def project_to_output_distribution(inputs,
output_spec,
project_to_discrete,
project_to_continuous,
outer_rank=1,
scope='project_to_output'):
"""Project a batch of inputs to a distribution object.
Args:
inputs: An input Tensor of shape [batch_size, None].
output_spec: A single output spec.
project_to_discrete: The method to use for projecting a discrete output.
project_to_continuous: The method to use for projecting a continuous output.
outer_rank: The number of outer dimensions of inputs to consider batch
dimensions and to treat as batch dimensions of output distribution.
scope: The variable scope.
Returns:
A distribution object corresponding to the arguments and output spec
provided.
Raises:
ValueError: If the distribution type of output_spec is unclear.
"""
with tf.variable_scope(scope):
if tensor_spec.is_discrete(output_spec):
return project_to_discrete(inputs, output_spec, outer_rank=outer_rank)
elif tensor_spec.is_continuous(output_spec):
return project_to_continuous(inputs, output_spec, outer_rank=outer_rank)
else:
raise ValueError('Output spec corresponds to unknown distribution.')
def testExclusive(self, dtype):
if dtype == tf.string:
self.skipTest("Not compatible with string type.")
spec = tensor_spec.TensorSpec((2, 3), dtype=dtype)
self.assertIs(
tensor_spec.is_discrete(spec) ^ tensor_spec.is_continuous(spec),
True)
def calc_default_target_entropy(spec):
"""Calc default target entropy
Args:
spec (TensorSpec): action spec
Returns:
"""
dims = np.product(spec.shape.as_list())
if tensor_spec.is_continuous(spec):
e = -1
else:
min_prob = 0.01
p = min_prob
q = 1 - p
e = -p * np.log(p) - q * np.log(q)
return e * dims
def calc_default_target_entropy(spec):
"""Calc default target entropy
Args:
spec (TensorSpec): action spec
Returns:
"""
zeros = np.zeros(spec.shape)
min_max = np.broadcast(spec.minimum, spec.maximum, zeros)
cont = tensor_spec.is_continuous(spec)
min_prob = 0.01
log_mp = np.log(min_prob)
# continuous: suppose the prob concentrates on a delta of 0.01*(M-m)
# discrete: ignore the entry of 0.99 and uniformly distribute probs on rest
e = np.sum([(np.log(M - m) + log_mp
if cont else min_prob * (np.log(M - m) - log_mp))
for m, M, _ in min_max])
return e
def calc_default_max_entropy(spec, fraction=0.8):
"""Calc default max entropy
Args:
spec (TensorSpec): action spec
fraction (float): this fraction of the theoretical entropy upper bound
will be used as the max entropy
Returns:
A default max entropy for adjusting the entropy weight
"""
assert fraction <= 1.0 and fraction > 0
zeros = np.zeros(spec.shape)
min_max = np.broadcast(spec.minimum, spec.maximum, zeros)
cont = tensor_spec.is_continuous(spec)
# use uniform distributions to compute upper bounds
e = np.sum([(np.log(M - m) * (fraction if M - m > 1 else 1.0 / fraction)
if cont else np.log(M - m + 1) * fraction)
for m, M, _ in min_max])
return e
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 testExclusive(self, dtype):
spec = tensor_spec.TensorSpec((2, 3), dtype=dtype)
self.assertIs(
tensor_spec.is_discrete(spec) ^ tensor_spec.is_continuous(spec),
True)
def testIsContinuous(self, dtype):
spec = tensor_spec.TensorSpec((2, 3), dtype=dtype)
self.assertIs(tensor_spec.is_continuous(spec), dtype.is_floating)
def __init__(self,
action_spec,
actor_network: Network,
critic_network: Network,
critic_loss=None,
target_entropy=None,
initial_log_alpha=0.0,
target_update_tau=0.05,
target_update_period=1,
dqda_clipping=None,
actor_optimizer=None,
critic_optimizer=None,
alpha_optimizer=None,
gradient_clipping=None,
train_step_counter=None,
debug_summaries=False,
name="SacAlgorithm"):
"""Create a SacAlgorithm
Args:
action_spec (nested BoundedTensorSpec): representing the actions.
actor_network (Network): The network will be called with
call(observation, step_type).
critic_network (Network): The network will be called with
call(observation, action, step_type).
critic_loss (None|OneStepTDLoss): an object for calculating critic loss.
