def testClipToBounds(self):
value = tf.constant([1, 2, 4, -3])
spec = tensor_spec.BoundedTensorSpec((4, ), tf.float32, [0, 0, 0, 0],
[3, 3, 3, 3])
expected_clipped_value = np.array([1, 2, 3, 0])
clipped_value = common.clip_to_spec(value, spec)
clipped_value_ = self.evaluate(clipped_value)
self.assertAllClose(expected_clipped_value, clipped_value_)
def testScaledDistribution(self):
output_spec = tensor_spec.BoundedTensorSpec([1], tf.float32, -2, 4)
network = normal_projection_network.NormalProjectionNetwork(
output_spec, init_means_output_factor=10, state_dependent_std=True,
scale_distribution=True)
inputs = _get_inputs(batch_size=100, num_input_dims=5)
distributions, _ = network(inputs, outer_rank=1)
self.evaluate(tf.compat.v1.global_variables_initializer())
sample = self.evaluate(distributions.sample())
clipped = self.evaluate(common.clip_to_spec(sample, output_spec))
np.testing.assert_almost_equal(clipped, sample)
def clip_action(action, action_spec):
if isinstance(action_spec, tensor_spec.BoundedTensorSpec):
return common.clip_to_spec(action, action_spec)
return action
def _sample(dist, action_spec):
action = dist.sample(seed=seed_stream())
if self._clip:
return common_utils.clip_to_spec(action, action_spec)
return action
def _add_ou_noise(action, ou_process, action_spec):
noisy_action = action + ou_process()
if self._clip:
return common.clip_to_spec(noisy_action, action_spec)
return noisy_action
def _sample(dist, action_spec):
action = dist.sample(seed=seed)
if self._clip:
return common.clip_to_spec(action, action_spec)
return action
