def test_without_std_share_network_shapes(self, output_dim, hidden_dim):
model = GaussianGRUModel(output_dim=output_dim,
hidden_dim=hidden_dim,
std_share_network=False,
hidden_nonlinearity=None,
recurrent_nonlinearity=None,
hidden_w_init=self.default_initializer,
recurrent_w_init=self.default_initializer,
output_w_init=self.default_initializer)
step_hidden_var = tf.compat.v1.placeholder(shape=(self.batch_size,
hidden_dim),
name='step_hidden',
dtype=tf.float32)
model.build(self.input_var, self.step_input_var, step_hidden_var)
# output layer is a tf.keras.layers.Dense object,
# which cannot be access by tf.compat.v1.variable_scope.
# A workaround is to access in tf.compat.v1.global_variables()
for var in tf.compat.v1.global_variables():
if 'output_layer/kernel' in var.name:
std_share_output_weights = var
if 'output_layer/bias' in var.name:
std_share_output_bias = var
if 'log_std_param/parameter' in var.name:
log_std_param = var
assert std_share_output_weights.shape[1] == output_dim
assert std_share_output_bias.shape == output_dim
assert log_std_param.shape == output_dim
def test_dist(self):
model = GaussianGRUModel(output_dim=1, hidden_dim=1)
step_hidden_var = tf.compat.v1.placeholder(shape=(self.batch_size, 1),
name='step_hidden',
dtype=tf.float32)
model.build(self.input_var, self.step_input_var, step_hidden_var)
assert isinstance(model.networks['default'].dist,
tfp.distributions.MultivariateNormalDiag)
def test_without_std_share_network_output_values(self, mock_normal,
output_dim, hidden_dim,
init_std):
mock_normal.return_value = 0.5
model = GaussianGRUModel(output_dim=output_dim,
hidden_dim=hidden_dim,
std_share_network=False,
hidden_nonlinearity=None,
recurrent_nonlinearity=None,
hidden_w_init=self.default_initializer,
recurrent_w_init=self.default_initializer,
output_w_init=self.default_initializer,
init_std=init_std)
step_hidden_var = tf.compat.v1.placeholder(shape=(self.batch_size,
hidden_dim),
name='step_hidden',
dtype=tf.float32)
(mean_var, step_mean_var, log_std_var, step_log_std_var, step_hidden,
hidden_init_var, dist) = model.build(self.input_var,
self.step_input_var,
step_hidden_var)
hidden1 = hidden2 = np.full((self.batch_size, hidden_dim),
hidden_init_var.eval())
mean, log_std = self.sess.run(
[mean_var, log_std_var],
feed_dict={self.input_var: self.obs_inputs})
for i in range(self.time_step):
mean1, log_std1, hidden1 = self.sess.run(
[step_mean_var, step_log_std_var, step_hidden],
feed_dict={
self.step_input_var: self.obs_input,
step_hidden_var: hidden1
})
hidden2 = recurrent_step_gru(input_val=self.obs_input,
num_units=hidden_dim,
step_hidden=hidden2,
w_x_init=0.1,
w_h_init=0.1,
b_init=0.,
nonlinearity=None,
gate_nonlinearity=None)
output_nonlinearity = np.full(
(np.prod(hidden2.shape[1:]), output_dim), 0.1)
output2 = np.matmul(hidden2, output_nonlinearity)
assert np.allclose(mean1, output2)
expected_log_std = np.full((self.batch_size, output_dim),
np.log(init_std))
assert np.allclose(log_std1, expected_log_std)
assert np.allclose(hidden1, hidden2)
def test_without_std_share_network_is_pickleable(self, mock_normal,
output_dim, hidden_dim):
mock_normal.return_value = 0.5
model = GaussianGRUModel(output_dim=output_dim,
hidden_dim=hidden_dim,
std_share_network=False,
hidden_nonlinearity=None,
recurrent_nonlinearity=None,
hidden_w_init=self.default_initializer,
recurrent_w_init=self.default_initializer,
output_w_init=self.default_initializer)
step_hidden_var = tf.compat.v1.placeholder(shape=(self.batch_size,
hidden_dim),
name='step_hidden',
dtype=tf.float32)
(mean_var, step_mean_var, log_std_var, step_log_std_var, step_hidden,
_, _) = model.build(self.input_var, self.step_input_var,
step_hidden_var)
