def __init__(
self,
env_spec,
hidden_sizes=(32, 32),
hidden_nonlinearity=NL.rectify,
hidden_W_init=LI.HeUniform(),
hidden_b_init=LI.Constant(0.),
output_nonlinearity=NL.tanh,
output_W_init=LI.Uniform(-3e-3, 3e-3),
output_b_init=LI.Uniform(-3e-3, 3e-3),
bn=False):
Serializable.quick_init(self, locals())
l_obs = L.InputLayer(shape=(None, env_spec.observation_space.flat_dim))
l_hidden = l_obs
if bn:
l_hidden = batch_norm(l_hidden)
for idx, size in enumerate(hidden_sizes):
l_hidden = L.DenseLayer(
l_hidden,
num_units=size,
W=hidden_W_init,
b=hidden_b_init,
nonlinearity=hidden_nonlinearity,
name="h%d" % idx
)
if bn:
l_hidden = batch_norm(l_hidden)
l_output = L.DenseLayer(
l_hidden,
num_units=env_spec.action_space.flat_dim,
W=output_W_init,
b=output_b_init,
nonlinearity=output_nonlinearity,
name="output"
)
# Note the deterministic=True argument. It makes sure that when getting
# actions from single observations, we do not update params in the
# batch normalization layers
action_var = L.get_output(l_output, deterministic=True)
self._output_layer = l_output
self._f_actions = ext.compile_function([l_obs.input_var], action_var)
super(DeterministicMLPPolicy, self).__init__(env_spec)
LasagnePowered.__init__(self, [l_output])
def __init__(
self,
name,
env_spec,
conv_filters,
conv_filter_sizes,
conv_strides,
conv_pads,
hidden_sizes=[],
hidden_nonlinearity=NL.rectify,
output_nonlinearity=NL.softmax,
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:
"""
Serializable.quick_init(self, locals())
assert isinstance(env_spec.action_space, Discrete)
self._env_spec = env_spec
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=NL.softmax,
name="prob_network",
)
self._l_prob = prob_network.output_layer
self._l_obs = prob_network.input_layer
self._f_prob = ext.compile_function(
[prob_network.input_layer.input_var],
L.get_output(prob_network.output_layer))
self._dist = Categorical(env_spec.action_space.n)
super(CategoricalConvPolicy, self).__init__(env_spec)
LasagnePowered.__init__(self, [prob_network.output_layer])
def update_opt(self,
loss,
target,
inputs,
extra_inputs=None,
gradients=None,
*args,
**kwargs):
"""
:param loss: Symbolic expression for the loss function.
:param target: A parameterized object to optimize over. It should implement methods of the
:class:`rllab_maml.core.paramerized.Parameterized` class.
:param leq_constraint: A constraint provided as a tuple (f, epsilon), of the form f(*inputs) <= epsilon.
:param inputs: A list of symbolic variables as inputs
:param gradients: symbolic expressions for the gradients of trainable parameters of the target. By default
this will be computed by calling theano.grad
:return: No return value.
"""
self._target = target
def get_opt_output(gradients):
if gradients is None:
gradients = theano.grad(loss,
target.get_params(trainable=True))
flat_grad = flatten_tensor_variables(gradients)
return [loss.astype('float64'), flat_grad.astype('float64')]
if extra_inputs is None:
extra_inputs = list()
self._opt_fun = lazydict(
f_loss=lambda: compile_function(inputs + extra_inputs, loss),
f_opt=lambda: compile_function(
inputs=inputs + extra_inputs,
outputs=get_opt_output(gradients),
))
def update_opt(self, loss, target, leq_constraint, inputs, constraint_name="constraint", *args, **kwargs):
"""
:param loss: Symbolic expression for the loss function.
:param target: A parameterized object to optimize over. It should implement methods of the
:class:`rllab_maml.core.paramerized.Parameterized` class.
:param leq_constraint: A constraint provided as a tuple (f, epsilon), of the form f(*inputs) <= epsilon.
:param inputs: A list of symbolic variables as inputs
:return: No return value.
