def test_std_network_output_values_with_batch(input_dim, output_dim,
hidden_sizes, learn_std):
init_std = 2.
module = GaussianMLPModule(input_dim=input_dim,
output_dim=output_dim,
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_,
learn_std=learn_std)
batch_size = 5
dist = module(torch.ones([batch_size, input_dim]))
exp_mean = torch.full(
(batch_size, output_dim),
input_dim * (torch.Tensor(hidden_sizes).prod().item()),
dtype=torch.float)
exp_variance = init_std**2
assert dist.mean.equal(exp_mean)
assert dist.variance.equal(
torch.full((batch_size, output_dim), exp_variance, dtype=torch.float))
assert dist.rsample().shape == (batch_size, output_dim)
def test_softplus_min_std(input_dim, output_dim, hidden_sizes):
min_value = 2.
module = GaussianMLPModule(input_dim=input_dim,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
init_std=1.,
min_std=min_value,
hidden_nonlinearity=None,
std_parameterization='softplus',
hidden_w_init=nn.init.zeros_,
output_w_init=nn.init.zeros_)
dist = module(torch.ones(input_dim))
exp_variance = torch.Tensor([min_value]).exp().add(1.).log()**2
assert dist.variance.equal(torch.full((output_dim, ), exp_variance[0]))
def test_exp_max_std(input_dim, output_dim, hidden_sizes):
max_value = 1.
module = GaussianMLPModule(input_dim=input_dim,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
init_std=10.,
max_std=max_value,
hidden_nonlinearity=None,
std_parameterization='exp',
hidden_w_init=nn.init.zeros_,
output_w_init=nn.init.zeros_)
dist = module(torch.ones(input_dim))
exp_variance = max_value**2
assert dist.variance.equal(torch.full((output_dim, ), exp_variance))
def test_softplus_max_std(input_dim, output_dim, hidden_sizes):
max_value = 1.
module = GaussianMLPModule(input_dim=input_dim,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
init_std=10,
max_std=max_value,
hidden_nonlinearity=None,
std_parameterization='softplus',
hidden_w_init=nn.init.ones_,
output_w_init=nn.init.ones_)
dist = module(torch.ones(input_dim))
exp_variance = torch.Tensor([max_value]).exp().add(1.).log()**2
assert torch.equal(
dist.variance,
torch.full((output_dim, ), exp_variance[0], dtype=torch.float))
def test_std_network_output_values(input_dim, output_dim, hidden_sizes):
init_std = 2.
module = GaussianMLPModule(input_dim=input_dim,
output_dim=output_dim,
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 = module(torch.ones(input_dim))
exp_mean = torch.full(
(output_dim, ), input_dim * (torch.Tensor(hidden_sizes).prod().item()))
exp_variance = init_std**2
assert dist.mean.equal(exp_mean)
assert dist.variance.equal(torch.full((output_dim, ), exp_variance))
assert dist.rsample().shape == (output_dim, )
def test_softplus_std_network_output_values(input_dim, output_dim,
hidden_sizes):
init_std = 2.
module = GaussianMLPModule(input_dim=input_dim,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
init_std=init_std,
hidden_nonlinearity=None,
std_parameterization='softplus',
hidden_w_init=nn.init.ones_,
output_w_init=nn.init.ones_)
dist = module(torch.ones(input_dim))
exp_mean = input_dim * torch.Tensor(hidden_sizes).prod().item()
exp_variance = torch.Tensor([init_std]).exp().add(1.).log()**2
assert dist.mean.equal(
torch.full((output_dim, ), exp_mean, dtype=torch.float))
assert dist.variance.equal(
torch.full((output_dim, ), exp_variance[0], dtype=torch.float))
assert dist.rsample().shape == (output_dim, )
def test_unknown_std_parameterization():
with pytest.raises(NotImplementedError):
GaussianMLPModule(input_dim=1,
output_dim=1,
std_parameterization='unknown')