def _initialize(self):
input_var = tf.compat.v1.placeholder(tf.float32,
shape=(None, ) +
self._input_shape)
with tf.compat.v1.variable_scope(self._variable_scope):
self.model.build(input_var)
ys_var = tf.compat.v1.placeholder(dtype=tf.float32,
name='ys',
shape=(None, self._output_dim))
old_prob_var = tf.compat.v1.placeholder(dtype=tf.float32,
name='old_prob',
shape=(None,
self._output_dim))
y_hat = self.model.networks['default'].y_hat
old_info_vars = dict(prob=old_prob_var)
info_vars = dict(prob=y_hat)
self._dist = Categorical(self._output_dim)
mean_kl = tf.reduce_mean(
self._dist.kl_sym(old_info_vars, info_vars))
loss = -tf.reduce_mean(
self._dist.log_likelihood_sym(ys_var, info_vars))
predicted = tf.one_hot(tf.argmax(y_hat, axis=1),
depth=self._output_dim)
self._f_predict = tensor_utils.compile_function([input_var],
predicted)
self._f_prob = tensor_utils.compile_function([input_var], y_hat)
self._optimizer.update_opt(loss=loss,
target=self,
network_output=[y_hat],
inputs=[input_var, ys_var])
self._tr_optimizer.update_opt(
loss=loss,
target=self,
network_output=[y_hat],
inputs=[input_var, ys_var, old_prob_var],
leq_constraint=(mean_kl, self._max_kl_step))
def distribution(self):
"""Policy distribution.
Returns:
garage.tf.distributions.Categorical: Policy distribution.
"""
return Categorical(self._action_dim)
def __init__(
self,
env_spec,
conv_filters,
conv_filter_sizes,
conv_strides,
conv_pads,
hidden_sizes=[],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.softmax,
prob_network=None,
name="CategoricalConvPolicy",
):
"""
:param env_spec: A spec for the mdp.
:param hidden_sizes: list of sizes for the fully connected
hidden layers
:param hidden_nonlinearity: nonlinearity used for each hidden layer
:param prob_network: manually specified network for this policy, other
network params
are ignored
:return:
"""
assert isinstance(env_spec.action_space, Discrete)
Serializable.quick_init(self, locals())
self._name = name
self._env_spec = env_spec
self._prob_network_name = "prob_network"
with tf.variable_scope(name, "CategoricalConvPolicy"):
if prob_network is None:
prob_network = ConvNetwork(
input_shape=env_spec.observation_space.shape,
output_dim=env_spec.action_space.n,
conv_filters=conv_filters,
conv_filter_sizes=conv_filter_sizes,
conv_strides=conv_strides,
conv_pads=conv_pads,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
name="conv_prob_network",
)
with tf.name_scope(self._prob_network_name):
out_prob = L.get_output(prob_network.output_layer)
self._l_prob = prob_network.output_layer
self._l_obs = prob_network.input_layer
self._f_prob = tensor_utils.compile_function(
[prob_network.input_layer.input_var], [out_prob])
self._dist = Categorical(env_spec.action_space.n)
super(CategoricalConvPolicy, self).__init__(env_spec)
LayersPowered.__init__(self, [prob_network.output_layer])
def __init__(
self,
env_spec,
name='CategoricalMLPPolicy',
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
prob_network=None,
):
"""
CategoricalMLPPolicy.
A policy that uses a MLP to estimate a categorical distribution.
Args:
env_spec (garage.envs.env_spec.EnvSpec): Environment specification.
hidden_sizes (list[int]): Output dimension of dense layer(s).
For example, (32, 32) means the MLP of this policy consists
of two hidden layers, each with 32 hidden units.
hidden_nonlinearity: Activation function for
intermediate dense layer(s).
prob_network (tf.Tensor): manually specified network for this
policy. If None, a MLP with the network parameters will be
created. If not None, other network params are ignored.
