def test_stepsequence_from_pandas(mock_data, given_rewards: bool):
rewards, states, observations, actions, hidden, policy_infos = mock_data
states = np.asarray(states)
observations = np.asarray(observations)
actions = to.stack(actions).numpy()
rewards = np.asarray(rewards)
# Create fake observed data set. The labels must match the labels of the spaces. The order can be mixed.
content = dict(
s0=states[:, 0],
s1=states[:, 1],
s2=states[:, 2],
o3=observations[:, 3],
o0=observations[:, 0],
o2=observations[:, 2],
o1=observations[:, 1],
a1=actions[:, 1],
a0=actions[:, 0],
# Some content that was not in
steps=np.arange(0, states.shape[0]),
infos=[dict(foo="bar")] * 6,
)
if given_rewards:
content["rewards"] = rewards
df = pd.DataFrame(dict([(k, pd.Series(v)) for k, v in content.items()]))
env = MockEnv(
state_space=InfBoxSpace(shape=states[0].shape,
labels=["s0", "s1", "s2"]),
obs_space=InfBoxSpace(shape=observations[0].shape,
labels=["o0", "o1", "o2", "o3"]),
act_space=InfBoxSpace(shape=actions[0].shape, labels=["a0", "a1"]),
)
reconstructed = StepSequence.from_pandas(df, env.spec)
assert len(reconstructed.rewards) == len(rewards)
assert np.allclose(reconstructed.states, states)
assert np.allclose(reconstructed.observations, observations)
assert np.allclose(reconstructed.actions, actions)
def test_reset():
mockenv = MockEnv(act_space=BoxSpace(-1, 1, shape=(2, )),
obs_space=BoxSpace(-1, 1, shape=(2, )))
wenv = DownsamplingWrapper(mockenv, factor=2)
# Perform some actions
wenv.step(np.array([0, 4]))
assert mockenv.last_act == [0, 4]
wenv.step(np.array([4, 4]))
assert mockenv.last_act == [0, 4]
wenv.step(np.array([4, 4]))
assert mockenv.last_act == [4, 4]
# The next action would be [4, 4] again, but now we reset
wenv.reset()
assert wenv._act_last is None
assert wenv._cnt == 0
wenv.step(np.array([1, 2]))
assert mockenv.last_act == [1, 2]
wenv.step(np.array([2, 3]))
assert mockenv.last_act == [1, 2]
def test_domain_param():
mockenv = MockEnv(act_space=BoxSpace(-1, 1, shape=(2, )))
wenv = ActDelayWrapper(mockenv, delay=1)
# Reset to initialize buffer
wenv.reset()
# Perform some actions
wenv.step(np.array([0, 1]))
assert mockenv.last_act == [0, 0]
wenv.step(np.array([2, 4]))
assert mockenv.last_act == [0, 1]
# change the delay and reset
wenv.domain_param = {'act_delay': 2}
wenv.reset()
wenv.step(np.array([1, 2]))
assert mockenv.last_act == [0, 0]
wenv.step(np.array([2, 3]))
assert mockenv.last_act == [0, 0]
wenv.step(np.array([8, 9]))
assert mockenv.last_act == [1, 2]
def test_indi_nonlin_layer(in_features, same_nonlin, bias, weight):
if not same_nonlin and in_features > 1:
nonlin = in_features*[to.tanh]
else:
nonlin = to.sigmoid
layer = IndiNonlinLayer(in_features, nonlin, bias, weight)
assert isinstance(layer, nn.Module)
i = to.randn(in_features)
o = layer(i)
assert isinstance(o, to.Tensor)
assert i.shape == o.shape
@pytest.mark.parametrize(
'env', [MockEnv(obs_space=InfBoxSpace(shape=1), act_space=InfBoxSpace(shape=1))]
)
@pytest.mark.parametrize(
'policy', [
# Two-headed policies are not supported
'rnn_policy',
'lstm_policy',
'gru_policy',
'adn_policy',
'nf_policy',
],
ids=['rnn', 'lstm', 'gru', 'adn', 'nf'],
indirect=True
)
@pytest.mark.parametrize('windowed', [True, False],
ids=['windowed', 'not_windowed'])