def test_get_actions(self, batch_size, hidden_sizes):
"""Test get_actions function."""
env_spec = GymEnv(DummyBoxEnv())
obs_dim = env_spec.observation_space.flat_dim
act_dim = env_spec.action_space.flat_dim
obs = torch.ones([batch_size, obs_dim], dtype=torch.float32)
init_std = 2.
policy = GaussianMLPPolicy(env_spec=env_spec,
hidden_sizes=hidden_sizes,
init_std=init_std,
hidden_nonlinearity=None,
std_parameterization='exp',
hidden_w_init=nn.init.ones_,
output_w_init=nn.init.ones_)
dist = policy(obs)[0]
expected_mean = torch.full([batch_size, act_dim],
obs_dim *
(torch.Tensor(hidden_sizes).prod().item()),
dtype=torch.float)
expected_variance = init_std**2
action, prob = policy.get_actions(obs)
assert np.array_equal(prob['mean'], expected_mean.numpy())
assert dist.variance.equal(
torch.full((batch_size, act_dim),
expected_variance,
dtype=torch.float))
assert action.shape == (batch_size, act_dim)
def test_get_action_dict_space(self):
"""Test if observations from dict obs spaces are properly flattened."""
env = GymEnv(DummyDictEnv(obs_space_type='box', act_space_type='box'))
policy = GaussianMLPPolicy(env_spec=env.spec,
hidden_nonlinearity=None,
hidden_sizes=(1, ),
hidden_w_init=nn.init.ones_,
output_w_init=nn.init.ones_)
obs = env.reset()[0]
action, _ = policy.get_action(obs)
assert env.action_space.shape == action.shape
actions, _ = policy.get_actions(np.array([obs, obs]))
for action in actions:
assert env.action_space.shape == action.shape
actions, _ = policy.get_actions(np.array([obs, obs]))
for action in actions:
assert env.action_space.shape == action.shape
def test_policy_get_actions(mock_model, input_dim, output_dim, hidden_sizes):
action = torch.randn((output_dim, ))
mock_dist = mock.MagicMock()
mock_dist.rsample.return_value = action
mock_model.return_value = mock_dist
env_spec = mock.MagicMock()
env_spec.observation_space.flat_dim = input_dim
env_spec.action_space.flat_dim = output_dim
policy = GaussianMLPPolicy(env_spec, mock_model)
input = torch.ones(input_dim)
sample = policy.get_actions(input)
assert np.array_equal(sample, action.detach().numpy())
def test_is_pickleable(self, batch_size, hidden_sizes):
env_spec = TfEnv(DummyBoxEnv())
obs_dim = env_spec.observation_space.flat_dim
obs = torch.ones([batch_size, obs_dim], dtype=torch.float32)
init_std = 2.
policy = GaussianMLPPolicy(env_spec=env_spec,
hidden_sizes=hidden_sizes,
init_std=init_std,
hidden_nonlinearity=None,
std_parameterization='exp',
hidden_w_init=nn.init.ones_,
output_w_init=nn.init.ones_)
output1_action, output1_prob = policy.get_actions(obs)
p = pickle.dumps(policy)
policy_pickled = pickle.loads(p)
output2_action, output2_prob = policy_pickled.get_actions(obs)
assert output1_prob['mean'].equal(output2_prob['mean'])
assert output1_action.shape == output2_action.shape
