# Training
n_steps = math.floor(cfg['environment']['max_time'] /
cfg['environment']['control_dt'])
total_steps = n_steps * env.num_envs
avg_rewards = []
avg_dones = []
fig, ax = plt.subplots(1,
2,
constrained_layout=True,
sharex=True,
figsize=[10.8, 4.8])
actor = ppo_module.Actor(
ppo_module.MLP(cfg['architecture']['policy_net'], nn.LeakyReLU, ob_dim,
act_dim),
ppo_module.MultivariateGaussianDiagonalCovariance(act_dim, 1.0), 'cuda')
critic = ppo_module.Critic(
ppo_module.MLP(cfg['architecture']['value_net'], nn.LeakyReLU, ob_dim, 1),
'cuda')
ppo = PPO.PPO(
actor=actor,
critic=critic,
num_envs=cfg['environment']['num_envs'],
num_transitions_per_env=n_steps,
num_learning_epochs=4,
gamma=0.996,
lam=0.95,
num_mini_batches=4,
"Can't find trained weight, please provide a trained weight with --weight switch\n"
)
else:
print("Loaded weight from {}\n".format(weight_path))
start = time.time()
env.reset()
reward_ll_sum = 0
done_sum = 0
average_dones = 0.
n_steps = math.floor(cfg['environment']['max_time'] /
cfg['environment']['control_dt'])
total_steps = n_steps * 1
start_step_id = 0
print("Visualizing and evaluating the policy: ", weight_path)
loaded_graph = ppo_module.MLP(cfg['architecture']['policy_net'],
torch.nn.LeakyReLU, ob_dim, act_dim)
loaded_graph.load_state_dict(
torch.load(weight_path)['actor_architecture_state_dict'])
env.load_scaling(weight_dir, int(iteration_number))
env.turn_on_visualization()
# max_steps = 1000000
max_steps = 1000 ## 10 secs
for step in range(max_steps):
time.sleep(0.01)
obs = env.observe(False)
action_ll = loaded_graph.architecture(torch.from_numpy(obs).cpu())
reward_ll, dones = env.step(action_ll.cpu().detach().numpy())
reward_ll_sum = reward_ll_sum + reward_ll[0]