# output layer is a tf.keras.layers.Dense object,
# which cannot be access by tf.compat.v1.variable_scope.
# A workaround is to access in tf.compat.v1.global_variables()
for var in tf.compat.v1.global_variables():
if 'output_layer/bias' in var.name:
var.load(tf.ones_like(var).eval())
hidden = np.zeros((self.batch_size, hidden_dim))
outputs1 = self.sess.run([mean_var, log_std_var],
feed_dict={self.input_var: self.obs_inputs})
output1 = self.sess.run([step_mean_var, step_log_std_var, step_hidden],
feed_dict={
self.step_input_var: self.obs_input,
step_hidden_var: hidden
}) # noqa: E126
h = pickle.dumps(model)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
model_pickled = pickle.loads(h)
input_var = tf.compat.v1.placeholder(tf.float32,
shape=(None, None,
self.feature_shape),
name='input')
step_input_var = tf.compat.v1.placeholder(
tf.float32,
shape=(None, self.feature_shape),
name='step_input')
step_hidden_var = tf.compat.v1.placeholder(shape=(self.batch_size,
hidden_dim),
name='initial_hidden',
dtype=tf.float32)
(mean_var2, step_mean_var2, log_std_var2, step_log_std_var2,
step_hidden2, _, _) = model_pickled.build(input_var,
step_input_var,
step_hidden_var)
outputs2 = sess.run([mean_var2, log_std_var2],
feed_dict={input_var: self.obs_inputs})
output2 = sess.run(
[step_mean_var2, step_log_std_var2, step_hidden2],
feed_dict={
step_input_var: self.obs_input,
step_hidden_var: hidden
})
assert np.array_equal(outputs1, outputs2)
assert np.array_equal(output1, output2)
class GaussianGRUPolicy(StochasticPolicy):
"""Models the action distribution using a Gaussian parameterized by a GRU.
Args:
env_spec (metarl.envs.env_spec.EnvSpec): Environment specification.
name (str): Model name, also the variable scope.
hidden_dim (int): Hidden dimension for GRU cell for mean.
hidden_nonlinearity (Callable): Activation function for intermediate
dense layer(s). It should return a tf.Tensor. Set it to
None to maintain a linear activation.
hidden_w_init (Callable): Initializer function for the weight
of intermediate dense layer(s). The function should return a
tf.Tensor.
hidden_b_init (Callable): Initializer function for the bias
of intermediate dense layer(s). The function should return a
tf.Tensor.
recurrent_nonlinearity (Callable): Activation function for recurrent
layers. It should return a tf.Tensor. Set it to None to
maintain a linear activation.
recurrent_w_init (Callable): Initializer function for the weight
of recurrent layer(s). The function should return a
tf.Tensor.
output_nonlinearity (Callable): Activation function for output dense
layer. It should return a tf.Tensor. Set it to None to
maintain a linear activation.
output_w_init (Callable): Initializer function for the weight
of output dense layer(s). The function should return a
tf.Tensor.
output_b_init (Callable): Initializer function for the bias
of output dense layer(s). The function should return a
tf.Tensor.
hidden_state_init (Callable): Initializer function for the
initial hidden state. The functino should return a tf.Tensor.
hidden_state_init_trainable (bool): Bool for whether the initial
hidden state is trainable.
learn_std (bool): Is std trainable.
std_share_network (bool): Boolean for whether mean and std share
the same network.
init_std (float): Initial value for std.
layer_normalization (bool): Bool for using layer normalization or not.
state_include_action (bool): Whether the state includes action.
If True, input dimension will be
(observation dimension + action dimension).