If None, a default OneStepTDLoss will be used.
initial_log_alpha (float): initial value for variable log_alpha
target_entropy (float|None): The target average policy entropy, for updating alpha.
target_update_tau (float): Factor for soft update of the target
networks.
target_update_period (int): Period for soft update of the target
networks.
dqda_clipping (float): when computing the actor loss, clips the
gradient dqda element-wise between [-dqda_clipping, dqda_clipping].
Does not perform clipping if dqda_clipping == 0.
actor_optimizer (tf.optimizers.Optimizer): The optimizer for actor.
critic_optimizer (tf.optimizers.Optimizer): The optimizer for critic.
alpha_optimizer (tf.optimizers.Optimizer): The optimizer for alpha.
gradient_clipping (float): Norm length to clip gradients.
train_step_counter (tf.Variable): An optional counter to increment
every time the a new iteration is started. If None, it will use
tf.summary.experimental.get_step(). If this is still None, a
counter will be created.
debug_summaries (bool): True if debug summaries should be created.
name (str): The name of this algorithm.
"""
critic_network1 = critic_network
critic_network2 = critic_network.copy(name='CriticNetwork2')
log_alpha = tfa_common.create_variable(name='log_alpha',
initial_value=initial_log_alpha,
dtype=tf.float32,
trainable=True)
super().__init__(
action_spec,
train_state_spec=SacState(
share=SacShareState(actor=actor_network.state_spec),
actor=SacActorState(critic1=critic_network.state_spec,
critic2=critic_network.state_spec),
critic=SacCriticState(
critic1=critic_network.state_spec,
critic2=critic_network.state_spec,
target_critic1=critic_network.state_spec,
target_critic2=critic_network.state_spec)),
action_distribution_spec=actor_network.output_spec,
predict_state_spec=actor_network.state_spec,
optimizer=[actor_optimizer, critic_optimizer, alpha_optimizer],
get_trainable_variables_func=[
lambda: actor_network.trainable_variables, lambda:
(critic_network1.trainable_variables + critic_network2.
trainable_variables), lambda: [log_alpha]
],
gradient_clipping=gradient_clipping,
train_step_counter=train_step_counter,
debug_summaries=debug_summaries,
name=name)
self._log_alpha = log_alpha
self._actor_network = actor_network
self._critic_network1 = critic_network1
self._critic_network2 = critic_network2
self._target_critic_network1 = self._critic_network1.copy(
name='TargetCriticNetwork1')
self._target_critic_network2 = self._critic_network2.copy(
name='TargetCriticNetwork2')
self._actor_optimizer = actor_optimizer
self._critic_optimizer = critic_optimizer
self._alpha_optimizer = alpha_optimizer
if critic_loss is None:
critic_loss = OneStepTDLoss(debug_summaries=debug_summaries)
self._critic_loss = critic_loss
flat_action_spec = tf.nest.flatten(self._action_spec)
self._is_continuous = tensor_spec.is_continuous(flat_action_spec[0])
if target_entropy is None:
target_entropy = np.sum(
list(
map(dist_utils.calc_default_target_entropy,
flat_action_spec)))
self._target_entropy = target_entropy
self._dqda_clipping = dqda_clipping
self._update_target = common.get_target_updater(
models=[self._critic_network1, self._critic_network2],
target_models=[
self._target_critic_network1, self._target_critic_network2
],
tau=target_update_tau,
period=target_update_period)
tfa_common.soft_variables_update(
self._critic_network1.variables,
self._target_critic_network1.variables,
tau=1.0)
tfa_common.soft_variables_update(
self._critic_network2.variables,
self._target_critic_network2.variables,
tau=1.0)
def create_sac_algorithm(env,
actor_fc_layers=(100, 100),
critic_fc_layers=(100, 100),
use_rnns=False,
alpha_learning_rate=5e-3,
actor_learning_rate=5e-3,
critic_learning_rate=5e-3,
debug_summaries=False):
"""Create a simple SacAlgorithm.
