def testSimple(self): converter = data_converter.AsTrajectory(self._data_context) traj = tensor_spec.sample_spec_nest(self._data_context.trajectory_spec, outer_dims=[2, 3]) converted = converter(traj) (traj, converted) = self.evaluate((traj, converted)) tf.nest.map_structure(self.assertAllEqual, converted, traj)
def testFromBatchTimeTransition(self): converter = data_converter.AsTrajectory(self._data_context) traj = tensor_spec.sample_spec_nest(self._data_context.trajectory_spec, outer_dims=[2, 3]) transition = trajectory.to_transition(traj, traj) converted = converter(transition) (traj, converted) = self.evaluate((traj, converted)) tf.nest.map_structure(self.assertAllEqual, converted, traj)
def testNoTimeDimensionRaises(self): converter = data_converter.AsTrajectory(self._data_context) traj = tensor_spec.sample_spec_nest(self._data_context.trajectory_spec, outer_dims=[3]) with self.assertRaisesRegex( ValueError, r'must have two outer dimensions: batch size and time'): converter(traj)
def testNoTimeDimensionRaises(self): converter = data_converter.AsTrajectory(self._data_context) traj = tensor_spec.sample_spec_nest(self._data_context.trajectory_spec, outer_dims=[3]) with self.assertRaisesRegex( ValueError, r'tensors must have shape \`\[B, T\] \+ spec.shape\`'): converter(traj)
def testInvalidTimeDimensionRaises(self): converter = data_converter.AsTrajectory(self._data_context, sequence_length=4) traj = tensor_spec.sample_spec_nest(self._data_context.trajectory_spec, outer_dims=[2, 3]) with self.assertRaisesRegex( ValueError, r'has a time axis dim value \'3\' vs the expected \'4\''): converter(traj)
def testPrunes(self): converter = data_converter.AsTrajectory(self._data_context) my_spec = self._data_context.trajectory_spec.replace( action={ 'action1': tf.TensorSpec((), tf.float32), 'action2': tf.TensorSpec([4], tf.int32) }) traj = tensor_spec.sample_spec_nest(my_spec, outer_dims=[2, 3]) converted = converter(traj) expected = tf.nest.map_structure(lambda x: x, traj) del expected.action['action2'] (expected, converted) = self.evaluate((expected, converted)) tf.nest.map_structure(self.assertAllEqual, converted, expected)
def __init__(self, time_step_spec=None, action_spec=None, training_data_spec=None, train_sequence_length=None): if time_step_spec is None: obs_spec = {'obs': tf.TensorSpec([], tf.float32)} time_step_spec = ts.time_step_spec(obs_spec) action_spec = action_spec or () policy = random_tf_policy.RandomTFPolicy(time_step_spec, action_spec) super(MyAgent, self).__init__(time_step_spec=time_step_spec, action_spec=action_spec, policy=policy, collect_policy=policy, train_sequence_length=train_sequence_length, training_data_spec=training_data_spec) self._as_trajectory = data_converter.AsTrajectory( self.data_context, sequence_length=train_sequence_length)
def __init__(self, mixture_distribution: types.Distribution, agents: Sequence[tf_agent.TFAgent], name: Optional[Text] = None): """Initializes an instance of `MixtureAgent`. Args: mixture_distribution: An instance of `tfd.Categorical` distribution. This distribution is used to draw sub-policies by the mixture policy. The parameters of the distribution is trained by the mixture agent. agents: List of instances of TF-Agents bandit agents. These agents will be trained and used to select actions. The length of this list should match that of `mixture_weights`. name: The name of this instance of `MixtureAgent`. """ tf.Module.__init__(self, name=name) time_step_spec = agents[0].time_step_spec action_spec = agents[0].action_spec self._original_info_spec = agents[0].policy.info_spec error_message = None for agent in agents[1:]: if action_spec != agent.action_spec: error_message = 'Inconsistent action specs.' if time_step_spec != agent.time_step_spec: error_message = 'Inconsistent time step specs.' if self._original_info_spec != agent.policy.info_spec: error_message = 'Inconsistent info specs.' if error_message is not None: raise ValueError(error_message) self._agents = agents self._num_agents = len(agents) self._mixture_distribution = mixture_distribution policies = [agent.collect_policy for agent in agents] policy = mixture_policy.MixturePolicy(mixture_distribution, policies) super(MixtureAgent, self).__init__(time_step_spec, action_spec, policy, policy, train_sequence_length=None) self._as_trajectory = data_converter.AsTrajectory(self.data_context, sequence_length=None)
def __init__(self, time_step_spec: types.TimeStep, action_spec: types.BoundedTensorSpec, learning_rate: float, name: Optional[Text] = None): """Initialize an instance of `Exp3Agent`. Args: time_step_spec: A `TimeStep` spec describing the expected `TimeStep`s. action_spec: A scalar `BoundedTensorSpec` with `int32` or `int64` dtype describing the number of actions for this agent. learning_rate: A float valued scalar. A higher value will force the agent to converge on a single action more quickly. A lower value will encourage more exploration. This value corresponds to the `inverse_temperature` argument passed to `CategoricalPolicy`. name: a name for this instance of `Exp3Agent`. """ tf.Module.__init__(self, name=name) common.tf_agents_gauge.get_cell('TFABandit').set(True) self._num_actions = policy_utilities.get_num_actions_from_tensor_spec( action_spec) self._weights = tf.compat.v2.Variable(tf.zeros(self._num_actions), name='weights') self._learning_rate = tf.compat.v2.Variable(learning_rate, name='learning_rate') policy = categorical_policy.CategoricalPolicy( weights=self._weights, time_step_spec=time_step_spec, action_spec=action_spec, inverse_temperature=self._learning_rate) # TODO(b/127462472): consider policy=GreedyPolicy(collect_policy). super(Exp3Agent, self).__init__(time_step_spec=time_step_spec, action_spec=policy.action_spec, policy=policy, collect_policy=policy, train_sequence_length=None, validate_args=False) self._as_trajectory = data_converter.AsTrajectory(self.data_context, sequence_length=None)
