# Wrappers
env = ResizeWrapper(env)
env = NormalizeWrapper(env)
env = ImgWrapper(env) # to make the images from 160x120x3 into 3x160x120
env = ActionWrapper(env)
# env = DtRewardWrapper(env) # not during testing
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(file_name, directory="./pytorch_models")
with torch.no_grad():
while True:
obs = env.reset()
env.render()
rewards = []
while True:
action = policy.predict(np.array(obs))
obs, rew, done, misc = env.step(action)
rewards.append(rew)
env.render()
if done:
break
print("mean episode reward:", np.mean(rewards))
obs = env.get_features()
EPISODES, STEPS = 20, 1000
DEBUG = False
# please notice
logger = Logger(env, log_file=f'train-{int(EPISODES*STEPS/1000)}k.log')
start_time = time.time()
print(
f"[INFO]Starting to get logs for {EPISODES} episodes each {STEPS} steps..")
with torch.no_grad():
# let's collect our samples
for episode in range(0, EPISODES):
for steps in range(0, STEPS):
# we use our 'expert' to predict the next action.
action = expert.predict(np.array(obs))
# Apply the action
observation, reward, done, info = env.step(action)
# Get features(state representation) for RL agent
obs = env.get_features()
if done:
print(f"#Episode: {episode}\t | #Step: {steps}")
break
closest_point, _ = env.closest_curve_point(env.cur_pos,
env.cur_angle)
if closest_point is None:
done = True
break
# Cut the horizon: obs.shape = (480,640,3) --> (300,640,3)