Args:
env (TFEnvironment): A TFEnvironment
actor_fc_layers (list[int]): list of fc layers parameters for actor network
critic_fc_layers (list[int]): list of fc layers parameters for critic network
use_rnns (bool): True if rnn should be used
alpha_learning_rate (float): learning rate for alpha
actor_learning_rate (float) : learning rate for actor network
critic_learning_rate (float) : learning rate for critic network
debug_summaries (bool): True if debug summaries should be created
"""
observation_spec = env.observation_spec()
action_spec = env.action_spec()
is_continuous = tensor_spec.is_continuous(tf.nest.flatten(action_spec)[0])
if use_rnns:
actor_net = ActorDistributionRnnNetwork(
observation_spec,
action_spec,
input_fc_layer_params=actor_fc_layers,
output_fc_layer_params=())
if is_continuous:
critic_net = CriticRnnNetwork(
(observation_spec, action_spec),
observation_fc_layer_params=(),
action_fc_layer_params=(),
output_fc_layer_params=(),
joint_fc_layer_params=critic_fc_layers)
else:
critic_net = QRnnNetwork(observation_spec,
action_spec,
output_fc_layer_params=(),
input_fc_layer_params=critic_fc_layers)
else:
actor_net = ActorDistributionNetwork(observation_spec,
action_spec,
fc_layer_params=actor_fc_layers)
if is_continuous:
critic_net = CriticNetwork((observation_spec, action_spec),
joint_fc_layer_params=critic_fc_layers)
else:
critic_net = QNetwork(observation_spec,
action_spec,
fc_layer_params=critic_fc_layers)
actor_optimizer = tf.optimizers.Adam(learning_rate=actor_learning_rate)
critic_optimizer = tf.optimizers.Adam(learning_rate=critic_learning_rate)
alpha_optimizer = tf.optimizers.Adam(learning_rate=alpha_learning_rate)
return SacAlgorithm(action_spec=action_spec,
actor_network=actor_net,
critic_network=critic_net,
actor_optimizer=actor_optimizer,
critic_optimizer=critic_optimizer,
alpha_optimizer=alpha_optimizer,
debug_summaries=debug_summaries)
def testIsContinuous(self, dtype):
spec = array_spec.ArraySpec((2, 3), dtype=dtype)
self.assertIs(tensor_spec.is_continuous(spec),
issubclass(np.dtype(dtype).type, np.floating))
def _validate_action_spec(action_spec):
if not tensor_spec.is_continuous(action_spec):
raise ValueError(
'OU Noise is applicable only to continuous actions.')
def normal(inputs,
output_spec,
outer_rank=1,
projection_layer=default_fully_connected,
mean_transform=tanh_squash_to_spec,
std_initializer=tf.zeros_initializer(),
std_transform=tf.exp,
distribution_cls=tfp.distributions.Normal):
"""Project a batch of inputs to a batch of means and standard deviations.
Given an output spec for a single tensor continuous action, produces a
neural net layer converting inputs to a normal distribution matching
the spec. The mean is derived from a fully connected linear layer as
mean_transform(layer_output, output_spec). The std is fixed to a single
trainable tensor (thus independent of the inputs). Specifically, std is
parameterized as std_transform(variable).
Args:
inputs: An input Tensor of shape [batch_size, ?].
output_spec: An output spec (either BoundedArraySpec or BoundedTensorSpec).
outer_rank: The number of outer dimensions of inputs to consider batch
dimensions and to treat as batch dimensions of output distribution.
projection_layer: Function taking in inputs, num_elements, scope and
returning a projection of inputs to a Tensor of width num_elements.
mean_transform: A function taking in layer output and the output_spec,
returning the means. Defaults to tanh_squash_to_spec.
std_initializer: Initializer for std_dev variables.
std_transform: The function applied to the trainable std variable. For
example, tf.exp (default), tf.nn.softplus.
distribution_cls: The distribution class to use for output distribution.
Default is tfp.distributions.Normal.
Returns:
A tf.distribution.Normal object in which the standard deviation is not
dependent on input.
Raises:
ValueError: If output_spec is invalid.