def __init__(self, time_step_spec: types.TimeStep, action_spec: types.BoundedTensorSpec, variable_collection: Optional[ BernoulliBanditVariableCollection] = None, dtype: tf.DType = tf.float32, batch_size: Optional[int] = 1, observation_and_action_constraint_splitter: Optional[ types.Splitter] = None, emit_policy_info: Sequence[Text] = (), name: Optional[Text] = None): """Creates a Bernoulli Thompson Sampling Agent. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A nest of `BoundedTensorSpec` representing the actions. variable_collection: Instance of `BernoulliBanditVariableCollection`. Collection of variables to be updated by the agent. If `None`, a new instance of `BernoulliBanditVariableCollection` will be created. dtype: The type of the variables. Should be one of `tf.float32` or `tf.float64`. batch_size: optional int with the batch size. It defaults to 1. observation_and_action_constraint_splitter: A function used for masking valid/invalid actions with each state of the environment. The function takes in a full observation and returns a tuple consisting of 1) the part of the observation intended as input to the bandit agent and policy, and 2) the boolean mask. This function should also work with a `TensorSpec` as input, and should output `TensorSpec` objects for the observation and mask. emit_policy_info: (tuple of strings) what side information we want to get as part of the policy info. Allowed values can be found in `policy_utilities.PolicyInfo`. name: Python str name of this agent. All variables in this module will fall under that name. Defaults to the class name. Raises: ValueError: If the action spec contains more than one action or it is not a bounded scalar int32 spec with minimum 0. TypeError: if variable_collection is not an instance of `BernoulliBanditVariableCollection`. """ tf.Module.__init__(self, name=name) common.tf_agents_gauge.get_cell('TFABandit').set(True) self._observation_and_action_constraint_splitter = ( observation_and_action_constraint_splitter) self._num_actions = policy_utilities.get_num_actions_from_tensor_spec( action_spec) self._dtype = dtype if variable_collection is None: variable_collection = BernoulliBanditVariableCollection( num_actions=self._num_actions, dtype=dtype) elif not isinstance(variable_collection, BernoulliBanditVariableCollection): raise TypeError('Parameter `variable_collection` should be ' 'of type `BernoulliBanditVariableCollection`.') self._variable_collection = variable_collection self._alpha = variable_collection.alpha self._beta = variable_collection.beta self._batch_size = batch_size policy = bernoulli_policy.BernoulliThompsonSamplingPolicy( time_step_spec, action_spec, self._alpha, self._beta, observation_and_action_constraint_splitter, emit_policy_info=emit_policy_info) super(BernoulliThompsonSamplingAgent, self).__init__(time_step_spec, action_spec, policy, collect_policy=policy, train_sequence_length=None) self._as_trajectory = data_converter.AsTrajectory(self.data_context, sequence_length=None)
def __init__(self, exploration_policy, time_step_spec: types.TimeStep, action_spec: types.BoundedTensorSpec, variable_collection: Optional[ LinearBanditVariableCollection] = None, alpha: float = 1.0, gamma: float = 1.0, use_eigendecomp: bool = False, tikhonov_weight: float = 1.0, add_bias: bool = False, emit_policy_info: Sequence[Text] = (), emit_log_probability: bool = False, observation_and_action_constraint_splitter: Optional[ types.Splitter] = None, accepts_per_arm_features: bool = False, debug_summaries: bool = False, summarize_grads_and_vars: bool = False, enable_summaries: bool = True, dtype: tf.DType = tf.float32, name: Optional[Text] = None): """Initialize an instance of `LinearBanditAgent`. Args: exploration_policy: An Enum of type `ExplorationPolicy`. The kind of policy we use for exploration. Currently supported policies are `LinUCBPolicy` and `LinearThompsonSamplingPolicy`. time_step_spec: A `TimeStep` spec describing the expected `TimeStep`s. action_spec: A scalar `BoundedTensorSpec` with `int32` or `int64` dtype describing the number of actions for this agent. variable_collection: Instance of `LinearBanditVariableCollection`. Collection of variables to be updated by the agent. If `None`, a new instance of `LinearBanditVariableCollection` will be created. alpha: (float) positive scalar. This is the exploration parameter that multiplies the confidence intervals. gamma: a float forgetting factor in [0.0, 1.0]. When set to 1.0, the algorithm does not forget. use_eigendecomp: whether to use eigen-decomposition or not. The default solver is Conjugate Gradient. tikhonov_weight: (float) tikhonov regularization term. add_bias: If true, a bias term will be added to the linear reward estimation. emit_policy_info: (tuple of strings) what side information we want to get as part of the policy info. Allowed values can be found in `policy_utilities.PolicyInfo`. emit_log_probability: Whether the policy emits log-probabilities or not. Since the policy is deterministic, the probability is just 1. observation_and_action_constraint_splitter: A function used for masking valid/invalid actions with each state of the environment. The function takes in a full observation and returns a tuple consisting of 1) the part of the observation intended as input to the bandit agent and policy, and 2) the boolean mask. This function should also work with a `TensorSpec` as input, and should output `TensorSpec` objects for the observation and mask. accepts_per_arm_features: (bool) Whether the agent accepts per-arm features. debug_summaries: A Python bool, default False. When True, debug summaries are gathered. summarize_grads_and_vars: A Python bool, default False. When True, gradients and network variable summaries are written during training. enable_summaries: A Python bool, default True. When False, all summaries (debug or otherwise) should not be written. dtype: The type of the parameters stored and updated by the agent. Should be one of `tf.float32` and `tf.float64`. Defaults to `tf.float32`. name: a name for this instance of `LinearBanditAgent`. Raises: ValueError if dtype is not one of `tf.float32` or `tf.float64`. TypeError if variable_collection is not an instance of `LinearBanditVariableCollection`. """ tf.Module.