def test_zero_normal_projection_net(self, state_dependent_std):
"""A zero-weight net generates zero actions."""
input_spec = TensorSpec((10, ), torch.float32)
action_spec = TensorSpec((8, ), torch.float32)
embedding = input_spec.ones(outer_dims=(2, ))
net = NormalProjectionNetwork(
input_size=input_spec.shape[0],
action_spec=action_spec,
projection_output_init_gain=0,
std_bias_initializer_value=0,
squash_mean=False,
state_dependent_std=state_dependent_std,
std_transform=math_ops.identity)
out = net(embedding)[0].sample((10, ))
self.assertTrue(isinstance(net.output_spec, DistributionSpec))
self.assertEqual(tuple(out.size()), (
10,
2,
) + action_spec.shape)
self.assertTrue(torch.all(out == 0))
def test_squash_mean_normal_projection_net(self, network_ctor):
"""A net with `sqaush_mean=True` should generate means within the action
spec."""
input_spec = TensorSpec((10, ), torch.float32)
embedding = input_spec.ones(outer_dims=(100, ))
action_spec = TensorSpec((8, ), torch.float32)
# For squashing mean, we need a bounded action spec
self.assertRaises(AssertionError, network_ctor, input_spec,
action_spec)
action_spec = BoundedTensorSpec((2, ),
torch.float32,
minimum=(0, -0.01),
maximum=(0.01, 0))
net = network_ctor(
input_spec.shape[0], action_spec, projection_output_init_gain=1.0)
dist, _ = net(embedding)
self.assertTrue(dist.mean.std() > 0)
self.assertTrue(
torch.all(dist.mean > torch.as_tensor(action_spec.minimum)))
self.assertTrue(
torch.all(dist.mean < torch.as_tensor(action_spec.maximum)))