"""
def __init__(self,
env_spec,
hidden_dims=[32],
name='GaussianGRUPolicy',
hidden_nonlinearity=tf.nn.tanh,
hidden_w_init=tf.glorot_uniform_initializer(),
hidden_b_init=tf.zeros_initializer(),
recurrent_nonlinearity=tf.nn.sigmoid,
recurrent_w_init=tf.glorot_uniform_initializer(),
output_nonlinearity=None,
output_w_init=tf.glorot_uniform_initializer(),
output_b_init=tf.zeros_initializer(),
hidden_state_init=tf.zeros_initializer(),
hidden_state_init_trainable=False,
learn_std=True,
std_share_network=False,
init_std=1.0,
layer_normalization=False,
state_include_action=True):
if not isinstance(env_spec.action_space, akro.Box):
raise ValueError('GaussianGRUPolicy only works with '
'akro.Box action space, but not {}'.format(
env_spec.action_space))
super().__init__(name, env_spec)
self._obs_dim = env_spec.observation_space.flat_dim
self._action_dim = env_spec.action_space.flat_dim
self._hidden_dims = hidden_dims
self._state_include_action = state_include_action
if state_include_action:
self._input_dim = self._obs_dim + self._action_dim
else:
self._input_dim = self._obs_dim
self.model = GaussianGRUModel(
output_dim=self._action_dim,
hidden_dims=hidden_dims,
name='GaussianGRUModel',
hidden_nonlinearity=hidden_nonlinearity,
hidden_w_init=hidden_w_init,
hidden_b_init=hidden_b_init,
recurrent_nonlinearity=recurrent_nonlinearity,
recurrent_w_init=recurrent_w_init,
output_nonlinearity=output_nonlinearity,
output_w_init=output_w_init,
output_b_init=output_b_init,
hidden_state_init=hidden_state_init,
hidden_state_init_trainable=hidden_state_init_trainable,
layer_normalization=layer_normalization,
learn_std=learn_std,
std_share_network=std_share_network,
init_std=init_std)
self._prev_actions = None
self._prev_hiddens = None
self._initialize()
def _initialize(self):
obs_ph = tf.compat.v1.placeholder(tf.float32,
shape=(None, None, self._input_dim))
step_input_var = tf.compat.v1.placeholder(shape=(None,
self._input_dim),
name='step_input',
dtype=tf.float32)
step_hidden_var = tf.compat.v1.placeholder(
shape=(None, self._hidden_dims[0]),
name='step_hidden_input',
dtype=tf.float32)
with tf.compat.v1.variable_scope(self.name) as vs:
self._variable_scope = vs
self.model.build(obs_ph, step_input_var, step_hidden_var)
self._f_step_mean_std = (
tf.compat.v1.get_default_session().make_callable(
[
self.model.networks['default'].step_mean,
self.model.networks['default'].step_log_std,
self.model.networks['default'].step_hidden
],
feed_list=[step_input_var, step_hidden_var]))
@property
def vectorized(self):
"""bool: Whether the policy is vectorized or not."""
return True
def dist_info_sym(self, obs_var, state_info_vars, name=None):
"""Build a symbolic graph of the distribution parameters.
Args:
obs_var (tf.Tensor): Tensor input for symbolic graph.
state_info_vars (dict): Extra state information, e.g.
previous action.
name (str): Name for symbolic graph.
Returns:
dict[tf.Tensor]: Outputs of the symbolic graph of distribution
parameters.
"""
if self._state_include_action:
prev_action_var = state_info_vars['prev_action']
prev_action_var = tf.cast(prev_action_var, tf.float32)
all_input_var = tf.concat(axis=2,
values=[obs_var, prev_action_var])
else:
all_input_var = obs_var
with tf.compat.v1.variable_scope(self._variable_scope):
mean_var, _, log_std_var, _, _, _, _ = self.model.build(
all_input_var,
self.model.networks['default'].step_input,
self.model.networks['default'].step_hidden_input,
name=name)
return dict(mean=mean_var, log_std=log_std_var)
def reset(self, dones=None):
"""Reset the policy.
Note:
If `dones` is None, it will be by default `np.array([True])` which
implies the policy will not be "vectorized", i.e. number of
parallel environments for training data sampling = 1.
Args:
dones (numpy.ndarray): Bool that indicates terminal state(s).
"""
if dones is None:
dones = np.array([True])
if self._prev_actions is None or len(dones) != len(self._prev_actions):
self._prev_actions = np.zeros(
(len(dones), self.action_space.flat_dim))
self._prev_hiddens = np.zeros((len(dones), self._hidden_dims[0]))
self._prev_actions[dones] = 0.
self._prev_hiddens[dones] = self.model.networks[
'default'].init_hidden.eval()
def get_action(self, observation):
"""Get a single action from this policy for the input observation.
