def test_gru_network():
from rllab.core.network import GRUNetwork
import lasagne.layers as L
from rllab.misc import ext
import numpy as np
network = GRUNetwork(
input_shape=(2, 3),
output_dim=5,
hidden_dim=4,
)
f_output = ext.compile_function(inputs=[network.input_layer.input_var],
outputs=L.get_output(network.output_layer))
assert f_output(np.zeros((6, 8, 2, 3))).shape == (6, 8, 5)
def __init__(self,
env_spec,
hidden_sizes=(32, ),
state_include_action=True,
hidden_nonlinearity=NL.tanh):
"""
:param env_spec: A spec for the env.
:param hidden_sizes: list of sizes for the fully connected hidden layers
:param hidden_nonlinearity: nonlinearity used for each hidden layer
:return:
"""
assert isinstance(env_spec.action_space, Discrete)
Serializable.quick_init(self, locals())
super(CategoricalGRUPolicy, self).__init__(env_spec)
assert len(hidden_sizes) == 1
if state_include_action:
input_shape = (env_spec.observation_space.flat_dim +
env_spec.action_space.flat_dim, )
else:
input_shape = (env_spec.observation_space.flat_dim, )
prob_network = GRUNetwork(
input_shape=input_shape,
output_dim=env_spec.action_space.n,
hidden_dim=hidden_sizes[0],
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=NL.softmax,
)
self._prob_network = prob_network
self._state_include_action = state_include_action
self._f_step_prob = ext.compile_function(
[
prob_network.step_input_layer.input_var,
prob_network.step_prev_hidden_layer.input_var
],
L.get_output([
prob_network.step_output_layer, prob_network.step_hidden_layer
]))
self._prev_action = None
self._prev_hidden = None
self._hidden_sizes = hidden_sizes
self._dist = RecurrentCategorical()
self.reset()
LasagnePowered.__init__(self, [prob_network.output_layer])
def __init__(
self,
input_shape,
output_dim,
predict_all=False, # CF
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 = GRUNetwork(
input_shape=input_shape,
output_dim=output_dim,
hidden_dim=hidden_sizes[0], # this gives 32 by default
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.itensor3("ys")
old_prob_var = TT.tensor3("old_prob")
x_mean_var = theano.shared(
np.zeros(
(
1,
1,
) + input_shape
), # this syntax makes the shape (1,1,*input_shape,). The first is traj
name="x_mean",
broadcastable=(
True,
True,
) + (False, ) * len(input_shape))
x_std_var = theano.shared(np.ones((
1,
1,
) + input_shape),
name="x_std",
broadcastable=(
True,
True,
) + (False, ) * len(input_shape))
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
prob_var_all = L.get_output(
l_prob, {prob_network.input_layer: normalized_xs_var})
if predict_all:
prob_var = prob_var_all
else:
# take only last dim but keep the shape
prob_var_last = TT.reshape(
prob_var_all[:, -1, :],
(TT.shape(prob_var_all)[0], 1, TT.shape(prob_var_all)[2]))
# padd along the time dimension to obtain the same shape as before
padded_prob_var = TT.tile(prob_var_last,
(1, TT.shape(prob_var_all)[1], 1))
# give it the standard name
prob_var = padded_prob_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_flat = special.to_onehot_sym(
TT.flatten(TT.argmax(prob_var, axis=-1)), output_dim)
predicted = TT.reshape(predicted_flat, TT.shape(prob_var))
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__(
self,
env_spec,
hidden_dim=32,
feature_network=None,
state_include_action=True,
hidden_nonlinearity=NL.tanh):
"""
:param env_spec: A spec for the env.
:param hidden_dim: dimension of hidden layer
:param hidden_nonlinearity: nonlinearity used for each hidden layer
:return:
"""
assert isinstance(env_spec.action_space, Discrete)
Serializable.quick_init(self, locals())
super(CategoricalGRUPolicy, self).__init__(env_spec)
obs_dim = env_spec.observation_space.flat_dim
action_dim = env_spec.action_space.flat_dim
if state_include_action:
input_dim = obs_dim + action_dim
else:
input_dim = obs_dim
l_input = L.InputLayer(
shape=(None, None, input_dim),
name="input"
)
if feature_network is None:
feature_dim = input_dim
l_flat_feature = None
l_feature = l_input
else:
feature_dim = feature_network.output_layer.output_shape[-1]
l_flat_feature = feature_network.output_layer
l_feature = OpLayer(
l_flat_feature,
extras=[l_input],
name="reshape_feature",
op=lambda flat_feature, input: TT.reshape(
flat_feature,
[input.shape[0], input.shape[1], feature_dim]
),
shape_op=lambda _, input_shape: (input_shape[0], input_shape[1], feature_dim)
)
prob_network = GRUNetwork(
input_shape=(feature_dim,),
input_layer=l_feature,
output_dim=env_spec.action_space.n,
hidden_dim=hidden_dim,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=TT.nnet.softmax,
name="prob_network"
)
self.prob_network = prob_network
self.feature_network = feature_network
self.l_input = l_input
self.state_include_action = state_include_action
flat_input_var = TT.matrix("flat_input")
if feature_network is None:
feature_var = flat_input_var
else:
feature_var = L.get_output(l_flat_feature, {feature_network.input_layer: flat_input_var})
self.f_step_prob = ext.compile_function(
[
flat_input_var,
prob_network.step_prev_hidden_layer.input_var
],
L.get_output([
prob_network.step_output_layer,
prob_network.step_hidden_layer
], {prob_network.step_input_layer: feature_var})
)
self.input_dim = input_dim
self.action_dim = action_dim
self.hidden_dim = hidden_dim
self.prev_action = None
self.prev_hidden = None
self.dist = RecurrentCategorical(env_spec.action_space.n)
out_layers = [prob_network.output_layer]
if feature_network is not None:
out_layers.append(feature_network.output_layer)
LasagnePowered.__init__(self, out_layers)
def __init__(
self,
env_spec,
hidden_sizes=(32, ),
state_include_action=True,
hidden_nonlinearity=NL.tanh,
learn_std=True,
init_std=1.0,
output_nonlinearity=None,
):
"""
:param env_spec: A spec for the env.