__init__(self, name=name) common.tf_agents_gauge.get_cell('TFABandit').set(True) self._num_actions = policy_utilities.get_num_actions_from_tensor_spec( action_spec) self._num_models = 1 if accepts_per_arm_features else self._num_actions self._observation_and_action_constraint_splitter = ( observation_and_action_constraint_splitter) self._time_step_spec = time_step_spec self._accepts_per_arm_features = accepts_per_arm_features self._add_bias = add_bias if observation_and_action_constraint_splitter is not None: context_spec, _ = observation_and_action_constraint_splitter( time_step_spec.observation) else: context_spec = time_step_spec.observation (self._global_context_dim, self._arm_context_dim) = bandit_spec_utils.get_context_dims_from_spec( context_spec, accepts_per_arm_features) if self._add_bias: # The bias is added via a constant 1 feature. self._global_context_dim += 1 self._overall_context_dim = self._global_context_dim + self._arm_context_dim self._alpha = alpha if variable_collection is None: variable_collection = LinearBanditVariableCollection( context_dim=self._overall_context_dim, num_models=self._num_models, use_eigendecomp=use_eigendecomp, dtype=dtype) elif not isinstance(variable_collection, LinearBanditVariableCollection): raise TypeError('Parameter `variable_collection` should be ' 'of type `LinearBanditVariableCollection`.') self._variable_collection = variable_collection self._cov_matrix_list = variable_collection.cov_matrix_list self._data_vector_list = variable_collection.data_vector_list self._eig_matrix_list = variable_collection.eig_matrix_list self._eig_vals_list = variable_collection.eig_vals_list # We keep track of the number of samples per arm. self._num_samples_list = variable_collection.num_samples_list self._gamma = gamma if self._gamma < 0.0 or self._gamma > 1.0: raise ValueError( 'Forgetting factor `gamma` must be in [0.0, 1.0].') self._dtype = dtype if dtype not in (tf.float32, tf.float64): raise ValueError( 'Agent dtype should be either `tf.float32 or `tf.float64`.') self._use_eigendecomp = use_eigendecomp self._tikhonov_weight = tikhonov_weight if exploration_policy == ExplorationPolicy.linear_ucb_policy: exploration_strategy = lin_policy.ExplorationStrategy.optimistic elif exploration_policy == ( ExplorationPolicy.linear_thompson_sampling_policy): exploration_strategy = lin_policy.ExplorationStrategy.sampling else: raise ValueError( 'Linear bandit agent with policy %s not implemented' % exploration_policy) policy = lin_policy.LinearBanditPolicy( action_spec=action_spec, cov_matrix=self._cov_matrix_list, data_vector=self._data_vector_list, num_samples=self._num_samples_list, time_step_spec=time_step_spec, exploration_strategy=exploration_strategy, alpha=alpha, eig_vals=self._eig_vals_list if self._use_eigendecomp else (), eig_matrix=self._eig_matrix_list if self._use_eigendecomp else (), tikhonov_weight=self._tikhonov_weight, add_bias=add_bias, emit_policy_info=emit_policy_info, emit_log_probability=emit_log_probability, accepts_per_arm_features=accepts_per_arm_features, observation_and_action_constraint_splitter=( observation_and_action_constraint_splitter)) training_data_spec = None if accepts_per_arm_features: training_data_spec = bandit_spec_utils.drop_arm_observation( policy.trajectory_spec) super(LinearBanditAgent, self).__init__(time_step_spec=time_step_spec, action_spec=action_spec, policy=policy, collect_policy=policy, training_data_spec=training_data_spec, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars, enable_summaries=enable_summaries, train_sequence_length=None) self._as_trajectory = data_converter.AsTrajectory(self.data_context, sequence_length=None)
def __init__( self, time_step_spec: Optional[ts.TimeStep], action_spec: Optional[types.NestedBoundedTensorSpec], scalarizer: multi_objective_scalarizer.Scalarizer, objective_network_and_loss_fn_sequence: Sequence[Tuple[ Network, Callable[..., tf.Tensor]]], optimizer: tf.keras.optimizers.Optimizer, observation_and_action_constraint_splitter: types.Splitter = None, accepts_per_arm_features: bool = False, # Params for training. gradient_clipping: Optional[float] = None, # Params for debugging. debug_summaries: bool = False, summarize_grads_and_vars: bool = False, enable_summaries: bool = True, emit_policy_info: Tuple[Text] = (), train_step_counter: Optional[tf.Variable] = None, name: Optional[Text] = None): """Creates a Greedy Multi-objective Neural Agent. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A nest of `BoundedTensorSpec` representing the actions. scalarizer: A `tf_agents.bandits.multi_objective.multi_objective_scalarizer.Scalarizer` object that implements scalarization of multiple objectives into a single scalar reward. objective_network_and_loss_fn_sequence: A Sequence of Tuples (`tf_agents.network.Network`, error loss function) to be used by the agent. Each network `net` will be called as `net(observation, training=...)