"""
assert isinstance(env_spec.action_space, akro.Discrete)
Serializable.quick_init(self, locals())
self.name = name
self._prob_network_name = 'prob_network'
with tf.variable_scope(name, 'CategoricalMLPPolicy'):
if prob_network is None:
prob_network = MLP(
input_shape=(env_spec.observation_space.flat_dim, ),
output_dim=env_spec.action_space.n,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=tf.nn.softmax,
name=self._prob_network_name,
)
self._l_prob = prob_network.output_layer
self._l_obs = prob_network.input_layer
with tf.name_scope(self._prob_network_name):
prob_network_outputs = L.get_output(prob_network.output_layer)
self._f_prob = tensor_utils.compile_function(
[prob_network.input_layer.input_var], prob_network_outputs)
self._dist = Categorical(env_spec.action_space.n)
super(CategoricalMLPPolicy, self).__init__(env_spec)
LayersPowered.__init__(self, [prob_network.output_layer])
def __init__(
self,
env_spec,
name="CategoricalMLPPolicy",
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
prob_network=None,
):
"""
:param env_spec: A spec for the mdp.
:param hidden_sizes: list of sizes for the fully connected
hidden layers
:param hidden_nonlinearity: nonlinearity used for each hidden layer
:param prob_network: manually specified network for this policy, other
network params
are ignored
:return:
"""
assert isinstance(env_spec.action_space, Discrete)
Serializable.quick_init(self, locals())
self.name = name
self._prob_network_name = "prob_network"
with tf.variable_scope(name, "CategoricalMLPPolicy"):
if prob_network is None:
prob_network = MLP(
input_shape=(env_spec.observation_space.flat_dim, ),
output_dim=env_spec.action_space.n,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=tf.nn.softmax,
name=self._prob_network_name,
)
self._l_prob = prob_network.output_layer
self._l_obs = prob_network.input_layer
with tf.name_scope(self._prob_network_name):
prob_network_outputs = L.get_output(prob_network.output_layer)
self._f_prob = tensor_utils.compile_function(
[prob_network.input_layer.input_var], prob_network_outputs)
self._dist = Categorical(env_spec.action_space.n)
super(CategoricalMLPPolicy, self).__init__(env_spec)
LayersPowered.__init__(self, [prob_network.output_layer])
def distribution(self):
"""Policy distribution."""
return Categorical(self.action_dim)
class CategoricalMLPRegressorWithModel(StochasticRegressor2):
"""
CategoricalMLPRegressor with garage.tf.models.NormalizedInputMLPModel.
A class for performing regression (or classification, really) by fitting
a Categorical distribution to the outputs. Assumes that the output will
always be a one hot vector
Args:
input_shape (tuple[int]): Input shape of the training data. Since an
MLP model is used, implementation assumes flattened inputs. The
input shape of each data point should thus be of shape (x, ).
output_dim (int): Output dimension of the model.
name (str): Model name, also the variable scope.
hidden_sizes (list[int]): Output dimension of dense layer(s) for
the MLP for the network. For example, (32, 32) means the MLP
consists of two hidden layers, each with 32 hidden units.
hidden_nonlinearity (callable): Activation function for intermediate
dense layer(s). It should return a tf.Tensor. Set it to
None to maintain a tanh activation.
hidden_w_init (callable): Initializer function for the weight
of intermediate dense layer(s). The function should return a
tf.Tensor. Default is Glorot uniform initializer.
hidden_b_init (callable): Initializer function for the bias
of intermediate dense layer(s). The function should return a
tf.Tensor. Default is zero initializer.
output_nonlinearity (callable): Activation function for output dense
layer. It should return a tf.Tensor. Set it to None to
maintain a softmax activation.
output_w_init (callable): Initializer function for the weight
of output dense layer(s). The function should return a
tf.Tensor. Default is Glorot uniform initializer.
output_b_init (callable): Initializer function for the bias
of output dense layer(s). The function should return a
tf.Tensor. Default is zero initializer.
optimizer (garage.tf.Optimizer): Optimizer for minimizing the negative
log-likelihood. Defaults to LbsgsOptimizer
optimizer_args (dict): Arguments for the optimizer. Default is None,
which means no arguments.
tr_optimizer (garage.tf.Optimizer): Optimizer for trust region
approximation. Defaults to ConjugateGradientOptimizer.
tr_optimizer_args (dict): Arguments for the trust region optimizer.
Default is None, which means no arguments.
use_trust_region (bool): Whether to use trust region constraint.
max_kl_step (float): KL divergence constraint for each iteration.
normalize_inputs (bool): Bool for normalizing inputs or not.
layer_normalization (bool): Bool for using layer normalization or not.
