def testUpdate(self): tf.compat.v1.set_random_seed(1) actor_network = DummyActorNet(self._action_spec) value_network = DummyValueNet() policy = ppo_policy.PPOPolicy(self._time_step_spec, self._action_spec, actor_network=actor_network, value_network=value_network) new_policy = ppo_policy.PPOPolicy(self._time_step_spec, self._action_spec, actor_network=actor_network, value_network=value_network) action_step = policy.action(self._time_step) new_action_step = new_policy.action(self._time_step) self.assertEqual(action_step.action.shape, new_action_step.action.shape) self.assertEqual(action_step.action.dtype, new_action_step.action.dtype) self.evaluate(tf.compat.v1.global_variables_initializer()) self.evaluate(new_policy.update(policy)) actions_, new_actions_ = self.evaluate( [action_step.action, new_action_step.action]) self.assertAllEqual(actions_, new_actions_)
def test_same_policy_same_output(self): if not tf.executing_eagerly(): self.skipTest( 'Skipping test: sequential networks not supported in TF1') observation_tensor_spec = tf.TensorSpec(shape=[1], dtype=tf.float32) action_tensor_spec = tensor_spec.BoundedTensorSpec((8, ), tf.float32, -1, 1) value_net = value_network.ValueNetwork(observation_tensor_spec, fc_layer_params=(1, )) actor_net_lib = ppo_actor_network.PPOActorNetwork() actor_net_lib.seed_stream_class = DeterministicSeedStream actor_net_sequential = actor_net_lib.create_sequential_actor_net( fc_layer_units=(1, ), action_tensor_spec=action_tensor_spec, seed=1) actor_net_actor_dist = actor_distribution_network.ActorDistributionNetwork( observation_tensor_spec, action_tensor_spec, fc_layer_params=(1, ), activation_fn=tf.nn.tanh, kernel_initializer=tf.keras.initializers.Orthogonal(seed=1), seed_stream_class=DeterministicSeedStream, seed=1) tf.random.set_seed(111) seq_policy = ppo_policy.PPOPolicy( ts.time_step_spec(observation_tensor_spec), action_tensor_spec, actor_net_sequential, value_net, collect=True) tf.random.set_seed(111) actor_dist_policy = ppo_policy.PPOPolicy( ts.time_step_spec(observation_tensor_spec), action_tensor_spec, actor_net_actor_dist, value_net, collect=True) sample_timestep = ts.TimeStep(step_type=tf.constant([1, 1], dtype=tf.int32), reward=tf.constant([1, 1], dtype=tf.float32), discount=tf.constant([1, 1], dtype=tf.float32), observation=tf.constant( [[1], [2]], dtype=tf.float32)) seq_policy_step = seq_policy._distribution(sample_timestep, policy_state=()) act_dist_policy_step = actor_dist_policy._distribution(sample_timestep, policy_state=()) seq_scale = seq_policy_step.info['dist_params']['scale_diag'] act_dist_scale = act_dist_policy_step.info['dist_params']['scale'] self.assertAllEqual(seq_scale, act_dist_scale) self.assertAllEqual(seq_policy_step.info['dist_params']['loc'], act_dist_policy_step.info['dist_params']['loc'])
def _make_policy(self, collect): return ppo_policy.PPOPolicy( time_step_spec=self.time_step_spec(), action_spec=self.action_spec(), actor_network=self._actor_net, value_network=self._value_net, observation_normalizer=self._observation_normalizer, clip=False, collect=collect)
def testBuild(self, network_cls): actor_network = network_cls(self._action_spec) value_network = DummyValueNet() policy = ppo_policy.PPOPolicy(self._time_step_spec, self._action_spec, actor_network=actor_network, value_network=value_network) self.assertEqual(policy.time_step_spec, self._time_step_spec) self.assertEqual(policy.action_spec, self._action_spec)
def testGaussianDistribution(self): actor_network = DummyActorDistributionNet(self._action_spec) value_network = DummyValueNet() policy = ppo_policy.PPOPolicy(self._time_step_spec, self._action_spec, actor_network=actor_network, value_network=value_network) distribution_step = policy.distribution(self._time_step) self.assertIsInstance(distribution_step.action, tfp.distributions.Normal)
def testReset(self, network_cls): actor_network = network_cls(self._action_spec) value_network = DummyValueNet() policy = ppo_policy.PPOPolicy(self._time_step_spec, self._action_spec, actor_network=actor_network, value_network=value_network) policy_state = policy.get_initial_state(batch_size=1) # Dummy network has no policy_state so expect empty tuple from reset. self.assertEqual((), policy_state)