Args:
observation (numpy.ndarray): Observation from environment.
Returns:
tuple[numpy.ndarray, dict]: Predicted action and agent info.
action (numpy.ndarray): Predicted action.
agent_info (dict): Distribution obtained after observing the
given observation, with keys
* mean: (numpy.ndarray)
* log_std: (numpy.ndarray)
* prev_action: (numpy.ndarray), only present if
self._state_include_action is True.
"""
actions, agent_infos = self.get_actions([observation])
return actions[0], {k: v[0] for k, v in agent_infos.items()}
def get_actions(self, observations):
"""Get multiple actions from this policy for the input observations.
Args:
observations (numpy.ndarray): Observations from environment.
Returns:
tuple[numpy.ndarray, dict]: Prediction actions and agent infos.
actions (numpy.ndarray): Predicted actions.
agent_infos (dict): Distribution obtained after observing the
given observation, with keys
* mean: (numpy.ndarray)
* log_std: (numpy.ndarray)
* prev_action: (numpy.ndarray), only present if
self._state_include_action is True.
"""
# flat_obs = self.observation_space.flatten_n(observations)
if self._state_include_action:
assert self._prev_actions is not None
all_input = np.concatenate([observations, self._prev_actions],
axis=-1)
else:
all_input = observations
means, log_stds, hidden_vec = self._f_step_mean_std(
all_input, self._prev_hiddens)
rnd = np.random.normal(size=means.shape)
samples = rnd * np.exp(log_stds) + means
# samples = self.action_space.unflatten_n(samples)
prev_actions = self._prev_actions
self._prev_actions = samples
self._prev_hiddens = hidden_vec
agent_infos = dict(mean=means, log_std=log_stds)
if self._state_include_action:
agent_infos['prev_action'] = np.copy(prev_actions)
return samples, agent_infos
@property
def recurrent(self):
"""bool: Whether this policy is recurrent or not."""
return True
@property
def distribution(self):
"""metarl.tf.distributions.DiagonalGaussian: Policy distribution."""
return self.model.networks['default'].dist
@property
def state_info_specs(self):
"""list: State info specification."""
if self._state_include_action:
return [
('prev_action', (self._action_dim, )),
]
return []
def __getstate__(self):
"""See `Object.__getstate__`."""
new_dict = super().__getstate__()
del new_dict['_f_step_mean_std']
return new_dict
def __setstate__(self, state):
"""See `Object.__setstate__`."""
super().__setstate__(state)
self._initialize()
class GaussianGRUPolicy(StochasticPolicy):
"""Gaussian GRU Policy.
A policy represented by a Gaussian distribution
which is parameterized by a Gated Recurrent Unit (GRU).
Args:
env_spec (metarl.envs.env_spec.EnvSpec): Environment specification.
name (str): Model name, also the variable scope.
hidden_dim (int): Hidden dimension for GRU cell for mean.
hidden_nonlinearity (Callable): Activation function for intermediate
dense layer(s). It should return a tf.Tensor. Set it to
None to maintain a linear activation.
hidden_w_init (Callable): Initializer function for the weight
of intermediate dense layer(s). The function should return a
tf.Tensor.
hidden_b_init (Callable): Initializer function for the bias
of intermediate dense layer(s). The function should return a
tf.Tensor.
recurrent_nonlinearity (Callable): Activation function for recurrent
layers. It should return a tf.Tensor. Set it to None to
maintain a linear activation.
recurrent_w_init (Callable): Initializer function for the weight
of recurrent layer(s). The function should return a
tf.Tensor.
output_nonlinearity (Callable): Activation function for output dense
layer. It should return a tf.Tensor. Set it to None to
maintain a linear activation.
output_w_init (Callable): Initializer function for the weight
of output dense layer(s). The function should return a
tf.Tensor.
output_b_init (Callable): Initializer function for the bias
of output dense layer(s). The function should return a
tf.Tensor.
hidden_state_init (Callable): Initializer function for the
initial hidden state. The functino should return a tf.Tensor.
hidden_state_init_trainable (bool): Bool for whether the initial
hidden state is trainable.
learn_std (bool): Is std trainable.
std_share_network (bool): Boolean for whether mean and std share
the same network.
init_std (float): Initial value for std.
layer_normalization (bool): Bool for using layer normalization or not.
state_include_action (bool): Whether the state includes action.