"""
constraint_term, constraint_value = leq_constraint
penalty_var = TT.scalar("penalty")
penalized_loss = loss + penalty_var * constraint_term
self._target = target
self._max_constraint_val = constraint_value
self._constraint_name = constraint_name
def get_opt_output():
flat_grad = flatten_tensor_variables(theano.grad(
penalized_loss, target.get_params(trainable=True), disconnected_inputs='ignore'
))
return [penalized_loss.astype('float64'), flat_grad.astype('float64')]
self._opt_fun = lazydict(
f_loss=lambda: compile_function(inputs, loss, log_name="f_loss"),
f_constraint=lambda: compile_function(inputs, constraint_term, log_name="f_constraint"),
f_penalized_loss=lambda: compile_function(
inputs=inputs + [penalty_var],
outputs=[penalized_loss, loss, constraint_term],
log_name="f_penalized_loss",
),
f_opt=lambda: compile_function(
inputs=inputs + [penalty_var],
outputs=get_opt_output(),
log_name="f_opt"
)
)
def __init__(
self,
env_spec,
hidden_sizes=(32, 32),
hidden_nonlinearity=NL.tanh,
num_seq_inputs=1,
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:
"""
Serializable.quick_init(self, locals())
assert isinstance(env_spec.action_space, Discrete)
if prob_network is None:
prob_network = MLP(
input_shape=(env_spec.observation_space.flat_dim *
num_seq_inputs, ),
output_dim=env_spec.action_space.n,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=NL.softmax,
)
self._l_prob = prob_network.output_layer
self._l_obs = prob_network.input_layer
self._f_prob = ext.compile_function(
[prob_network.input_layer.input_var],
L.get_output(prob_network.output_layer))
self._dist = Categorical(env_spec.action_space.n)
super(CategoricalMLPPolicy, self).__init__(env_spec)
LasagnePowered.__init__(self, [prob_network.output_layer])
def __init__(
self,
input_shape,
output_dim,
prob_network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=NL.rectify,
optimizer=None,
use_trust_region=True,
step_size=0.01,
normalize_inputs=True,
name=None,
):
"""
: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 step_size: KL divergence constraint for each iteration
"""
Serializable.quick_init(self, locals())
if optimizer is None:
if use_trust_region:
optimizer = PenaltyLbfgsOptimizer()
else:
optimizer = LbfgsOptimizer()
self.output_dim = output_dim
self._optimizer = optimizer
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=NL.softmax,
)
l_prob = prob_network.output_layer
LasagnePowered.__init__(self, [l_prob])
xs_var = prob_network.input_layer.input_var
ys_var = TT.imatrix("ys")
old_prob_var = TT.matrix("old_prob")
x_mean_var = theano.shared(np.zeros((1, ) + input_shape),
name="x_mean",
broadcastable=(True, ) +
(False, ) * len(input_shape))
x_std_var = theano.shared(np.ones((1, ) + input_shape),
name="x_std",
broadcastable=(True, ) +
(False, ) * len(input_shape))
normalized_xs_var = (xs_var - x_mean_var) / x_std_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 = TT.mean(dist.kl_sym(old_info_vars, info_vars))
loss = -TT.mean(dist.log_likelihood_sym(ys_var, info_vars))
predicted = special.to_onehot_sym(TT.argmax(prob_var, axis=1),
output_dim)
self._f_predict = ext.compile_function([xs_var], predicted)
self._f_prob = ext.compile_function([xs_var], prob_var)
self._prob_network = prob_network
self._l_prob = l_prob
optimizer_args = dict(
loss=loss,
target=self,
network_outputs=[prob_var],
)
if use_trust_region:
optimizer_args["leq_constraint"] = (mean_kl, step_size)
optimizer_args["inputs"] = [xs_var, ys_var, old_prob_var]
else:
optimizer_args["inputs"] = [xs_var, ys_var]
self._optimizer.update_opt(**optimizer_args)
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
def init_opt(self):
is_recurrent = int(self.policy.recurrent)
obs_var = self.env.observation_space.new_tensor_variable(
'obs',
extra_dims=1 + is_recurrent,
)
action_var = self.env.action_space.new_tensor_variable(
'action',
extra_dims=1 + is_recurrent,
)
advantage_var = ext.new_tensor('advantage',
ndim=1 + is_recurrent,
dtype=theano.config.floatX)
dist = self.policy.distribution # TODO - this needs to be the dist of the fast policies.
old_dist_info_vars = {
k: ext.new_tensor('old_%s' % k,
ndim=2 + is_recurrent,
dtype=theano.config.floatX)
for k in dist.dist_info_keys
}
old_dist_info_vars_list = [
old_dist_info_vars[k] for k in dist.dist_info_keys
]
if is_recurrent:
valid_var = TT.matrix('valid')
else:
valid_var = None
state_info_vars = {
k: ext.new_tensor(k,
ndim=2 + is_recurrent,
dtype=theano.config.floatX)
for k in self.policy.state_info_keys
}
state_info_vars_list = [
state_info_vars[k] for k in self.policy.state_info_keys
]
dist_info_vars = self.policy.dist_info_sym(obs_var, state_info_vars)
logli = dist.log_likelihood_sym(action_var, dist_info_vars)
kl = dist.kl_sym(old_dist_info_vars, dist_info_vars)
# formulate as a minimization problem
# The gradient of the surrogate objective is the policy gradient
if is_recurrent:
surr_obj = -TT.sum(
logli * advantage_var * valid_var) / TT.sum(valid_var)
mean_kl = TT.sum(kl * valid_var) / TT.sum(valid_var)
max_kl = TT.max(kl * valid_var)
else:
surr_obj = -TT.mean(logli * advantage_var)
mean_kl = TT.mean(kl)
max_kl = TT.max(kl)
input_list = [obs_var, action_var, advantage_var
] + state_info_vars_list
if is_recurrent:
input_list.append(valid_var)
self.optimizer.update_opt(surr_obj,
target=self.policy,
inputs=input_list)
f_kl = ext.compile_function(
inputs=input_list + old_dist_info_vars_list,
outputs=[mean_kl, max_kl],
)
self.opt_info = dict(f_kl=f_kl, )