obs_img = np.zeros((50, 50))
obs_dis = last_distance
obs_ori = orientation
new_obs_img = np.zeros((50, 50))
new_obs_dis = last_distance
new_obs_ori = orientation
##################################################
''' We create the policy network (the Actor model) '''
policy = TD3(action_dim, max_action)
##################################################
''' We create the Experience Replay memory'''
replay_buffer = ReplayBuffer(sample_size=sample_size)
##################################################
im = CoreImage("./images/MASK1.png")
imgCV2 = cv2.imread('./images/MASK1.png')
rows, cols, dims = imgCV2.shape
# Initializing the map
first_update = True
def init():
global sand
global goal_x
global goal_y
episode_reward = 0
maxepisode_timesteps = 500
torch.manual_seed(seed)
np.random.seed(seed)
state_dim = 5
action_dim = 1
max_action = 5
min_action = -5
# Getting our AI, which we call "brain", and that contains our neural network that represents our Q-function
#brain = Dqn(5,3,0.9)
action2rotation = [0, 5, -5]
#spacenetwork = ObsSpaceNetwork()
policy = TD3(state_dim, action_dim, max_action)
replay_buffer = ReplayBuffer()
last_reward = 0
scores = []
im = CoreImage("./images/MASK1.png")
total_timesteps = 0
timesteps_since_eval = 0
episode_num = 0
episode_timesteps = 0
done = True
t0 = time.time()
# textureMask = CoreImage(source="./kivytest/simplemask1.png")
# Initializing the map
first_update = True
i = 0
def update(self, dt):
global longueur
global largeur
longueur = self.width
largeur = self.height
if first_update:
init()
def evaluate_policy(policy, eval_episodes=10):
avg_reward = 0.
for _ in range(eval_episodes):
obs = reset(self)
done = False
while not done:
action = policy.select_action(obs)
obs, reward, done, _ = Car.move(action)
avg_reward += reward
avg_reward /= eval_episodes
print("---------------------------------------")
print("Average Reward over the Evaluation Step: %f" % (avg_reward))
print("---------------------------------------")
return avg_reward
file_name = "%s_%s_%s" % ("TD3", env_name, str(seed))
print("---------------------------------------")
print("Settings: %s" % (file_name))
print("-------------------------------------- -")
if not os.path.exists("./results"):
os.makedirs("./results")
if save_models and not os.path.exists("./pytorch_models"):
os.makedirs("./pytorch_models")
torch.manual_seed(seed)
np.random.seed(seed)
state_dim = [32, 32, 1]
action_dim = 1
max_action = 5
policy = TD3(state_dim, action_dim, max_action)
replay_buffer = ReplayBuffer()
evaluations = [evaluate_policy(policy)]
def mkdir(base, name):
path = os.path.join(base, name)
if not os.path.exists(path):
os.makedirs(path)
return path
work_dir = mkdir('exp', 'brs')
monitor_dir = mkdir(work_dir, 'monitor')
max_episode_steps = 400
total_timesteps = 0
timesteps_since_eval = 0
episode_num = 0
done = True
t0 = time.time()
# We start the main loop over 40,000 timesteps
while total_timesteps < max_timesteps:
# If the episode is done
if done:
# If we are not at the very beginning, we start the training process of the model
if (total_timesteps != 0 and total_timesteps > (batch_size)):
print("Total Timesteps: {} Episode Num: {} Reward: {}".
format(total_timesteps, episode_num, episode_reward))
policy.train(replay_buffer, episode_timesteps, batch_size,
discount, tau, policy_noise, noise_clip,
policy_freq)
# We evaluate the episode and we save the policy
if timesteps_since_eval >= eval_freq:
timesteps_since_eval %= eval_freq
evaluations.append(evaluate_policy(policy))
policy.save(file_name, directory="./pytorch_models")
np.save("./results/%s" % (file_name), evaluations)
# When the training step is done, we reset the state of the environment
obs = reset()
# Set the Done to False
done = False
# Set rewards and episode timesteps to zero
episode_reward = 0
episode_timesteps = 0
episode_num += 1
# Before 10000 timesteps, we play random actions
if total_timesteps < start_timesteps:
action = np.random.normal(0, 1,
size=1).clip(-1,
1).astype(np.float32)
else: # After 10000 timesteps, we switch to the model
action = policy.select_action(obs)
# If the explore_noise parameter is not 0, we add noise to the action and we clip it
if expl_noise != 0:
action = (action +
np.random.normal(0, expl_noise, size=1)).clip(
-1, 1)
# The agent performs the action in the environment, then reaches the next state and receives the reward
new_obs, reward, done, _ = move(action)
# We check if the episode is done
# done_bool = 0 if episode_timesteps + 1 == env._max_episode_steps else float(done)
if episode_timesteps + 1 == max_episode_steps:
done = True
done = float(done)
# We increase the total reward
episode_reward += reward
# We store the new transition into the Experience Replay memory (ReplayBuffer)
replay_buffer.add((obs, new_obs, action, reward, done_bool))
# We update the state, the episode timestep, the total timesteps, and the timesteps since the evaluation of the policy
obs = new_obs
episode_timesteps += 1
total_timesteps += 1
timesteps_since_eval += 1
t1 = time.time()
print("Total time taken: {}".format(t1 - t0))
evaluations.append(evaluate_policy(policy))
if save_models:
policy.save("%s" % (file_name), directory="./pytorch_models")
np.save("./results/%s" % (file_name), evaluations)
CarApp().stop()