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, gym.spaces.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 __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,
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: sizes list 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 set_param_values(self, flattened_params, **tags):
return LasagnePowered.set_param_values(self, flattened_params, **tags)
def get_param_values(self, **tags):
return LasagnePowered.get_param_values(self, **tags)
def __init__(
self,
input_shape,
output_dim,
mean_network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=NL.rectify,
optimizer=None,
use_trust_region=True,
step_size=0.01,
learn_std=True,
init_std=1.0,
adaptive_std=False,
std_share_network=False,
std_hidden_sizes=(32, 32),
std_nonlinearity=None,
normalize_inputs=True,
normalize_outputs=True,
name=None,
batchsize=None,
subsample_factor=1.,
):
"""
: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
:param learn_std: Whether to learn the standard deviations. Only
effective if adaptive_std is False. If adaptive_std is True, this
parameter is ignored, and the weights for the std network are always
learned.
:param adaptive_std: Whether to make the std a function of the states.
:param std_share_network: Whether to use the same network as the mean.
:param std_hidden_sizes: Number of hidden units of each layer of the
std network. Only used if `std_share_network` is False. It defaults to
the same architecture as the mean.
:param std_nonlinearity: Non-linearity used for each layer of the std
network. Only used if `std_share_network`
is False. It defaults to the same non-linearity as the mean.
"""
Serializable.quick_init(self, locals())
self._batchsize = batchsize
self._subsample_factor = subsample_factor
if optimizer is None:
if use_trust_region:
optimizer = PenaltyLbfgsOptimizer()
else:
optimizer = LbfgsOptimizer()
self._optimizer = optimizer
if mean_network is None:
mean_network = MLP(
input_shape=input_shape,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=None,
)
l_mean = mean_network.output_layer
if adaptive_std:
l_log_std = MLP(
input_shape=input_shape,
input_var=mean_network.input_layer.input_var,
output_dim=output_dim,
hidden_sizes=std_hidden_sizes,
hidden_nonlinearity=std_nonlinearity,
output_nonlinearity=None,
).output_layer
else:
l_log_std = ParamLayer(
mean_network.input_layer,
num_units=output_dim,
param=lasagne.init.Constant(np.log(init_std)),
name="output_log_std",
trainable=learn_std,
)
LasagnePowered.__init__(self, [l_mean, l_log_std])
xs_var = mean_network.input_layer.input_var
ys_var = TT.matrix("ys")
old_means_var = TT.matrix("old_means")
old_log_stds_var = TT.matrix("old_log_stds")
x_mean_var = theano.shared(np.zeros((1, ) + input_shape,
dtype=theano.config.floatX),
name="x_mean",
broadcastable=(True, ) +
(False, ) * len(input_shape))
x_std_var = theano.shared(np.ones((1, ) + input_shape,
dtype=theano.config.floatX),
name="x_std",
broadcastable=(True, ) +
(False, ) * len(input_shape))
y_mean_var = theano.shared(np.zeros((1, output_dim),
dtype=theano.config.floatX),
name="y_mean",
broadcastable=(True, False))
y_std_var = theano.shared(np.ones((1, output_dim),
dtype=theano.config.floatX),
name="y_std",
broadcastable=(True, False))
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
normalized_ys_var = (ys_var - y_mean_var) / y_std_var
normalized_means_var = L.get_output(
l_mean, {mean_network.input_layer: normalized_xs_var})
normalized_log_stds_var = L.get_output(
l_log_std, {mean_network.input_layer: normalized_xs_var})
means_var = normalized_means_var * y_std_var + y_mean_var
log_stds_var = normalized_log_stds_var + TT.log(y_std_var)
normalized_old_means_var = (old_means_var - y_mean_var) / y_std_var
normalized_old_log_stds_var = old_log_stds_var - TT.log(y_std_var)
dist = self._dist = DiagonalGaussian(output_dim)
normalized_dist_info_vars = dict(mean=normalized_means_var,
log_std=normalized_log_stds_var)
mean_kl = TT.mean(
dist.kl_sym(
dict(mean=normalized_old_means_var,
log_std=normalized_old_log_stds_var),
normalized_dist_info_vars,
))
loss = - \
TT.mean(dist.log_likelihood_sym(
normalized_ys_var, normalized_dist_info_vars))
