def projection_net_factory(action_spec):
return normal_projection_network.NormalProjectionNetwork(
action_spec,
mean_transform=None,
state_dependent_std=True,
std_transform=sac_agent.std_clip_transform,
scale_distribution=True)
def _normal_projection_net(action_spec, init_means_output_factor=0.1):
return normal_projection_network.NormalProjectionNetwork(
action_spec,
mean_transform=None,
state_dependent_std=True,
init_means_output_factor=init_means_output_factor,
std_transform=sac_agent.std_clip_transform,
scale_distribution=True)
def _normal_projection_net(action_spec,
init_action_stddev=0.35,
init_means_output_factor=0.1):
std_bias_initializer_value = np.log(np.exp(init_action_stddev) - 1)
return normal_projection_network.NormalProjectionNetwork(
action_spec,
init_means_output_factor=init_means_output_factor,
std_bias_initializer_value=std_bias_initializer_value)
def normal_projection_net(action_spec,
init_action_stddev=0.35,
init_means_output_factor=0.1):
del init_action_stddev
return normal_projection_network.NormalProjectionNetwork(
action_spec,
mean_transform=None,
state_dependent_std=True,
init_means_output_factor=init_means_output_factor,
std_transform=sac_agent.std_clip_transform)
def _critic_normal_projection_net(output_spec,
init_stddev=0.35,
init_means_output_factor=0.1):
std_bias_initializer_value = np.log(np.exp(init_stddev) - 1)
return normal_projection_network.NormalProjectionNetwork(
output_spec,
init_means_output_factor=init_means_output_factor,
std_bias_initializer_value=std_bias_initializer_value,
scale_distribution=False)
def normal_projection_net(action_spec,
init_action_stddev=0.35,
init_means_output_factor=0.1,
scale_distribution=True):
del init_action_stddev
return normal_projection_network.NormalProjectionNetwork(
action_spec,
state_dependent_std=True,
init_means_output_factor=init_means_output_factor,
std_transform=sac_agent.std_clip_transform,
scale_distribution=scale_distribution)
def normal_projection_net(action_spec,
init_action_stddev=0.35,
init_means_output_factor=0.1):
std_bias_initializer_value = np.log(np.exp(init_action_stddev) - 1)
return normal_projection_network.NormalProjectionNetwork(
action_spec,
mean_transform=None,
state_dependent_std=True,
init_means_output_factor=init_means_output_factor,
std_transform=sac_agent.std_clip_transform,
std_bias_initializer_value=std_bias_initializer_value,
scale_distribution=True,
)
def _normal_projection_net(action_spec,
init_action_stddev=0.35,
init_means_output_factor=0.1,
seed_stream_class=tfp.util.SeedStream,
seed=None):
std_bias_initializer_value = np.log(np.exp(init_action_stddev) - 1)
return normal_projection_network.NormalProjectionNetwork(
action_spec,
init_means_output_factor=init_means_output_factor,
std_bias_initializer_value=std_bias_initializer_value,
scale_distribution=False,
seed_stream_class=seed_stream_class,
seed=seed)
def testTrainableVariables(self):
output_spec = tensor_spec.BoundedTensorSpec([2], tf.float32, 0, 1)
network = normal_projection_network.NormalProjectionNetwork(output_spec)
inputs = _get_inputs(batch_size=3, num_input_dims=5)
network(inputs, outer_rank=1)
self.evaluate(tf.compat.v1.global_variables_initializer())
# Dense kernel, dense bias, std bias.
self.assertEqual(3, len(network.trainable_variables))
self.assertEqual((5, 2), network.trainable_variables[0].shape)
self.assertEqual((2,), network.trainable_variables[1].shape)
self.assertEqual((2,), network.trainable_variables[2].shape)
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 _critic_normal_projection_net(output_spec,
init_stddev=0.35,
init_means_output_factor=0.1):
del init_stddev
# std_bias_initializer_value = round(np.log(init_action_stddev + 1e-10), 3)
return normal_projection_network.NormalProjectionNetwork(
output_spec,
init_means_output_factor=init_means_output_factor,
# std_bias_initializer_value=std_bias_initializer_value,
mean_transform=None,
std_transform=sac_agent.std_clip_transform,
state_dependent_std=True,
scale_distribution=False)
def testBuildStateDepStddev(self):
output_spec = tensor_spec.BoundedTensorSpec([2], tf.float32, 0, 1)
network = normal_projection_network.NormalProjectionNetwork(
output_spec, state_dependent_std=True, scale_distribution=False)
inputs = _get_inputs(batch_size=3, num_input_dims=5)
distribution, _ = network(inputs, outer_rank=1)
self.evaluate(tf.compat.v1.global_variables_initializer())
self.assertEqual(tfp.distributions.Normal, type(distribution))
means, stds = distribution.loc, distribution.scale
self.assertAllEqual(means.shape.as_list(),
[3] + output_spec.shape.as_list())
self.assertAllEqual(stds.shape.as_list(), [3] + output_spec.shape.as_list())
def testBuild(self):
output_spec = tensor_spec.BoundedTensorSpec([2], tf.float32, 0, 1)
network = normal_projection_network.NormalProjectionNetwork(
output_spec, scale_distribution=False)
inputs = _get_inputs(batch_size=3, num_input_dims=5)
distribution, _ = network(inputs, outer_rank=1)
self.evaluate(tf.compat.v1.global_variables_initializer())
self.assertIsInstance(
distribution, tfp.distributions.MultivariateNormalDiag)
means, stds = distribution.mean(), distribution.stddev()
self.assertAllEqual(means.shape.as_list(),
[3] + output_spec.shape.as_list())
self.assertAllEqual(stds.shape.as_list(), [3] + output_spec.shape.as_list())
