def test_save_path(self):
n_store_episodes = 10
obs = np.ones(shape=self.env.observation_space.shape, dtype=np.float32)
for epi in range(n_store_episodes):
for i in range(self.replay_buffer.get_buffer_size()):
self.replay_buffer.add(obs=obs * i,
act=i,
rew=0.,
next_obs=obs * (i + 1),
done=False)
save_path(
self.replay_buffer.sample(
self.replay_buffer.get_buffer_size()),
os.path.join(self.output_dir,
"step_0_epi_{}_return_0.0.pkl").format(epi))
data = restore_latest_n_traj(self.output_dir)
self.assertEqual(
data["obses"].shape[0],
self.replay_buffer.get_buffer_size() * n_store_episodes)
max_steps = 10
data = restore_latest_n_traj(self.output_dir, 1, max_steps)
self.assertEqual(data["obses"].shape[0], max_steps)
self.assertEqual(data["acts"].shape[0], max_steps)
self.assertEqual(data["next_obses"].shape[0], max_steps)
if args.expert_path_dir is None:
print("Plaese generate demonstrations first")
print("python examples/run_sac.py --env-name=RoboschoolReacher-v1 --save-test-path --test-interval=50000")
exit()
units = [400, 300]
env = gym.make(args.env_name)
test_env = gym.make(args.env_name)
policy = DDPG(
state_shape=env.observation_space.shape,
action_dim=env.action_space.high.size,
max_action=env.action_space.high[0],
gpu=args.gpu,
actor_units=units,
critic_units=units,
n_warmup=10000,
batch_size=100)
irl = VAIL(
state_shape=env.observation_space.shape,
action_dim=env.action_space.high.size,
units=units,
enable_sn=args.enable_sn,
batch_size=32,
gpu=args.gpu)
expert_trajs = restore_latest_n_traj(
args.expert_path_dir, n_path=20, max_steps=1000)
trainer = IRLTrainer(policy, env, args, irl, expert_trajs["obses"],
expert_trajs["next_obses"], expert_trajs["acts"], test_env)
trainer()