def test_get_action(self, mock_normal, obs_dim, action_dim, hidden_dim):
mock_normal.return_value = 0.5
env = TfEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('metarl.tf.policies.'
'gaussian_gru_policy.GaussianGRUModel'),
new=SimpleGaussianGRUModel):
policy = GaussianGRUPolicy(env_spec=env.spec,
state_include_action=False)
policy.reset()
obs = env.reset()
expected_action = np.full(action_dim, 0.5 * np.exp(0.5) + 0.5)
action, agent_info = policy.get_action(obs)
assert env.action_space.contains(action)
assert np.allclose(action, expected_action)
expected_mean = np.full(action_dim, 0.5)
assert np.array_equal(agent_info['mean'], expected_mean)
expected_log_std = np.full(action_dim, 0.5)
assert np.array_equal(agent_info['log_std'], expected_log_std)
actions, agent_infos = policy.get_actions([obs])
for action, mean, log_std in zip(actions, agent_infos['mean'],
agent_infos['log_std']):
assert env.action_space.contains(action)
assert np.allclose(action, expected_action)
assert np.array_equal(mean, expected_mean)
assert np.array_equal(log_std, expected_log_std)
def test_get_action_state_include_action(self, obs_dim, action_dim,
hidden_dim):
env = MetaRLEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
obs_var = tf.compat.v1.placeholder(
tf.float32,
shape=[
None, None,
env.observation_space.flat_dim + np.prod(action_dim)
],
name='obs')
policy = GaussianGRUPolicy(env_spec=env.spec,
hidden_dim=hidden_dim,
state_include_action=True)
policy.build(obs_var)
policy.reset()
obs = env.reset()
action, _ = policy.get_action(obs.flatten())
assert env.action_space.contains(action)
policy.reset()
actions, _ = policy.get_actions([obs.flatten()])
for action in actions:
assert env.action_space.contains(action)