` and is expected to output a `tf.Tensor` of predicted values for a specific objective for all actions, shaped as [batch-size, number-of-actions]. Each network will be trained via minimizing the accompanying error loss function, which takes parameters labels, predictions, and weights (any function from tf.losses would work). optimizer: A 'tf.keras.optimizers.Optimizer' object, the optimizer to use for training. observation_and_action_constraint_splitter: A function used for masking valid/invalid actions with each state of the environment. The function takes in a full observation and returns a tuple consisting of 1) the part of the observation intended as input to the bandit agent and policy, and 2) the boolean mask of shape `[batch_size, num_actions]`. This function should also work with a `TensorSpec` as input, and should output `TensorSpec` objects for the observation and mask. accepts_per_arm_features: (bool) Whether the agent accepts per-arm features. gradient_clipping: A float representing the norm length to clip gradients (or None for no clipping.) debug_summaries: A Python bool, default False. When True, debug summaries are gathered. summarize_grads_and_vars: A Python bool, default False. When True, gradients and network variable summaries are written during training. enable_summaries: A Python bool, default True. When False, all summaries (debug or otherwise) should not be written. emit_policy_info: (tuple of strings) what side information we want to get as part of the policy info. Allowed values can be found in `policy_utilities.PolicyInfo`. train_step_counter: An optional `tf.Variable` to increment every time the train op is run. Defaults to the `global_step`. name: Python str name of this agent. All variables in this module will fall under that name. Defaults to the class name. Raises: ValueError: - If the action spec contains more than one action or or it is not a bounded scalar int32 spec with minimum 0. - If the length of `objective_network_and_loss_fn_sequence` is less than two. """ tf.Module.__init__(self, name=name) common.tf_agents_gauge.get_cell('TFABandit').set(True) self._observation_and_action_constraint_splitter = ( observation_and_action_constraint_splitter) self._num_actions = policy_utilities.get_num_actions_from_tensor_spec( action_spec) self._accepts_per_arm_features = accepts_per_arm_features self._num_objectives = len(objective_network_and_loss_fn_sequence) if self._num_objectives < 2: raise ValueError( 'Number of objectives should be at least two, but found to be {}' .format(self._num_objectives)) self._objective_networks, self._error_loss_fns = tuple( zip(*objective_network_and_loss_fn_sequence)) self._optimizer = optimizer self._gradient_clipping = gradient_clipping self._heteroscedastic = [ isinstance(network, heteroscedastic_q_network.HeteroscedasticQNetwork) for network in self._objective_networks ] policy = greedy_multi_objective_policy.GreedyMultiObjectiveNeuralPolicy( time_step_spec, action_spec, scalarizer, self._objective_networks, observation_and_action_constraint_splitter, accepts_per_arm_features=accepts_per_arm_features, emit_policy_info=emit_policy_info) training_data_spec = None if accepts_per_arm_features: training_data_spec = bandit_spec_utils.drop_arm_observation( policy.trajectory_spec) super(GreedyMultiObjectiveNeuralAgent, self).__init__(time_step_spec, action_spec, policy, collect_policy=policy, train_sequence_length=None, training_data_spec=training_data_spec, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars, enable_summaries=enable_summaries, train_step_counter=train_step_counter, validate_args=False) self._as_trajectory = data_converter.AsTrajectory(self.data_context, sequence_length=None)
def __init__( self, time_step_spec: ts.TimeStep, action_spec: types.NestedTensorSpec, cloning_network: network.Network, optimizer: types.Optimizer, num_outer_dims: Literal[1, 2] = 1, # pylint: disable=bad-whitespace epsilon_greedy: types.Float = 0.1, loss_fn: Optional[Callable[[types.NestedTensor, bool], types.Tensor]] = None, gradient_clipping: Optional[types.Float] = None, # Params for debugging. debug_summaries: bool = False, summarize_grads_and_vars: bool = False, train_step_counter: Optional[tf.Variable] = None, name: Optional[Text] = None): """Creates an instance of a Behavioral Cloning agent. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A nest of BoundedTensorSpec representing the actions. cloning_network: A `tf_agents.networks.Network` to be used by the agent. The network will be called as ``` network(observation, step_type=step_type, network_state=initial_state) ``` and must return a 2-tuple with elements `(output, next_network_state)` optimizer: The optimizer to use for training. num_outer_dims: The number of outer dimensions for the agent. Must be either 1 or 2. If 2, training will require both a batch_size and time dimension on every Tensor; if 1, training will require only a batch_size outer dimension. epsilon_greedy: probability of choosing a random action in the default epsilon-greedy collect policy (used only if actions are discrete) loss_fn: A function for computing the error between the output of the cloning network and the action that was taken. If None, the loss depends on the action dtype. The `loss_fn` is called with parameters: `(experience, training)`, and must return a loss value for each element of the batch. gradient_clipping: Norm length to clip gradients. debug_summaries: A bool to gather debug summaries. summarize_grads_and_vars: If True, gradient and network variable summaries will be written during training. 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. """ tf.Module.