If True, input dimension will be
(observation dimension + action dimension).
"""
def __init__(self,
env_spec,
hidden_dim=32,
name='GaussianGRUPolicy',
hidden_nonlinearity=tf.nn.tanh,
hidden_w_init=tf.initializers.glorot_uniform(),
hidden_b_init=tf.zeros_initializer(),
recurrent_nonlinearity=tf.nn.sigmoid,
recurrent_w_init=tf.initializers.glorot_uniform(),
output_nonlinearity=None,
output_w_init=tf.initializers.glorot_uniform(),
output_b_init=tf.zeros_initializer(),
hidden_state_init=tf.zeros_initializer(),
hidden_state_init_trainable=False,
learn_std=True,
std_share_network=False,
init_std=1.0,
layer_normalization=False,
state_include_action=True):
if not isinstance(env_spec.action_space, akro.Box):
raise ValueError('GaussianGRUPolicy only works with '
'akro.Box action space, but not {}'.format(
env_spec.action_space))
super().__init__(name, env_spec)
self._obs_dim = env_spec.observation_space.flat_dim
self._action_dim = env_spec.action_space.flat_dim
self._hidden_dim = hidden_dim
self._hidden_nonlinearity = hidden_nonlinearity
self._hidden_w_init = hidden_w_init
self._hidden_b_init = hidden_b_init
self._recurrent_nonlinearity = recurrent_nonlinearity
self._recurrent_w_init = recurrent_w_init
self._output_nonlinearity = output_nonlinearity
self._output_w_init = output_w_init
self._output_b_init = output_b_init
self._hidden_state_init = hidden_state_init
self._hidden_state_init_trainable = hidden_state_init_trainable
self._learn_std = learn_std
self._std_share_network = std_share_network
self._init_std = init_std
self._layer_normalization = layer_normalization
self._state_include_action = state_include_action
if state_include_action:
self._input_dim = self._obs_dim + self._action_dim
else:
self._input_dim = self._obs_dim
self._f_step_mean_std = None
self.model = GaussianGRUModel(
output_dim=self._action_dim,
hidden_dim=hidden_dim,
name='GaussianGRUModel',
hidden_nonlinearity=hidden_nonlinearity,
hidden_w_init=hidden_w_init,
hidden_b_init=hidden_b_init,
recurrent_nonlinearity=recurrent_nonlinearity,
recurrent_w_init=recurrent_w_init,
output_nonlinearity=output_nonlinearity,
output_w_init=output_w_init,
output_b_init=output_b_init,
hidden_state_init=hidden_state_init,
hidden_state_init_trainable=hidden_state_init_trainable,
layer_normalization=layer_normalization,
learn_std=learn_std,
std_share_network=std_share_network,
init_std=init_std)
self._prev_actions = None
self._prev_hiddens = None
def build(self, state_input, name=None):
"""Build model.
Args:
state_input (tf.Tensor): State input.
name (str): Name of the model, which is also the name scope.
"""
with tf.compat.v1.variable_scope(self.name) as vs:
self._variable_scope = vs
step_input_var = tf.compat.v1.placeholder(shape=(None,
self._input_dim),
name='step_input',
dtype=tf.float32)
step_hidden_var = tf.compat.v1.placeholder(
shape=(None, self._hidden_dim),
name='step_hidden_input',
dtype=tf.float32)
self.model.build(state_input,
step_input_var,
step_hidden_var,
name=name)
self._f_step_mean_std = (
tf.compat.v1.get_default_session().make_callable(
[
self.model.networks['default'].step_mean,
self.model.networks['default'].step_log_std,
self.model.networks['default'].step_hidden
],
feed_list=[step_input_var, step_hidden_var]))
@property
def vectorized(self):
"""Vectorized or not.
Returns:
Bool: True if primitive supports vectorized operations.
"""
return True
def reset(self, do_resets=None):
"""Reset the policy.