"""
def __init__(self,
input_shape,
output_dim,
name='CategoricalMLPRegressorWithModel',
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
hidden_w_init=tf.glorot_uniform_initializer(),
hidden_b_init=tf.zeros_initializer(),
output_nonlinearity=tf.nn.softmax,
output_w_init=tf.glorot_uniform_initializer(),
output_b_init=tf.zeros_initializer(),
optimizer=None,
optimizer_args=None,
tr_optimizer=None,
tr_optimizer_args=None,
use_trust_region=True,
max_kl_step=0.01,
normalize_inputs=True,
layer_normalization=False):
super().__init__(input_shape, output_dim, name)
self._use_trust_region = use_trust_region
self._max_kl_step = max_kl_step
self._normalize_inputs = normalize_inputs
with tf.compat.v1.variable_scope(self._name, reuse=False) as vs:
self._variable_scope = vs
if optimizer_args is None:
optimizer_args = dict()
if tr_optimizer_args is None:
tr_optimizer_args = dict()
if optimizer is None:
optimizer = LbfgsOptimizer(**optimizer_args)
else:
optimizer = optimizer(**optimizer_args)
if tr_optimizer is None:
tr_optimizer = ConjugateGradientOptimizer(**tr_optimizer_args)
else:
tr_optimizer = tr_optimizer(**tr_optimizer_args)
self._optimizer = optimizer
self._tr_optimizer = tr_optimizer
self.model = NormalizedInputMLPModel(
input_shape,
output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
hidden_w_init=hidden_w_init,
hidden_b_init=hidden_b_init,
output_nonlinearity=output_nonlinearity,
output_w_init=output_w_init,
output_b_init=output_b_init,
layer_normalization=layer_normalization)
self._initialize()
def _initialize(self):
input_var = tf.compat.v1.placeholder(tf.float32,
shape=(None, ) +
self._input_shape)
with tf.compat.v1.variable_scope(self._variable_scope):
self.model.build(input_var)
ys_var = tf.compat.v1.placeholder(dtype=tf.float32,
name='ys',
shape=(None, self._output_dim))
old_prob_var = tf.compat.v1.placeholder(dtype=tf.float32,
name='old_prob',
shape=(None,
self._output_dim))
y_hat = self.model.networks['default'].y_hat
old_info_vars = dict(prob=old_prob_var)
info_vars = dict(prob=y_hat)
self._dist = Categorical(self._output_dim)
mean_kl = tf.reduce_mean(
self._dist.kl_sym(old_info_vars, info_vars))
loss = -tf.reduce_mean(
self._dist.log_likelihood_sym(ys_var, info_vars))
predicted = tf.one_hot(tf.argmax(y_hat, axis=1),
depth=self._output_dim)
self._f_predict = tensor_utils.compile_function([input_var],
predicted)
self._f_prob = tensor_utils.compile_function([input_var], y_hat)
self._optimizer.update_opt(loss=loss,
target=self,
network_output=[y_hat],
inputs=[input_var, ys_var])
self._tr_optimizer.update_opt(
loss=loss,
target=self,
network_output=[y_hat],
inputs=[input_var, ys_var, old_prob_var],
leq_constraint=(mean_kl, self._max_kl_step))
def fit(self, xs, ys):
"""
Fit with input data xs and label ys.
Args:
xs (numpy.ndarray): Input data.
ys (numpy.ndarray): Label of input data.
"""
if self._normalize_inputs:
# recompute normalizing constants for inputs
self.model.networks['default'].x_mean.load(
np.mean(xs, axis=0, keepdims=True))
self.model.networks['default'].x_std.load(
np.std(xs, axis=0, keepdims=True))
if self._use_trust_region:
# To use trust region constraint and optimizer
old_prob = self._f_prob(xs)
inputs = [xs, ys, old_prob]
optimizer = self._tr_optimizer
else:
inputs = [xs, ys]
optimizer = self._optimizer
loss_before = optimizer.loss(inputs)
tabular.record('{}/LossBefore'.format(self._name), loss_before)
optimizer.optimize(inputs)
loss_after = optimizer.loss(inputs)
tabular.record('{}/LossAfter'.format(self._name), loss_after)
tabular.record('{}/dLoss'.format(self._name), loss_before - loss_after)
self.first_optimized = True
def predict(self, xs):
"""
Predict ys based on input xs.
Args:
xs (numpy.ndarray): Input data.
Return:
The predicted ys (one hot vectors).