__init__(self, name=name) self._cloning_network = cloning_network self._optimizer = optimizer self._gradient_clipping = gradient_clipping action_spec = tensor_spec.from_spec(action_spec) flat_action_spec = tf.nest.flatten(action_spec) continuous_specs = [ tensor_spec.is_continuous(s) for s in flat_action_spec ] if not flat_action_spec: raise ValueError( 'The `action_spec` must contain at least one action.') single_discrete_scalar_action = ( len(flat_action_spec) == 1 and flat_action_spec[0].shape.rank == 0 and not tensor_spec.is_continuous(flat_action_spec[0])) single_continuous_action = (len(flat_action_spec) == 1 and tensor_spec.is_continuous( flat_action_spec[0])) if (not loss_fn and not single_discrete_scalar_action and not single_continuous_action): raise ValueError( 'A `loss_fn` must be provided unless there is a single, scalar ' 'discrete action or a single (scalar or non-scalar) continuous ' 'action.') self._network_output_spec = cloning_network.create_variables( time_step_spec.observation) # If there is a mix of continuous and discrete actions we want to use an # actor policy so we can use the `setup_as_continuous` method as long as the # user provided a custom loss_fn which we verified above. if any(continuous_specs): policy, collect_policy = self._setup_as_continuous( time_step_spec, action_spec, loss_fn) else: policy, collect_policy = self._setup_as_discrete( time_step_spec, action_spec, loss_fn, epsilon_greedy) super(BehavioralCloningAgent, 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) self._as_trajectory = data_converter.AsTrajectory( self.data_context, sequence_length=None, num_outer_dims=num_outer_dims)
def __init__( self, time_step_spec: types.TimeStep, action_spec: types.BoundedTensorSpec, encoding_network: types.Network, encoding_network_num_train_steps: int, encoding_dim: int, optimizer: types.Optimizer, variable_collection: Optional[NeuralLinUCBVariableCollection] = None, alpha: float = 1.0, gamma: float = 1.0, epsilon_greedy: float = 0.0, observation_and_action_constraint_splitter: Optional[ types.Splitter] = None, accepts_per_arm_features: bool = False, distributed_train_encoding_network: bool = False, # Params for training. error_loss_fn: types.LossFn = tf.compat.v1.losses.mean_squared_error, gradient_clipping: Optional[float] = None, # Params for debugging. debug_summaries: bool = False, summarize_grads_and_vars: bool = False, train_step_counter: Optional[tf.Variable] = None, emit_policy_info: Sequence[Text] = (), emit_log_probability: bool = False, dtype: tf.DType = tf.float64, name: Optional[Text] = None): """Initialize an instance of `NeuralLinUCBAgent`. Args: time_step_spec: A `TimeStep` spec describing the expected `TimeStep`s. action_spec: A scalar `BoundedTensorSpec` with `int32` or `int64` dtype describing the number of actions for this agent. encoding_network: a Keras network that encodes the observations. encoding_network_num_train_steps: how many training steps to run for training the encoding network before switching to LinUCB. If negative, the encoding network is assumed to be already trained. encoding_dim: the dimension of encoded observations. optimizer: The optimizer to use for training. variable_collection: Instance of `NeuralLinUCBVariableCollection`. Collection of variables to be updated by the agent. If `None`, a new instance of `LinearBanditVariables` will be created. Note that this collection excludes the variables owned by the encoding network. alpha: (float) positive scalar. This is the exploration parameter that multiplies the confidence intervals. gamma: a float forgetting factor in [0.0, 1.0]. When set to 1.0, the algorithm does not forget. epsilon_greedy: A float representing the probability of choosing a random action instead of the greedy action. observation_and_action_constraint_splitter: A function used for masking valid/invalid actions with each state of the environment. The function takes in a full observation and returns a tuple consisting of 1) the part of the observation intended as input to the bandit agent and policy, and 2) the boolean mask. This function should also work with a `TensorSpec` as input, and should output `TensorSpec` objects for the observation and mask. accepts_per_arm_features: (bool) Whether the policy accepts per-arm features. distributed_train_encoding_network: (bool) whether to train the encoding network or not. This applies only in distributed training setting. When set to true this agent will train the encoding network. Otherwise, it will assume the encoding network is already trained and will train LinUCB on top of it. error_loss_fn: A function for computing the error loss, taking parameters labels, predictions, and weights (any function from tf.losses would work). The default is `tf.losses.mean_squared_error`. gradient_clipping: A float representing the norm length to clip gradients (or None for no clipping.) debug_summaries: A Python bool, default False. When True, debug summaries are gathered. summarize_grads_and_vars: A Python bool, default False. When True, gradients and network variable summaries are written during training. train_step_counter: An optional `tf.Variable` to increment every time the train op is run. Defaults to the `global_step`. emit_policy_info: (tuple of strings) what side information we want to get as part of the policy info. Allowed values can be found in `policy_utilities.PolicyInfo`. emit_log_probability: Whether the NeuralLinUCBPolicy emits log-probabilities or not. Since the policy is deterministic, the probability is just 1. dtype: The type of the parameters stored and updated by the agent. Should be one of `tf.float32` and `tf.float64`. Defaults to `tf.float64`. name: a name for this instance of `NeuralLinUCBAgent`. Raises: TypeError if variable_collection is not an instance of `NeuralLinUCBVariableCollection`. ValueError if dtype is not one of `tf.float32` or `tf.float64`. """ tf.Module.