def testActionBatch(self, network_cls): actor_network = network_cls(self._action_spec) value_network = DummyValueNet() policy = ppo_policy.PPOPolicy(self._time_step_spec, self._action_spec, actor_network=actor_network, value_network=value_network) action_step = policy.action(self._time_step_batch) self.assertEqual(action_step.action.shape.as_list(), [2, 1]) self.assertEqual(action_step.action.dtype, tf.float32) self.evaluate(tf.compat.v1.global_variables_initializer()) actions_ = self.evaluate(action_step.action) self.assertTrue(np.all(actions_ >= self._action_spec.minimum)) self.assertTrue(np.all(actions_ <= self._action_spec.maximum))
def testActionList(self, network_cls): action_spec = [self._action_spec] actor_network = network_cls(action_spec) value_network = DummyValueNet() policy = ppo_policy.PPOPolicy(self._time_step_spec, action_spec, actor_network=actor_network, value_network=value_network) action_step = policy.action(self._time_step) self.assertIsInstance(action_step.action, list) self.evaluate(tf.compat.v1.global_variables_initializer()) actions_ = self.evaluate(action_step.action) self.assertTrue(np.all(actions_ >= action_spec[0].minimum)) self.assertTrue(np.all(actions_ <= action_spec[0].maximum))
def testValueInPolicyInfo(self, network_cls): actor_network = network_cls(self._action_spec) value_network = DummyValueNet() policy = ppo_policy.PPOPolicy(self._time_step_spec, self._action_spec, actor_network=actor_network, value_network=value_network) policy_step = policy.action(self._time_step) self.assertEqual(policy_step.info['value_prediction'].shape.as_list(), [1, 1]) self.assertEqual(policy_step.info['value_prediction'].dtype, tf.float32) self.evaluate(tf.compat.v1.global_variables_initializer()) self.evaluate(policy_step.info['value_prediction'])
def testDeterministicDistribution(self): actor_network = DummyActorNet(self._action_spec) value_network = DummyValueNet() policy = ppo_policy.PPOPolicy(self._time_step_spec, self._action_spec, actor_network=actor_network, value_network=value_network) action_step = policy.action(self._time_step) distribution_step = policy.distribution(self._time_step) self.assertIsInstance(distribution_step.action, tfp.distributions.Deterministic) distribution_mean = distribution_step.action.mean() self.evaluate(tf.compat.v1.global_variables_initializer()) actions_ = self.evaluate(action_step.action) distribution_mean_ = self.evaluate(distribution_mean) self.assertNear(actions_, distribution_mean_, 1e-6)
def testValue(self, network_cls): actor_network = network_cls(self._action_spec) value_network = DummyValueNet() policy = ppo_policy.PPOPolicy(self._time_step_spec, self._action_spec, actor_network=actor_network, value_network=value_network) batch_size = self._time_step.step_type.shape[0].value policy_state = policy.get_initial_state(batch_size=batch_size) value_pred, unused_policy_state = policy.apply_value_network( self._time_step.observation, self._time_step.step_type, policy_state) self.assertEqual(value_pred.shape.as_list(), [1, 1]) self.assertEqual(value_pred.dtype, tf.float32) self.evaluate(tf.global_variables_initializer()) self.evaluate(value_pred)
def testNonLegacyDistribution(self): if not tf.executing_eagerly(): self.skipTest( 'Skipping test: sequential networks not supported in TF1') actor_network = create_sequential_actor_net() action_spec = {'my_action': self._action_spec} value_network = DummyValueNet() policy = ppo_policy.PPOPolicy(self._time_step_spec, action_spec, actor_network=actor_network, value_network=value_network) distribution_step = policy.distribution(self._time_step) self.assertIsInstance(distribution_step.action['my_action'], tfp.distributions.TransformedDistribution) expected_info_spec = { 'dist_params': { 'my_action': { 'bijector': { 'bijectors:0': {}, 'bijectors:1': {}, 'bijectors:2': {} }, 'distribution': { 'scale': tf.TensorSpec([1], tf.float32), 'loc': tf.TensorSpec([1], tf.float32) }, } }, 'value_prediction': tf.TensorSpec([1, 1], tf.float32) } tf.nest.map_structure( lambda v, s: self.assertEqual(tf.type_spec_from_value(v), s), distribution_step.info, expected_info_spec)
