def test_gaussian_dist_instance():
torch.manual_seed(0)
act_size = 4
dist_instance = GaussianDistInstance(torch.zeros(1, act_size),
torch.ones(1, act_size))
action = dist_instance.sample()
assert action.shape == (1, act_size)
for log_prob in dist_instance.log_prob(torch.zeros(
(1, act_size))).flatten():
# Log prob of standard normal at 0
assert log_prob == pytest.approx(-0.919, abs=0.01)
for ent in dist_instance.entropy().flatten():
# entropy of standard normal at 0, based on 1/2 + ln(sqrt(2pi)sigma)
assert ent == pytest.approx(1.42, abs=0.01)
def test_get_probs_and_entropy():
# Test continuous
# Add two dists to the list. This isn't done in the code but we'd like to support it.
dist_list = [
GaussianDistInstance(torch.zeros((1, 2)), torch.ones((1, 2))),
GaussianDistInstance(torch.zeros((1, 2)), torch.ones((1, 2))),
]
action_list = [torch.zeros((1, 2)), torch.zeros((1, 2))]
log_probs, entropies, all_probs = ModelUtils.get_probs_and_entropy(
action_list, dist_list)
assert log_probs.shape == (1, 2, 2)
assert entropies.shape == (1, 2, 2)
assert all_probs is None
for log_prob in log_probs.flatten():
# Log prob of standard normal at 0
assert log_prob == pytest.approx(-0.919, abs=0.01)
for ent in entropies.flatten():
# entropy of standard normal at 0
assert ent == pytest.approx(1.42, abs=0.01)
# Test continuous
# Add two dists to the list.
act_size = 2
test_prob = torch.tensor([[1.0 - 0.1 * (act_size - 1)] + [0.1] *
(act_size - 1)]) # High prob for first action
dist_list = [
CategoricalDistInstance(test_prob),
CategoricalDistInstance(test_prob)
]
action_list = [torch.tensor([0]), torch.tensor([1])]
log_probs, entropies, all_probs = ModelUtils.get_probs_and_entropy(
action_list, dist_list)
assert all_probs.shape == (1, len(dist_list * act_size))
assert entropies.shape == (1, len(dist_list))
# Make sure the first action has high probability than the others.
assert log_probs.flatten()[0] > log_probs.flatten()[1]
def test_get_probs_and_entropy():
inp_size = 4
act_size = 2
action_model, masks = create_action_model(inp_size, act_size)
_continuous_dist = GaussianDistInstance(torch.zeros((1, 2)),
torch.ones((1, 2)))
act_size = 2
test_prob = torch.tensor([[1.0 - 0.1 * (act_size - 1)] + [0.1] *
(act_size - 1)])
_discrete_dist_list = [
CategoricalDistInstance(test_prob),
CategoricalDistInstance(test_prob),
]
dist_tuple = DistInstances(_continuous_dist, _discrete_dist_list)
agent_action = AgentAction(torch.zeros(
(1, 2)), [torch.tensor([0]), torch.tensor([1])])
log_probs, entropies = action_model._get_probs_and_entropy(
agent_action, dist_tuple)
assert log_probs.continuous_tensor.shape == (1, 2)
assert len(log_probs.discrete_list) == 2
for _disc in log_probs.discrete_list:
assert _disc.shape == (1, )
assert len(log_probs.all_discrete_list) == 2
for _disc in log_probs.all_discrete_list:
assert _disc.shape == (1, 2)
for clp in log_probs.continuous_tensor[0]:
# Log prob of standard normal at 0
assert clp == pytest.approx(-0.919, abs=0.01)
assert log_probs.discrete_list[0] > log_probs.discrete_list[1]
for ent, val in zip(entropies[0], [1.4189, 0.6191, 0.6191]):
assert ent == pytest.approx(val, abs=0.01)