"""
return self._f_predict(xs)
def predict_log_likelihood(self, xs, ys):
"""
Predict log likelihood of output based on input xs and labels ys.
Args:
xs (numpy.ndarray): Input data.
ys (numpy.ndarray): Input labels in one hot representation.
Return:
The predicted log likelihoods.
"""
prob = self._f_prob(xs)
return self._dist.log_likelihood(ys, dict(prob=prob))
def dist_info_sym(self, x_var, name=None):
"""
Symbolic graph of the distribution.
Args:
x_var (tf.Tensor): Input tf.Tensor for the input data.
name (str): Name of the new graph.
Return:
tf.Tensor output of the symbolic graph of the distribution.
"""
with tf.compat.v1.variable_scope(self._variable_scope):
prob, _, _ = self.model.build(x_var, name=name)
return dict(prob=prob)
def log_likelihood_sym(self, x_var, y_var, name=None):
"""
Symbolic graph of the log likelihood.
Args:
x_var (tf.Tensor): Input tf.Tensor for the input data.
y_var (tf.Tensor): Input tf.Tensor for the one hot label of data.
name (str): Name of the new graph.
Return:
tf.Tensor output of the symbolic log likelihood.
"""
with tf.compat.v1.variable_scope(self._variable_scope):
prob, _, _ = self.model.build(x_var, name=name)
return self._dist.log_likelihood_sym(y_var, dict(prob=prob))
def get_params_internal(self, **args):
"""Get the params, which are the trainable variables."""
return self._variable_scope.trainable_variables()
def __getstate__(self):
"""Object.__getstate__."""
new_dict = super().__getstate__()
del new_dict['_f_predict']
del new_dict['_f_prob']
del new_dict['_dist']
return new_dict
def __setstate__(self, state):
"""Object.__setstate__."""
super().__setstate__(state)
self._initialize()
def __init__(
self,
input_shape,
output_dim,
name='CategoricalMLPRegressor',
prob_network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
optimizer=None,
tr_optimizer=None,
use_trust_region=True,
max_kl_step=0.01,
normalize_inputs=True,
no_initial_trust_region=True,
):
"""
:param input_shape: Shape of the input data.
:param output_dim: Dimension of output.
:param hidden_sizes: Number of hidden units of each layer of the mean
network.
:param hidden_nonlinearity: Non-linearity used for each layer of the
mean network.
:param optimizer: Optimizer for minimizing the negative log-likelihood.
:param use_trust_region: Whether to use trust region constraint.
:param max_kl_step: KL divergence constraint for each iteration
"""
Parameterized.__init__(self)
Serializable.quick_init(self, locals())
with tf.compat.v1.variable_scope(name, 'CategoricalMLPRegressor'):
if optimizer is None:
optimizer = LbfgsOptimizer()
if tr_optimizer is None:
tr_optimizer = ConjugateGradientOptimizer()
self.output_dim = output_dim
self.optimizer = optimizer
self.tr_optimizer = tr_optimizer
self._prob_network_name = 'prob_network'
if prob_network is None:
prob_network = MLP(input_shape=input_shape,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=tf.nn.softmax,
name=self._prob_network_name)
l_prob = prob_network.output_layer
LayersPowered.__init__(self, [l_prob])
xs_var = prob_network.input_layer.input_var
ys_var = tf.compat.v1.placeholder(dtype=tf.float32,
shape=[None, output_dim],
name='ys')
old_prob_var = tf.compat.v1.placeholder(dtype=tf.float32,
shape=[None, output_dim],
name='old_prob')
x_mean_var = tf.compat.v1.get_variable(
name='x_mean',
shape=(1, ) + input_shape,
initializer=tf.constant_initializer(0., dtype=tf.float32))
x_std_var = tf.compat.v1.get_variable(
name='x_std',
shape=(1, ) + input_shape,
initializer=tf.constant_initializer(1., dtype=tf.float32))
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
with tf.name_scope(self._prob_network_name,
values=[normalized_xs_var]):
prob_var = L.get_output(