def __init__(self, time_step_spec, action_spec, optimizer=None, actor_net=None, value_net=None, importance_ratio_clipping=0.0, lambda_value=0.95, discount_factor=0.99, entropy_regularization=0.0, policy_l2_reg=0.0, value_function_l2_reg=0.0, value_pred_loss_coef=0.5, num_epochs=25, use_gae=False, use_td_lambda_return=False, normalize_rewards=True, reward_norm_clipping=10.0, normalize_observations=True, log_prob_clipping=0.0, kl_cutoff_factor=2.0, kl_cutoff_coef=1000.0, initial_adaptive_kl_beta=1.0, adaptive_kl_target=0.01, adaptive_kl_tolerance=0.3, gradient_clipping=None, check_numerics=False, debug_summaries=False, summarize_grads_and_vars=False, train_step_counter=None, name=None): """Creates a PPO Agent. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A nest of BoundedTensorSpec representing the actions. optimizer: Optimizer to use for the agent. actor_net: A function actor_net(observations, action_spec) that returns tensor of action distribution params for each observation. Takes nested observation and returns nested action. value_net: A function value_net(time_steps) that returns value tensor from neural net predictions for each observation. Takes nested observation and returns batch of value_preds. importance_ratio_clipping: Epsilon in clipped, surrogate PPO objective. For more detail, see explanation at the top of the doc. lambda_value: Lambda parameter for TD-lambda computation. discount_factor: Discount factor for return computation. entropy_regularization: Coefficient for entropy regularization loss term. policy_l2_reg: Coefficient for l2 regularization of policy weights. value_function_l2_reg: Coefficient for l2 regularization of value function weights. value_pred_loss_coef: Multiplier for value prediction loss to balance with policy gradient loss. num_epochs: Number of epochs for computing policy updates. use_gae: If True (default False), uses generalized advantage estimation for computing per-timestep advantage. Else, just subtracts value predictions from empirical return. use_td_lambda_return: If True (default False), uses td_lambda_return for training value function. (td_lambda_return = gae_advantage + value_predictions) normalize_rewards: If true, keeps moving variance of rewards and normalizes incoming rewards. reward_norm_clipping: Value above and below to clip normalized reward. normalize_observations: If true, keeps moving mean and variance of observations and normalizes incoming observations. log_prob_clipping: +/- value for clipping log probs to prevent inf / NaN values. Default: no clipping. kl_cutoff_factor: If policy KL changes more than this much for any single timestep, adds a squared KL penalty to loss function. kl_cutoff_coef: Loss coefficient for kl cutoff term. initial_adaptive_kl_beta: Initial value for beta coefficient of adaptive kl penalty. adaptive_kl_target: Desired kl target for policy updates. If actual kl is far from this target, adaptive_kl_beta will be updated. adaptive_kl_tolerance: A tolerance for adaptive_kl_beta. Mean KL above (1 + tol) * adaptive_kl_target, or below (1 - tol) * adaptive_kl_target, will cause adaptive_kl_beta to be updated. gradient_clipping: Norm length to clip gradients. Default: no clipping. check_numerics: If true, adds tf.debugging.check_numerics to help find NaN / Inf values. For debugging only. debug_summaries: A bool to gather debug summaries. summarize_grads_and_vars: If true, gradient summaries will be written. train_step_counter: An optional counter to increment every time the train op is run. Defaults to the global_step. name: The name of this agent. All variables in this module will fall under that name. Defaults to the class name. Raises: ValueError: If the actor_net is not a DistributionNetwork. """ if not isinstance(actor_net, network.DistributionNetwork): raise ValueError( 'actor_net must be an instance of a DistributionNetwork.') tf.Module.__init__(self, name=name) self._optimizer = optimizer self._actor_net = actor_net self._value_net = value_net self._importance_ratio_clipping = importance_ratio_clipping self._lambda = lambda_value self._discount_factor = discount_factor self._entropy_regularization = entropy_regularization self._policy_l2_reg = policy_l2_reg self._value_function_l2_reg = value_function_l2_reg self._value_pred_loss_coef = value_pred_loss_coef self._num_epochs = num_epochs self._use_gae = use_gae self._use_td_lambda_return = use_td_lambda_return self._reward_norm_clipping = reward_norm_clipping self._log_prob_clipping = log_prob_clipping self._kl_cutoff_factor = kl_cutoff_factor self._kl_cutoff_coef = kl_cutoff_coef self._adaptive_kl_target = adaptive_kl_target self._adaptive_kl_tolerance = adaptive_kl_tolerance self._gradient_clipping = gradient_clipping or 0.0 self._check_numerics = check_numerics if initial_adaptive_kl_beta > 0.0: # TODO(kbanoop): Rename create_variable. self._adaptive_kl_beta = common.create_variable( 'adaptive_kl_beta', initial_adaptive_kl_beta, dtype=tf.float32) else: self._adaptive_kl_beta = None self._reward_normalizer = None if normalize_rewards: self._reward_normalizer = tensor_normalizer.StreamingTensorNormalizer( tensor_spec.TensorSpec([], tf.float32), scope='normalize_reward') self._observation_normalizer = None if normalize_observations: self._observation_normalizer = ( tensor_normalizer.StreamingTensorNormalizer( time_step_spec.observation, scope='normalize_observations')) policy = greedy_policy.GreedyPolicy( ppo_policy.PPOPolicy( time_step_spec=time_step_spec, action_spec=action_spec, actor_network=actor_net, value_network=value_net, observation_normalizer=self._observation_normalizer, clip=False, collect=False)) collect_policy = ppo_policy.PPOPolicy( time_step_spec=time_step_spec, action_spec=action_spec, actor_network=actor_net, value_network=value_net, observation_normalizer=self._observation_normalizer, clip=False, collect=True) self._action_distribution_spec = (self._actor_net.output_spec) super(PPOAgent, self).__init__(time_step_spec, action_spec, policy, collect_policy, train_sequence_length=None, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars, train_step_counter=train_step_counter)
def __init__(self, time_step_spec, action_spec, optimizer=None, actor_net=None, value_net=None, importance_ratio_clipping=0.0, lambda_value=0.95, discount_factor=0.99, entropy_regularization=0.0, policy_l2_reg=0.0, value_function_l2_reg=0.0, shared_vars_l2_reg=0.0, value_pred_loss_coef=0.5, num_epochs=25, use_gae=False, use_td_lambda_return=False, normalize_rewards=True, reward_norm_clipping=10.0, normalize_observations=True, log_prob_clipping=0.0, kl_cutoff_factor=2.0, kl_cutoff_coef=1000.0, initial_adaptive_kl_beta=1.0, adaptive_kl_target=0.01, adaptive_kl_tolerance=0.3, gradient_clipping=None, check_numerics=False, compute_value_and_advantage_in_train=False, debug_summaries=False, summarize_grads_and_vars=False, train_step_counter=None, name=None): """Creates a PPO Agent. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A nest of `BoundedTensorSpec` representing the actions. optimizer: Optimizer to use for the agent, default to using `tf.compat.v1.train.AdamOptimizer`. actor_net: A `network.DistributionNetwork` which maps observations to action distributions. Commonly, it is set to `actor_distribution_network.ActorDistributionNetwork`. value_net: A `Network` which returns the value prediction for input states, with `call(observation, step_type, network_state)`. Commonly, it is set to `value_network.ValueNetwork`. importance_ratio_clipping: Epsilon in clipped, surrogate PPO objective. For more detail, see explanation at the top of the doc. lambda_value: Lambda parameter for TD-lambda computation. discount_factor: Discount factor for return computation. Default to `0.99` which is the value used for all environments from (Schulman, 2017). entropy_regularization: Coefficient for entropy regularization loss term. Default to `0.0` because no entropy bonus was used in (Schulman, 2017). policy_l2_reg: Coefficient for L2 regularization of unshared actor_net weights. Default to `0.0` because no L2 regularization was applied on the policy network weights in (Schulman, 2017). value_function_l2_reg: Coefficient for l2 regularization of unshared value function weights. Default to `0.0` because no L2 regularization was applied on the policy network weights in (Schulman, 2017). shared_vars_l2_reg: Coefficient for l2 regularization of weights shared between actor_net and value_net. Default to `0.0` because no L2 regularization was applied on the policy network or value network weights in (Schulman, 2017). value_pred_loss_coef: Multiplier for value prediction loss to balance with policy gradient loss. Default to `0.5`, which was used for all environments in the OpenAI baseline implementation. This parameters is irrelevant unless you are sharing part of actor_net and value_net. In that case, you would want to tune this coeeficient, whose value depends on the network architecture of your choice. num_epochs: Number of epochs for computing policy updates. (Schulman,2017) sets this to 10 for Mujoco, 15 for Roboschool