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,
device='cuda',
log_dir=saver.data_dir,
mini_batch_sampling='in_order',
)
if not test_mode:
for update in range(1000000):
ax[0].set(xlabel='iteration',
ylabel='avg performance',
title='average performance')
ax[1].set(xlabel='iteration',
ylabel='avg dones',
title='average dones')
ppo_module.MLP(cfg['architecture']['value_net'], nn.LeakyReLU, ob_dim, 1),
device)
saver = ConfigurationSaver(
log_dir=home_path + "/raisimGymTorch/data/" + task_name,
save_items=[task_path + "/cfg.yaml", task_path + "/Environment.hpp"])
tensorboard_launcher(saver.data_dir +
"/..") # press refresh (F5) after the first ppo update
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,
device=device,
log_dir=saver.data_dir,
shuffle_batch=False,
)
if mode == 'retrain':
load_param(weight_path, env, actor, critic, ppo.optimizer, saver.data_dir)
for update in range(1000000):
start = time.time()
env.reset()
reward_ll_sum = 0
done_sum = 0