__init__(self, name=name) common.tf_agents_gauge.get_cell('TFABandit').set(True) self._num_actions = policy_utilities.get_num_actions_from_tensor_spec( action_spec) self._num_models = 1 if accepts_per_arm_features else self._num_actions self._observation_and_action_constraint_splitter = ( observation_and_action_constraint_splitter) self._accepts_per_arm_features = accepts_per_arm_features self._alpha = alpha if variable_collection is None: variable_collection = NeuralLinUCBVariableCollection( self._num_models, encoding_dim, dtype) elif not isinstance(variable_collection, NeuralLinUCBVariableCollection): raise TypeError('Parameter `variable_collection` should be ' 'of type `NeuralLinUCBVariableCollection`.') self._variable_collection = variable_collection self._gamma = gamma if self._gamma < 0.0 or self._gamma > 1.0: raise ValueError('Forgetting factor `gamma` must be in [0.0, 1.0].') self._dtype = dtype if dtype not in (tf.float32, tf.float64): raise ValueError( 'Agent dtype should be either `tf.float32 or `tf.float64`.') self._epsilon_greedy = epsilon_greedy reward_layer = tf.keras.layers.Dense( self._num_models, kernel_initializer=tf.random_uniform_initializer( minval=-0.03, maxval=0.03), use_bias=False, activation=None, name='reward_layer') encoding_network.create_variables() self._encoding_network = encoding_network reward_layer.build(input_shape=tf.TensorShape([None, encoding_dim])) self._reward_layer = reward_layer self._encoding_network_num_train_steps = encoding_network_num_train_steps self._encoding_dim = encoding_dim self._optimizer = optimizer self._error_loss_fn = error_loss_fn self._gradient_clipping = gradient_clipping train_step_counter = tf.compat.v1.train.get_or_create_global_step() self._distributed_train_encoding_network = ( distributed_train_encoding_network) policy = neural_linucb_policy.NeuralLinUCBPolicy( encoding_network=self._encoding_network, encoding_dim=self._encoding_dim, reward_layer=self._reward_layer, epsilon_greedy=self._epsilon_greedy, actions_from_reward_layer=self.actions_from_reward_layer, cov_matrix=self.cov_matrix, data_vector=self.data_vector, num_samples=self.num_samples, time_step_spec=time_step_spec, alpha=alpha, emit_policy_info=emit_policy_info, emit_log_probability=emit_log_probability, accepts_per_arm_features=accepts_per_arm_features, distributed_use_reward_layer=distributed_train_encoding_network, observation_and_action_constraint_splitter=( observation_and_action_constraint_splitter)) training_data_spec = None if accepts_per_arm_features: training_data_spec = bandit_spec_utils.drop_arm_observation( policy.trajectory_spec) super(NeuralLinUCBAgent, self).__init__( time_step_spec=time_step_spec, action_spec=policy.action_spec, policy=policy, collect_policy=policy, train_sequence_length=None, training_data_spec=training_data_spec, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars, train_step_counter=train_step_counter, validate_args=False) self._as_trajectory = data_converter.AsTrajectory( self.data_context, sequence_length=None)
def __init__( self, time_step_spec: types.TimeStep, action_spec: types.BoundedTensorSpec, reward_network: types.Network, optimizer: types.Optimizer, observation_and_action_constraint_splitter: Optional[ types.Splitter] = None, accepts_per_arm_features: bool = False, constraints: Iterable[constr.BaseConstraint] = (), # Params for training. error_loss_fn: types.LossFn = tf.compat.v1.losses.mean_squared_error, gradient_clipping: Optional[float] = None, # Params for debugging. debug_summaries: bool = False, summarize_grads_and_vars: bool = False, enable_summaries: bool = True, emit_policy_info: Tuple[Text, ...] = (), train_step_counter: Optional[tf.Variable] = None, laplacian_matrix: Optional[types.Float] = None, laplacian_smoothing_weight: float = 0.001, name: Optional[Text] = None): """Creates a Greedy Reward Network Prediction Agent. In some use cases, the actions are not independent and they are related to each other (e.g., when the actions are ordinal integers). Assuming that the relations between arms can be modeled by a graph, we may want to enforce that the estimated reward function is smooth over the graph. This implies that the estimated rewards `r_i` and `r_j` for two related actions `i` and `j`, should be close to each other. To quantify this smoothness criterion we use the Laplacian matrix `L` of the graph over the actions. When the laplacian smoothing is enabled, the loss is extended to: ``` Loss_new := Loss + lambda r^T * L * r, ``` where `r` is the estimated reward vector for all actions. The second term is the laplacian smoothing regularization term and `lambda` is the weight that determines how strongly we enforce the regularization. For more details, please see: "Bandits on graphs and structures", Michal Valko https://hal.inria.fr/tel-01359757/document Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A nest of `BoundedTensorSpec` representing the actions. reward_network: A `tf_agents.network.Network` to be used by the agent. The network will be called with call(observation, step_type) and it is expected to provide a reward prediction for all actions. optimizer: The optimizer to use for training. observation_and_action_constraint_splitter: A function used for masking valid/invalid actions with each state of the environment. The function takes in a full observation and returns a tuple consisting of 1) the part of the observation intended as input to the bandit agent and policy, and 2) the boolean mask. This function should also work with a `TensorSpec` as input, and should output `TensorSpec` objects for the observation and mask. accepts_per_arm_features: (bool) Whether the policy accepts per-arm features. constraints: iterable of constraints objects that are instances of `tf_agents.bandits.agents.NeuralConstraint`. error_loss_fn: A function for computing the error loss, taking parameters labels, predictions, and weights (any function from tf.losses would work). The default is `tf.losses.mean_squared_error`. gradient_clipping: A float representing the norm length to clip gradients (or None for no clipping.) debug_summaries: A Python bool, default False. When True, debug summaries are gathered. summarize_grads_and_vars: A Python