:param hidden_sizes: list of sizes for the fully connected hidden layers
:param hidden_nonlinearity: nonlinearity used for each hidden layer
:return:
"""
Serializable.quick_init(self, locals())
super(GaussianGRUPolicy, self).__init__(env_spec)
assert len(hidden_sizes) == 1
if state_include_action:
obs_dim = env_spec.observation_space.flat_dim + env_spec.action_space.flat_dim
else:
obs_dim = env_spec.observation_space.flat_dim
action_dim = env_spec.action_space.flat_dim
mean_network = GRUNetwork(
input_shape=(obs_dim, ),
output_dim=action_dim,
hidden_dim=hidden_sizes[0],
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
)
l_mean = mean_network.output_layer
obs_var = mean_network.input_var
l_log_std = ParamLayer(
mean_network.input_layer,
num_units=action_dim,
param=lasagne.init.Constant(np.log(init_std)),
name="output_log_std",
trainable=learn_std,
)
l_step_log_std = ParamLayer(
mean_network.step_input_layer,
num_units=action_dim,
param=l_log_std.param,
name="step_output_log_std",
trainable=learn_std,
)
self._mean_network = mean_network
self._l_log_std = l_log_std
self._state_include_action = state_include_action
self._f_step_mean_std = ext.compile_function(
[
mean_network.step_input_layer.input_var,
mean_network.step_prev_hidden_layer.input_var
],
L.get_output([
mean_network.step_output_layer, l_step_log_std,
mean_network.step_hidden_layer
]))
self._prev_action = None
self._prev_hidden = None
self._hidden_sizes = hidden_sizes
self._dist = RecurrentDiagonalGaussian(action_dim)
self.reset()
self.set_greedy(False)
LasagnePowered.__init__(self, [mean_network.output_layer, l_log_std])
def __init__(
self,
input_shape,
output_dim,
predict_all=False,
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 predict_all: use the prediction made at every step about the latent variables (not only the last step)
: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
p_network = GRUNetwork(
input_shape=input_shape,
output_dim=output_dim,
hidden_dim=hidden_sizes[0],
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=NL.sigmoid,
)
l_p = p_network.output_layer # this is every intermediate latent state! but I only care about last
LasagnePowered.__init__(self, [l_p])
xs_var = p_network.input_layer.input_var
ys_var = TT.itensor3("ys") # this is 3D: (traj, time, lat_dim)
old_p_var = TT.tensor3("old_p")
x_mean_var = theano.shared(np.zeros((
1,
1,
) + input_shape),
name="x_mean",
broadcastable=(
True,
True,
) + (False, ) * len(input_shape))
x_std_var = theano.shared(np.ones((
1,
1,
) + input_shape),
name="x_std",
broadcastable=(
True,
True,
) + (False, ) * len(input_shape))
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
# this is the previous p_var, from which I only want the last time-step padded along all time-steps
p_var_all = L.get_output(l_p,
{p_network.input_layer: normalized_xs_var})
# take only last dim but keep the shape
p_var_last = TT.reshape(
p_var_all[:, -1, :],
(TT.shape(p_var_all)[0], 1, TT.shape(p_var_all)[2]))
# padd along the time dimension to obtain the same shape as before
padded_p = TT.tile(p_var_last, (1, TT.shape(p_var_all)[1], 1))
# give it the standard name
if predict_all:
p_var = p_var_all
else:
p_var = padded_p
old_info_vars = dict(p=old_p_var)
info_vars = dict(
p=p_var
) # posterior of the latent at every step, wrt obs-act. Same along batch if recurrent
dist = self._dist = Bernoulli(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)) # regressor just wants to min -loglik of data ys
predicted = p_var >= 0.5
self._f_predict = ext.compile_function([xs_var], predicted)
self._f_p = ext.compile_function(
[xs_var], p_var
) # for consistency with gauss_mlp_reg this should be ._f_pdists
self._l_p = l_p
optimizer_args = dict(
loss=loss,
target=self,
network_outputs=[p_var],
)
if use_trust_region:
optimizer_args["leq_constraint"] = (mean_kl, step_size)
optimizer_args["inputs"] = [xs_var, ys_var, old_p_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