and 3 for Atari. use_gae: If True (default False), uses generalized advantage estimation for computing per-timestep advantage. Else, just subtracts value predictions from empirical return. use_td_lambda_return: If True (default False), uses td_lambda_return for training value function; here: `td_lambda_return = gae_advantage + value_predictions`. `use_gae` must be set to `True` as well to enable TD -lambda returns. If `use_td_lambda_return` is set to True while `use_gae` is False, the empirical return will be used and a warning will be logged. normalize_rewards: If true, keeps moving variance of rewards and normalizes incoming rewards. While not mentioned directly in (Schulman, 2017), reward normalization was implemented in OpenAI baselines and (Ilyas et al., 2018) pointed out that it largely improves performance. You may refer to Figure 1 of https://arxiv.org/pdf/1811.02553.pdf for a comparison with and without reward scaling. reward_norm_clipping: Value above and below to clip normalized reward. Additional optimization proposed in (Ilyas et al., 2018) set to `5` or `10`. normalize_observations: If `True`, keeps moving mean and variance of observations and normalizes incoming observations. Additional optimization proposed in (Ilyas et al., 2018). log_prob_clipping: +/- value for clipping log probs to prevent inf / NaN values. Default: no clipping. kl_cutoff_factor: Only meaningful when `kl_cutoff_coef > 0.0`. A multipler used for calculating the KL cutoff ( = `kl_cutoff_factor * adaptive_kl_target`). If policy KL averaged across the batch changes more than the cutoff, a squared cutoff loss would be added to the loss function. kl_cutoff_coef: kl_cutoff_coef and kl_cutoff_factor are additional params if one wants to use a KL cutoff loss term in addition to the adaptive KL loss term. Default to 0.0 to disable the KL cutoff loss term as this was not used in the paper. kl_cutoff_coef is the coefficient to mulitply by the KL cutoff loss term, before adding to the total loss function. initial_adaptive_kl_beta: Initial value for beta coefficient of adaptive KL penalty. This initial value is not important in practice because the algorithm quickly adjusts to it. A common default is 1.0. adaptive_kl_target: Desired KL target for policy updates. If actual KL is far from this target, adaptive_kl_beta will be updated. You should tune this for your environment. 0.01 was found to perform well for Mujoco. adaptive_kl_tolerance: A tolerance for adaptive_kl_beta. Mean KL above `(1 + tol) * adaptive_kl_target`, or below `(1 - tol) * adaptive_kl_target`, will cause `adaptive_kl_beta` to be updated. `0.5` was chosen heuristically in the paper, but the algorithm is not very sensitive to it. gradient_clipping: Norm length to clip gradients. Default: no clipping. check_numerics: If true, adds `tf.debugging.check_numerics` to help find NaN / Inf values. For debugging only. compute_value_and_advantage_in_train: A bool to indicate where value prediction and advantage calculation happen. If True, both happen in agent.train(). If False, value prediction is computed during data collection. This argument must be set to `False` if mini batch learning is enabled. debug_summaries: A bool to gather debug summaries. summarize_grads_and_vars: If true, gradient summaries will be written. train_step_counter: An optional counter to increment every time the train op is run. Defaults to the global_step. name: The name of this agent. All variables in this module will fall under that name. Defaults to the class name. Raises: ValueError: If the actor_net is not a DistributionNetwork or value_net is not a Network. """ if not isinstance(actor_net, network.DistributionNetwork): raise ValueError( 'actor_net must be an instance of a network.DistributionNetwork.') if not isinstance(value_net, network.Network): raise ValueError('value_net must be an instance of a network.Network.') actor_net.create_variables() value_net.create_variables() tf.Module.