bool, default False. When True, gradients and network variable summaries are written during training. enable_summaries: A Python bool, default True. When False, all summaries (debug or otherwise) should not be written. emit_policy_info: (tuple of strings) what side information we want to get as part of the policy info. Allowed values can be found in `policy_utilities.PolicyInfo`. train_step_counter: An optional `tf.Variable` to increment every time the train op is run. Defaults to the `global_step`. laplacian_matrix: A float `Tensor` or a numpy array shaped `[num_actions, num_actions]`. This holds the Laplacian matrix used to regularize the smoothness of the estimated expected reward function. This only applies to problems where the actions have a graph structure. If `None`, the regularization is not applied. laplacian_smoothing_weight: A float that determines the weight of the regularization term. Note that this has no effect if `laplacian_matrix` above is `None`. name: Python str name of this agent. All variables in this module will fall under that name. Defaults to the class name. Raises: ValueError: If the action spec contains more than one action or or it is not a bounded scalar int32 spec with minimum 0. InvalidArgumentError: if the Laplacian provided is not None and not valid. """ tf.Module.__init__(self, name=name) common.tf_agents_gauge.get_cell('TFABandit').set(True) self._observation_and_action_constraint_splitter = ( observation_and_action_constraint_splitter) self._num_actions = policy_utilities.get_num_actions_from_tensor_spec( action_spec) self._accepts_per_arm_features = accepts_per_arm_features self._constraints = constraints reward_network.create_variables() self._reward_network = reward_network self._optimizer = optimizer self._error_loss_fn = error_loss_fn self._gradient_clipping = gradient_clipping self._heteroscedastic = isinstance( reward_network, heteroscedastic_q_network.HeteroscedasticQNetwork) self._laplacian_matrix = None if laplacian_matrix is not None: self._laplacian_matrix = tf.convert_to_tensor( laplacian_matrix, dtype=tf.float32) # Check the validity of the laplacian matrix. tf.debugging.assert_near( 0.0, tf.norm(tf.reduce_sum(self._laplacian_matrix, 1))) tf.debugging.assert_near( 0.0, tf.norm(tf.reduce_sum(self._laplacian_matrix, 0))) self._laplacian_smoothing_weight = laplacian_smoothing_weight policy = greedy_reward_policy.GreedyRewardPredictionPolicy( time_step_spec, action_spec, reward_network, observation_and_action_constraint_splitter, constraints=constraints, accepts_per_arm_features=accepts_per_arm_features, emit_policy_info=emit_policy_info) training_data_spec = None if accepts_per_arm_features: training_data_spec = bandit_spec_utils.drop_arm_observation( policy.trajectory_spec) super(GreedyRewardPredictionAgent, self).__init__( time_step_spec, action_spec, policy, collect_policy=policy, train_sequence_length=None, training_data_spec=training_data_spec, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars, enable_summaries=enable_summaries, train_step_counter=train_step_counter) self._as_trajectory = data_converter.AsTrajectory( self.data_context, sequence_length=None)
def __init__(self, time_step_spec: ts.TimeStep, action_spec: types.TensorSpec, actor_network: network.Network, optimizer: types.Optimizer, value_network: Optional[network.Network] = None, value_estimation_loss_coef: types.Float = 0.2, advantage_fn: Optional[AdvantageFnType] = None, use_advantage_loss: bool = True, gamma: types.Float = 1.0, normalize_returns: bool = True, gradient_clipping: Optional[types.Float] = None, debug_summaries: bool = False, summarize_grads_and_vars: bool = False, entropy_regularization: Optional[types.Float] = None, train_step_counter: Optional[tf.Variable] = None, name: Optional[Text] = None): """Creates a REINFORCE Agent. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A nest of BoundedTensorSpec representing the actions. actor_network: A tf_agents.network.Network to be used by the agent. The network will be called with call(observation, step_type). optimizer: Optimizer for the actor network. value_network: (Optional) A `tf_agents.network.Network` to be used by the agent. The network will be called with call(observation, step_type) and returns a floating point value tensor. value_estimation_loss_coef: (Optional) Multiplier for value prediction loss to balance with policy gradient loss. advantage_fn: A function `A(returns, value_preds)` that takes returns and value function predictions as input and returns advantages. The default is `A(returns, value_preds) = returns - value_preds` if a value network is specified and `use_advantage_loss=True`, otherwise `A(returns, value_preds) = returns`. use_advantage_loss: Whether to use value function predictions for computing returns. `use_advantage_loss=False` is equivalent to setting `advantage_fn=lambda returns, value_preds: returns`. gamma: A discount factor for future rewards. normalize_returns: Whether to normalize returns across episodes when computing the loss. gradient_clipping: Norm length to clip gradients. debug_summaries: A bool to gather debug summaries. summarize_grads_and_vars: If True, gradient and network variable summaries will be written during training. entropy_regularization: Coefficient for entropy regularization loss term. 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. """ tf.Module.__init__(self, name=name) actor_network.create_variables() self._actor_network = actor_network if value_network: value_network.create_variables() self._value_network = value_network collect_policy = actor_policy.ActorPolicy( time_step_spec=time_step_spec, action_spec=action_spec, actor_network=self._actor_network, clip=True) policy = greedy_policy.GreedyPolicy(collect_policy) self._optimizer = optimizer self._gamma = gamma self._normalize_returns = normalize_returns self._gradient_clipping = gradient_clipping self._entropy_regularization = entropy_regularization self._value_estimation_loss_coef = value_estimation_loss_coef self._baseline = self._value_network is not None self._advantage_fn = advantage_fn if self._advantage_fn is None: if use_advantage_loss and self._baseline: self._advantage_fn = lambda returns, value_preds: returns - value_preds else: self._advantage_fn = lambda returns, _: returns super(ReinforceAgent, 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) self._as_trajectory = data_converter.AsTrajectory(self.data_context)