l_prob, {prob_network.input_layer: normalized_xs_var})
old_info_vars = dict(prob=old_prob_var)
info_vars = dict(prob=prob_var)
dist = self._dist = Categorical(output_dim)
mean_kl = tf.reduce_mean(dist.kl_sym(old_info_vars, info_vars))
loss = -tf.reduce_mean(dist.log_likelihood_sym(ys_var, info_vars))
predicted = tf.one_hot(tf.argmax(prob_var, axis=1),
depth=output_dim)
self.prob_network = prob_network
self.f_predict = tensor_utils.compile_function([xs_var], predicted)
self.f_prob = tensor_utils.compile_function([xs_var], prob_var)
self.l_prob = l_prob
self.optimizer.update_opt(loss=loss,
target=self,
network_outputs=[prob_var],
inputs=[xs_var, ys_var])
self.tr_optimizer.update_opt(loss=loss,
target=self,
network_outputs=[prob_var],
inputs=[xs_var, ys_var, old_prob_var],
leq_constraint=(mean_kl, max_kl_step))
self.use_trust_region = use_trust_region
self.name = name
self.normalize_inputs = normalize_inputs
self.x_mean_var = x_mean_var
self.x_std_var = x_std_var
self.first_optimized = not no_initial_trust_region
class RecurrentCategorical(Distribution):
def __init__(self, dim, name="RecurrentCategorical"):
self._cat = Categorical(dim, name)
self._dim = dim
self._name = name
@property
def dim(self):
return self._dim
def kl_sym(self, old_dist_info_vars, new_dist_info_vars, name=None):
"""
Compute the symbolic KL divergence of two categorical distributions
"""
with tf.name_scope(name, "kl_sym",
[old_dist_info_vars, new_dist_info_vars]):
old_prob_var = old_dist_info_vars["prob"]
new_prob_var = new_dist_info_vars["prob"]
# Assume layout is N * T * A
return tf.reduce_sum(
old_prob_var *
(tf.log(old_prob_var + TINY) - tf.log(new_prob_var + TINY)),
axis=2)
def kl(self, old_dist_info, new_dist_info):
"""
Compute the KL divergence of two categorical distributions
"""
old_prob = old_dist_info["prob"]
new_prob = new_dist_info["prob"]
return np.sum(old_prob *
(np.log(old_prob + TINY) - np.log(new_prob + TINY)),
axis=2)
def likelihood_ratio_sym(self,
x_var,
old_dist_info_vars,
new_dist_info_vars,
name=None):
with tf.name_scope(name, "likelihood_ratio_sym",
[x_var, old_dist_info_vars, new_dist_info_vars]):
old_prob_var = old_dist_info_vars["prob"]
new_prob_var = new_dist_info_vars["prob"]
# Assume layout is N * T * A
a_dim = tf.shape(x_var)[2]
flat_ratios = self._cat.likelihood_ratio_sym(
tf.reshape(x_var, tf.stack([-1, a_dim])),
dict(prob=tf.reshape(old_prob_var, tf.stack([-1, a_dim]))),
dict(prob=tf.reshape(new_prob_var, tf.stack([-1, a_dim]))))
return tf.reshape(flat_ratios, tf.shape(old_prob_var)[:2])
def entropy(self, dist_info):
probs = dist_info["prob"]
return -np.sum(probs * np.log(probs + TINY), axis=2)
def entropy_sym(self, dist_info_vars, name=None):
with tf.name_scope(name, "entropy_sym", [dist_info_vars]):
probs = dist_info_vars["prob"]
return -tf.reduce_sum(probs * tf.log(probs + TINY), 2)
def log_likelihood_sym(self, xs, dist_info_vars, name=None):
with tf.name_scope(name, "log_likelihood_sym", [xs, dist_info_vars]):
probs = dist_info_vars["prob"]
# Assume layout is N * T * A
a_dim = tf.shape(probs)[2]
flat_logli = self._cat.log_likelihood_sym(
tf.reshape(xs, tf.stack([-1, a_dim])),
dict(prob=tf.reshape(probs, tf.stack((-1, a_dim)))))
return tf.reshape(flat_logli, tf.shape(probs)[:2])
def log_likelihood(self, xs, dist_info):
probs = dist_info["prob"]
# Assume layout is N * T * A
a_dim = tf.shape(probs)[2]
flat_logli = self._cat.log_likelihood_sym(
xs.reshape((-1, a_dim)), dict(prob=probs.reshape((-1, a_dim))))
return flat_logli.reshape(probs.shape[:2])
@property
def dist_info_specs(self):
return [("prob", (self.dim, ))]
def __init__(self, dim, name="RecurrentCategorical"):
self._cat = Categorical(dim, name)
self._dim = dim
self._name = name