__init__(self, name=name) self._optimizer = optimizer self._actor_net = actor_net self._value_net = value_net self._importance_ratio_clipping = importance_ratio_clipping self._lambda = lambda_value self._discount_factor = discount_factor self._entropy_regularization = entropy_regularization self._policy_l2_reg = policy_l2_reg self._value_function_l2_reg = value_function_l2_reg self._shared_vars_l2_reg = shared_vars_l2_reg self._value_pred_loss_coef = value_pred_loss_coef self._num_epochs = num_epochs self._use_gae = use_gae self._use_td_lambda_return = use_td_lambda_return self._reward_norm_clipping = reward_norm_clipping self._log_prob_clipping = log_prob_clipping self._kl_cutoff_factor = kl_cutoff_factor self._kl_cutoff_coef = kl_cutoff_coef self._adaptive_kl_target = adaptive_kl_target self._adaptive_kl_tolerance = adaptive_kl_tolerance self._gradient_clipping = gradient_clipping or 0.0 self._check_numerics = check_numerics self._compute_value_and_advantage_in_train = ( compute_value_and_advantage_in_train) if initial_adaptive_kl_beta > 0.0: # TODO(kbanoop): Rename create_variable. self._adaptive_kl_beta = common.create_variable( 'adaptive_kl_beta', initial_adaptive_kl_beta, dtype=tf.float32) else: self._adaptive_kl_beta = None self._reward_normalizer = None if normalize_rewards: self._reward_normalizer = tensor_normalizer.StreamingTensorNormalizer( tensor_spec.TensorSpec([], tf.float32), scope='normalize_reward') self._observation_normalizer = None if normalize_observations: self._observation_normalizer = ( tensor_normalizer.StreamingTensorNormalizer( time_step_spec.observation, scope='normalize_observations')) policy = greedy_policy.GreedyPolicy( ppo_policy.PPOPolicy( time_step_spec=time_step_spec, action_spec=action_spec, actor_network=actor_net, value_network=value_net, observation_normalizer=self._observation_normalizer, clip=False, collect=False)) collect_policy = ppo_policy.PPOPolicy( time_step_spec=time_step_spec, action_spec=action_spec, actor_network=actor_net, value_network=value_net, observation_normalizer=self._observation_normalizer, clip=False, collect=True, compute_value_and_advantage_in_train=( self._compute_value_and_advantage_in_train), ) self._action_distribution_spec = (self._actor_net.output_spec) super(PPOAgent, self).__init__( time_step_spec, action_spec, policy, collect_policy, train_sequence_length=None, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars, train_step_counter=train_step_counter)
def testPolicyStepWithActionMaskTurnedOn(self): # Creat specs with action constraints (mask). num_categories = 5 observation_tensor_spec = ( tensor_spec.TensorSpec(shape=(3, ), dtype=tf.int64, name='network_spec'), tensor_spec.TensorSpec(shape=(num_categories, ), dtype=tf.bool, name='mask_spec'), ) network_spec, _ = observation_tensor_spec action_tensor_spec = tensor_spec.BoundedTensorSpec((1, ), tf.int32, 0, num_categories - 1) # Create policy with splitter. def splitter_fn(observation_and_mask): return observation_and_mask[0], observation_and_mask[1] actor_network = mask_splitter_network.MaskSplitterNetwork( splitter_fn, actor_distribution_network.ActorDistributionNetwork( network_spec, action_tensor_spec), passthrough_mask=True) value_network = mask_splitter_network.MaskSplitterNetwork( splitter_fn, value_net.ValueNetwork(network_spec)) policy = ppo_policy.PPOPolicy( ts.time_step_spec(observation_tensor_spec), action_tensor_spec, actor_network=actor_network, value_network=value_network, clip=False) # Take a step. mask = np.array([True, False, True, False, True], dtype=np.bool) self.assertLen(mask, num_categories) time_step = ts.TimeStep(step_type=tf.constant([1], dtype=tf.int32), reward=tf.constant([1], dtype=tf.float32), discount=tf.constant([1], dtype=tf.float32), observation=(tf.constant( [[1, 2, 3], [4, 5, 6]], dtype=tf.int64), tf.constant([mask.tolist()], dtype=tf.bool))) action_step = policy.action(time_step) # Check the shape and type of the resulted action step. self.assertEqual(action_step.action.shape.as_list(), [2, 1]) self.assertEqual(action_step.action.dtype, tf.int32) self.evaluate(tf.compat.v1.global_variables_initializer()) # Check the actions in general and with respect to masking. actions = self.evaluate(action_step.action) self.assertTrue(np.all(actions >= action_tensor_spec.minimum)) self.assertTrue(np.all(actions <= action_tensor_spec.maximum)) # Check the logits. logits = np.array(self.evaluate( action_step.info['dist_params']['logits']), dtype=np.float32) masked_actions = np.array(range(len(mask)))[~mask] self.assertTrue( np.all(logits[:, :, masked_actions] == np.finfo(np.float32).min)) valid_actions = np.array(range(len(mask)))[mask] self.assertTrue( np.all(logits[:, :, valid_actions] > np.finfo(np.float32).min))