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=0.0, kl_cutoff_coef=0.0, initial_adaptive_kl_beta=0.0, adaptive_kl_target=0.0, adaptive_kl_tolerance=0.0, gradient_clipping=None, value_clipping=None, check_numerics=False, # TODO(b/150244758): Change the default to False once we move # clients onto Reverb. compute_value_and_advantage_in_train=True, update_normalizers_in_train=True, debug_summaries=False, summarize_grads_and_vars=False, train_step_counter=None, name='AttentionPPOAgent'): """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). If true, and the observation spec is not tf.float32 (such as Atari), please manually convert the observation spec received from the environment to tf.float32 before creating the networks. Otherwise, the normalized input to the network (float32) will have a different dtype as what the network expects, resulting in a mismatch error. Example usage: ```python observation_tensor_spec, action_spec, time_step_tensor_spec = ( spec_utils.get_tensor_specs(env)) normalized_observation_tensor_spec = tf.nest.map_structure( lambda s: tf.TensorSpec( dtype=tf.float32, shape=s.shape, name=s.name ), observation_tensor_spec ) actor_net = actor_distribution_network.ActorDistributionNetwork( normalized_observation_tensor_spec, ...) value_net = value_network.ValueNetwork( normalized_observation_tensor_spec, ...) # Note that the agent still uses the original time_step_tensor_spec # from the environment. agent = ppo_clip_agent.PPOClipAgent( time_step_tensor_spec, action_spec, actor_net, value_net, ...) ``` 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. value_clipping: Difference between new and old value predictions are clipped to this threshold. Value clipping could be helpful when training very deep networks. 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. update_normalizers_in_train: A bool to indicate whether normalizers are updated as parts of the `train` method. Set to `False` if mini batch learning is enabled, or if `train` is called on multiple iterations of the same trajectories. In that case, you would need to use `PPOLearner` (which updates all the normalizers outside of the agent). This ensures that normalizers are updated in the same way as (Schulman, 2017). 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: TypeError: if `actor_net` or `value_net` is not of type `tf_agents.networks.Network`. """ if not isinstance(actor_net, network.Network): raise TypeError( 'actor_net must be an instance of a network.Network.') if not isinstance(value_net, network.Network): raise TypeError( 'value_net must be an instance of a network.Network.') # PPOPolicy validates these, so we skip validation here. actor_net.create_variables(time_step_spec.observation) value_net.create_variables(time_step_spec.observation) 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._value_clipping = value_clipping or 0.0 self._check_numerics = check_numerics self._compute_value_and_advantage_in_train = ( compute_value_and_advantage_in_train) self.update_normalizers_in_train = update_normalizers_in_train if not isinstance(self._optimizer, tf.keras.optimizers.Optimizer): logging.warning( 'Only tf.keras.optimizers.Optimizers are well supported, got a ' 'non-TF2 optimizer: %s', self._optimizer) self._initial_adaptive_kl_beta = initial_adaptive_kl_beta if initial_adaptive_kl_beta > 0.0: 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')) self._advantage_normalizer = tensor_normalizer.StreamingTensorNormalizer( tensor_spec.TensorSpec([], tf.float32), scope='normalize_advantages') policy = greedy_policy.GreedyPolicy( attention_ppo_policy.AttentionPPOPolicy( 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 = attention_ppo_policy.AttentionPPOPolicy( 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), ) if isinstance(self._actor_net, network.DistributionNetwork): # Legacy behavior self._action_distribution_spec = self._actor_net.output_spec else: self._action_distribution_spec = self._actor_net.create_variables( time_step_spec.observation) # Set training_data_spec to collect_data_spec with augmented policy info, # iff return and normalized advantage are saved in preprocess_sequence. if self._compute_value_and_advantage_in_train: training_data_spec = None else: training_policy_info = collect_policy.trajectory_spec.policy_info.copy( ) training_policy_info.update({ 'value_prediction': collect_policy.trajectory_spec.policy_info['value_prediction'], 'return': tensor_spec.TensorSpec(shape=[], dtype=tf.float32), 'advantage': tensor_spec.TensorSpec(shape=[], dtype=tf.float32), }) training_data_spec = collect_policy.trajectory_spec.replace( policy_info=training_policy_info) super(ppo_agent.PPOAgent, self).__init__(time_step_spec, action_spec, policy, collect_policy, train_sequence_length=None, training_data_spec=training_data_spec, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars, train_step_counter=train_step_counter) # This must be built after super() which sets up self.data_context. self._collected_as_transition = data_converter.AsTransition( self.collect_data_context, squeeze_time_dim=False) self._as_trajectory = data_converter.AsTrajectory(self.data_context, sequence_length=None)