Note:
If `do_resets` is None, it will be by default `np.array([True])`
which implies the policy will not be "vectorized", i.e. number of
parallel environments for training data sampling = 1.
Args:
do_resets (numpy.ndarray): Bool that indicates terminal state(s).
"""
if do_resets is None:
do_resets = np.array([True])
if self._prev_actions is None or len(do_resets) != len(
self._prev_actions):
self._prev_actions = np.zeros(
(len(do_resets), self.action_space.flat_dim))
self._prev_hiddens = np.zeros((len(do_resets), self._hidden_dim))
self._prev_actions[do_resets] = 0.
self._prev_hiddens[do_resets] = self.model.networks[
'default'].init_hidden.eval()
def get_action(self, observation):
"""Get single action from this policy for the input observation.
Args:
observation (numpy.ndarray): Observation from environment.
Returns:
numpy.ndarray: Actions
dict: Predicted action and agent information.
Note:
It returns an action and a dict, with keys
- mean (numpy.ndarray): Mean of the distribution.
- log_std (numpy.ndarray): Log standard deviation of the
distribution.
- prev_action (numpy.ndarray): Previous action, only present if
self._state_include_action is True.
"""
actions, agent_infos = self.get_actions([observation])
return actions[0], {k: v[0] for k, v in agent_infos.items()}
def get_actions(self, observations):
"""Get multiple actions from this policy for the input observations.
Args:
observations (numpy.ndarray): Observations from environment.
Returns:
numpy.ndarray: Actions
dict: Predicted action and agent information.
Note:
It returns an action and a dict, with keys
- mean (numpy.ndarray): Means of the distribution.
- log_std (numpy.ndarray): Log standard deviations of the
distribution.
- prev_action (numpy.ndarray): Previous action, only present if
self._state_include_action is True.
"""
if self._state_include_action:
assert self._prev_actions is not None
all_input = np.concatenate([observations, self._prev_actions],
axis=-1)
else:
all_input = observations
means, log_stds, hidden_vec = self._f_step_mean_std(
all_input, self._prev_hiddens)
rnd = np.random.normal(size=means.shape)
samples = rnd * np.exp(log_stds) + means
samples = self.action_space.unflatten_n(samples)
prev_actions = self._prev_actions
self._prev_actions = samples
self._prev_hiddens = hidden_vec
agent_infos = dict(mean=means, log_std=log_stds)
if self._state_include_action:
agent_infos['prev_action'] = np.copy(prev_actions)
return samples, agent_infos
@property
def distribution(self):
"""Policy distribution.
Returns:
tfp.Distribution.MultivariateNormalDiag: Policy distribution.
"""
return self.model.networks['default'].dist
@property
def state_info_specs(self):
"""State info specifcation.
Returns:
List[str]: keys and shapes for the information related to the
policy's state when taking an action.
"""
if self._state_include_action:
return [
('prev_action', (self._action_dim, )),
]
return []
def clone(self, name):
"""Return a clone of the policy.
It only copies the configuration of the primitive,
not the parameters.
Args:
name (str): Name of the newly created policy. It has to be
different from source policy if cloned under the same
computational graph.
Returns:
metarl.tf.policies.GaussianGRUPolicy: Newly cloned policy.
"""
return self.__class__(
name=name,
env_spec=self._env_spec,
hidden_dim=self._hidden_dim,
hidden_nonlinearity=self._hidden_nonlinearity,
hidden_w_init=self._hidden_w_init,
hidden_b_init=self._hidden_b_init,
recurrent_nonlinearity=self._recurrent_nonlinearity,
recurrent_w_init=self._recurrent_w_init,
output_nonlinearity=self._output_nonlinearity,
output_w_init=self._output_w_init,
output_b_init=self._output_b_init,
hidden_state_init=self._hidden_state_init,
hidden_state_init_trainable=self._hidden_state_init_trainable,
learn_std=self._learn_std,
std_share_network=self._std_share_network,
init_std=self._init_std,
layer_normalization=self._layer_normalization,
state_include_action=self._state_include_action)
def __getstate__(self):
"""Object.__getstate__.
Returns:
dict: the state to be pickled for the instance.
"""
new_dict = super().__getstate__()
del new_dict['_f_step_mean_std']
return new_dict