def train():
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
if sys.platform == "darwin":
binary_path = "./bin/Reacher.app"
elif sys.platform == "linux":
binary_path = "./bin/Reacher_Linux_NoVis/Reacher.x86_64"
else:
binary_path = "./bin/Reacher_Windows_x86_64/Reacher.exe"
env = UnityEnvironment(file_name=binary_path)
n_agent, state_dim, action_dim = get_env_info(env)
model = GaussianActorCriticNetwork(state_dim,
action_dim,
hiddens=[512, 256])
model = model.to(device)
agent = PPOAgent(env,
model,
tmax=128,
n_epoch=10,
batch_size=128,
eps=0.1,
device=device)
n_step = 2000
n_episodes = 0
for step in range(n_step):
agent.step()
scores = agent.scores_by_episode
if n_episodes < len(scores):
n_episodes = len(scores)
print(f" episode #{n_episodes} : score = {scores[-1]:.2f}", end="")
if 100 <= n_episodes:
rewards_ma = np.mean(scores[-100:])
print(f", mean score of last 100 episodes = {rewards_ma:.2f}")
if 30. <= rewards_ma:
torch.save(model.state_dict(), "bestmodel.pth")
with open('rewards.pickle', 'wb') as fp:
pickle.dump(scores, fp)
print("\n ==== Achieved criteria! Model is saved.")
break
else:
print()
sys.stdout.flush()
print("Finished.")
def view():
# for different environments please change these parameters
if sys.platform == "darwin":
binary_path = "./bin/Reacher.app"
elif sys.platform == "linux":
binary_path = "./bin/Reacher_Linux/Reacher.x86_64"
else:
binary_path = "./bin/Reacher_Windows_x86_64/Reacher.exe"
env = UnityEnvironment(file_name=binary_path)
n_agent, state_dim, action_dim = get_env_info(env)
model = GaussianActorCriticNetwork(state_dim, action_dim, hiddens=[512, 256])
model = model.to(device)
# load best Model
state_dict = torch.load("model.pth",
map_location=lambda storage, loc: storage)
model.load_state_dict(state_dict)
# Reset Env
brain_name = env.brain_names[0]
env_info = env.reset(train_mode=False)[brain_name]
states = to_tensor(env_info.vector_observations)
n_step = 5000
model.eval()
for step in range(n_step):
# draw action from model
actions, _, _, _ = model(states)
# one step forward
actions_np = actions.cpu().numpy()
env_info = env.step(actions_np)[brain_name]
states = to_tensor(env_info.vector_observations)