# Launch the env with our helper function
env = launch_env()
# 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
noise_sigma = 0.2
noise = OUNoise(mu=np.zeros(action_dim), sigma=noise_sigma)
# Load VAE
image_dimensions = 3 * 160 * 120
feature_dimensions = 1000
encoding_dimensions = 40
vae = VAE(image_dimensions, feature_dimensions, encoding_dimensions, 'selu')
if(use_pr):
replay_buffer = utils.PrioritizedReplayBuffer(args.replay_buffer_max_size)
else:
replay_buffer = utils.ReplayBuffer(args.replay_buffer_max_size)
# Initialize policy
policy = DDPG(state_dim, action_dim, max_action, replay_buffer, net_type="vae", vae=vae)
# Evaluate untrained policy
evaluations= [evaluate_policy(env, policy)]
exp.metric("rewards", evaluations[0])
total_timesteps = 0
timesteps_since_eval = 0
episode_num = 0
done = True
episode_reward = None
env_counter = 0
while total_timesteps < args.max_timesteps:
# Wrappers
env = ResizeWrapper(env)
env = NormalizeWrapper(env)
env = ImgWrapper(env) # to make the images from 160x120x3 into 3x160x120
env = ActionWrapper(env)
env = DtRewardWrapper(env)
# Set seeds
seed(args.seed)
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")
replay_buffer = utils.ReplayBuffer(args.replay_buffer_max_size)
# Evaluate untrained policy
evaluations = [evaluate_policy(env, policy)]
exp.metric("rewards", evaluations[0])
total_timesteps = 0
timesteps_since_eval = 0
episode_num = 0
done = True
episode_reward = None
env_counter = 0
while total_timesteps < args.max_timesteps:
camera_height=480,
accept_start_angle_deg=4,
full_transparency=True,
distortion=True,
randomize_maps_on_reset=True,
draw_curve=False,
draw_bbox=False,
frame_skip=4,
draw_DDPG_features=False)
state_dim = env.get_features().shape[0]
action_dim = env.action_space.shape[0]
max_action = float(env.action_space.high[0])
# Initialize policy
expert = DDPG(state_dim, action_dim, max_action, net_type="dense")
expert.load("model-here",
directory="../duckietown_rl/pytorch_models",
for_inference=True)
# Initialize the environment
env.reset()
# Get features(state representation) for RL agent
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()