"""
if not tensor_spec.is_bounded(output_spec):
raise ValueError('Input output_spec is of invalid type '
'%s.' % type(output_spec))
if not tensor_spec.is_continuous(output_spec):
raise ValueError('Output is not continuous.')
batch_squash = utils.BatchSquash(outer_rank)
inputs = batch_squash.flatten(inputs)
means = projection_layer(inputs,
output_spec.shape.num_elements(),
scope='means')
stds = tf.contrib.layers.bias_add(
tf.zeros_like(means), # Independent of inputs.
initializer=std_initializer,
scope='stds',
activation_fn=None)
means = tf.reshape(means, [-1] + output_spec.shape.as_list())
means = mean_transform(means, output_spec)
means = tf.cast(means, output_spec.dtype)
stds = tf.reshape(stds, [-1] + output_spec.shape.as_list())
stds = std_transform(stds)
stds = tf.cast(stds, output_spec.dtype)
means, stds = batch_squash.unflatten(means), batch_squash.unflatten(stds)
return distribution_cls(means, stds)
def __init__(self,
x_spec,
y_spec,
model=None,
fc_layers=(256, ),
sampler='buffer',
buffer_size=65536,
optimizer: tf.optimizers.Optimizer = None,
estimator_type='DV',
averager=ScalarAdaptiveAverager(),
name="MIEstimator"):
"""Create a MIEstimator.
Args:
x_spec (nested TensorSpec): spec of x
y_spec (nested TensorSpec): spec of y
model (Network): can be called as model([x, y]) and return a Tensor
with shape=[batch_size, 1]. If None, a default MLP with
fc_layers will be created.
fc_layers (tuple[int]): size of hidden layers. Only used if model is
None.
sampler (str): type of sampler used to get samples from marginal
distribution, should be one of ['buffer', 'double_buffer',
'shuffle', 'shift']
buffer_size (int): capacity of buffer for storing y for sampler
'buffer' and 'double_buffer'
optimzer (tf.optimizers.Optimzer): optimizer
estimator_type (str): one of 'DV', 'KLD' or 'JSD'
averager (EMAverager): averager used to maintain a moving average
of exp(T). Only used for 'DV' estimator
name (str): name of this estimator
"""
assert estimator_type in ['ML', 'DV', 'KLD', 'JSD'
], "Wrong estimator_type %s" % estimator_type
super().__init__(train_state_spec=(), optimizer=optimizer, name=name)
self._x_spec = x_spec
self._y_spec = y_spec
if model is None:
if estimator_type == 'ML':
model = TFAEncodingNetwork(
name="MIEstimator",
input_tensor_spec=x_spec,
fc_layer_params=fc_layers,
preprocessing_combiner=NestConcatenate(axis=-1))
else:
model = EncodingNetwork(
name="MIEstimator",
input_tensor_spec=[x_spec, y_spec],
fc_layer_params=fc_layers,
last_layer_size=1)
self._model = model
self._type = estimator_type
if sampler == 'buffer':
self._y_buffer = DataBuffer(y_spec, capacity=buffer_size)
self._sampler = self._buffer_sampler
elif sampler == 'double_buffer':
self._x_buffer = DataBuffer(x_spec, capacity=buffer_size)
self._y_buffer = DataBuffer(y_spec, capacity=buffer_size)
self._sampler = self._double_buffer_sampler
elif sampler == 'shuffle':
self._sampler = self._shuffle_sampler
elif sampler == 'shift':
self._sampler = self._shift_sampler
else:
raise TypeError("Wrong type for sampler %s" % sampler)
if estimator_type == 'DV':
self._mean_averager = averager
if estimator_type == 'ML':
assert isinstance(
y_spec,
tf.TensorSpec), ("Currently, 'ML' does "
"not support nested y_spec: %s" % y_spec)
assert tensor_spec.is_continuous(y_spec), (
"Currently, 'ML' does "
"not support discreted y_spec: %s" % y_spec)
self._delta_loc_layer = tf.keras.layers.Dense(
y_spec.shape[-1],
kernel_initializer=tf.initializers.Zeros(),
bias_initializer=tf.initializers.Zeros(),
name='delta_loc_layer')
self._delta_scale_layer = tf.keras.layers.Dense(
y_spec.shape[-1],
kernel_initializer=tf.initializers.Zeros(),
bias_initializer=tf.keras.initializers.Constant(
value=math.log(math.e - 1)),
name='delta_scale_layer')