while True:
while not done:
action = policy.predict(np.array(obs))
# Perform action
obs, reward, done, _ = env.step(action)
env.render()
done = False
obs = env.reset()
def initWindowModel():
model = WindowModel(action_dim=2, max_action=1.)
try:
state_dict = torch.load('./models/windowimitate.pt')
model.load_state_dict(state_dict)
except:
print('failed to load model')
exit()
model.eval().to(device)
env = launch_env1()
env = ResizeWrapper(env)
env = NormalizeWrapper(env)
env = ImgWrapper(env)
env = ActionWrapper(env)
env = DtRewardWrapper(env)
env.reset()
env.render()
return env, model
def _dagger():
model = Model(action_dim=2, max_action=1.)
try:
state_dict = torch.load('./models/imitate.pt')
model.load_state_dict(state_dict)
except:
print('failed to load model')
exit()
model.eval().to(device)
env = launch_env1()
# Register a keyboard handler
env = ResizeWrapper(env)
env = NormalizeWrapper(env)
env = ImgWrapper(env)
env = ActionWrapper(env)
env = DtRewardWrapper(env)
obs = env.reset()
env.render()
key_handler = key.KeyStateHandler()
env.unwrapped.window.push_handlers(key_handler)
print(env.map_name)
raise Exception("asdfsadf")
obsHistory = []
actionHistory = []
while True:
obs = torch.from_numpy(obs).float().to(device).unsqueeze(0)
action = model(obs)
action = action.squeeze().data.cpu().numpy()
obs, reward, done, info = env.step(action)
print(key_handler)
daggerAction = np.array([0.0, 0.0])
if key_handler[key.UP]:
print("as===as=df=sad=f=asdf=sad=fs=adf")
daggerAction = np.array([1.00, 0.0])
#action = np.array([0.44, 0.0])
if key_handler[key.DOWN]:
print("as===as=df=sad=f=asdf=sad=fs=adf")
daggerAction = np.array([-1.00, 0])
#action = np.array([-0.44, 0])
if key_handler[key.LEFT]:
print("as===as=df=sad=f=asdf=sad=fs=adf")
daggerAction = np.array([0.35, +1])
if key_handler[key.RIGHT]:
print("as===as=df=sad=f=asdf=sad=fs=adf")
daggerAction = np.array([0.35, -1])
if key_handler[key.SPACE]:
obsHistoryArray = np.array(obsHistory)
actionHistoryArray = np.array(actionHistory)
np.save('./dagger/obs_{}.npy'.format(len(count)), obsHistoryArray)
np.save('./dagger/actions_{}.npy'.format(len(count)),
actionHistoryArray)
print(daggerAction)
obsHistory.append(obs)
actionHistory.append(daggerAction)
env.render()
if done:
if reward < 0:
print('*** FAILED ***')
time.sleep(0.7)
obs = env.reset()
env.render()