self._f_predict = compile_function([xs_var], means_var)
self._f_pdists = compile_function([xs_var], [means_var, log_stds_var])
self._l_mean = l_mean
self._l_log_std = l_log_std
optimizer_args = dict(
loss=loss,
target=self,
network_outputs=[normalized_means_var, normalized_log_stds_var],
)
if use_trust_region:
optimizer_args["leq_constraint"] = (mean_kl, step_size)
optimizer_args["inputs"] = [
xs_var, ys_var, old_means_var, old_log_stds_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._normalize_outputs = normalize_outputs
self._mean_network = mean_network
self._x_mean_var = x_mean_var
self._x_std_var = x_std_var
self._y_mean_var = y_mean_var
self._y_std_var = y_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_flat_dim = env_spec.action_space.flat_dim
if state_include_action:
input_dim = obs_dim + action_flat_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_flat_dim = action_flat_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: sizes list 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_flat_dim = env_spec.action_space.flat_dim
mean_network = GRUNetwork(
input_shape=(obs_dim, ),
output_dim=action_flat_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_flat_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_flat_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_flat_dim)
self.reset()
LasagnePowered.__init__(self, [mean_network.output_layer, l_log_std])
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__(
self,
env_spec,
hidden_sizes=(32, 32),
learn_std=True,
init_std=1.0,
adaptive_std=False,
std_share_network=False,
std_hidden_sizes=(32, 32),
min_std=1e-6,
std_hidden_nonlinearity=NL.tanh,
hidden_nonlinearity=NL.tanh,
output_nonlinearity=None,
mean_network=None,
std_network=None,
dist_cls=DiagonalGaussian,
):
"""
:param env_spec:
:param hidden_sizes: sizes list for the fully-connected hidden layers
:param learn_std: Is std trainable
:param init_std: Initial std
:param adaptive_std:
:param std_share_network:
:param std_hidden_sizes: sizes list for the fully-connected layers
for std
:param min_std: whether to make sure that the std is at least some
threshold value, to avoid numerical issues
:param std_hidden_nonlinearity:
:param hidden_nonlinearity: nonlinearity used for each hidden layer
:param output_nonlinearity: nonlinearity for the output layer
:param mean_network: custom network for the output mean
:param std_network: custom network for the output log std
:return:
"""
Serializable.quick_init(self, locals())
assert isinstance(env_spec.action_space, Box)
obs_dim = env_spec.observation_space.flat_dim
action_flat_dim = env_spec.action_space.flat_dim
# create network
if mean_network is None:
mean_network = MLP(
input_shape=(obs_dim, ),
output_dim=action_flat_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
)
self._mean_network = mean_network
l_mean = mean_network.output_layer
obs_var = mean_network.input_layer.input_var
if std_network is not None:
l_log_std = std_network.output_layer
else:
if adaptive_std:
std_network = MLP(
input_shape=(obs_dim, ),
input_layer=mean_network.input_layer,
output_dim=action_flat_dim,
hidden_sizes=std_hidden_sizes,
hidden_nonlinearity=std_hidden_nonlinearity,
output_nonlinearity=None,
)
l_log_std = std_network.output_layer
else:
l_log_std = ParamLayer(
mean_network.input_layer,
num_units=action_flat_dim,
param=lasagne.init.Constant(np.log(init_std)),
name="output_log_std",
trainable=learn_std,
)
self.min_std = min_std
mean_var, log_std_var = L.get_output([l_mean, l_log_std])
if self.min_std is not None:
log_std_var = TT.maximum(log_std_var, np.log(min_std))
self._mean_var, self._log_std_var = mean_var, log_std_var
self._l_mean = l_mean
self._l_log_std = l_log_std
self._dist = dist_cls(action_flat_dim)
LasagnePowered.__init__(self, [l_mean, l_log_std])
super(GaussianMLPPolicy, self).__init__(env_spec)
self._f_dist = ext.compile_function(
inputs=[obs_var],
outputs=[mean_var, log_std_var],
)
def __init__(self,
env_spec,
hidden_sizes=(32, 32),
hidden_nonlinearity=NL.rectify,
hidden_w_init=lasagne.init.HeUniform(),
hidden_b_init=lasagne.init.Constant(0.),
action_merge_layer=-2,
output_nonlinearity=None,
output_w_init=lasagne.init.Uniform(-3e-3, 3e-3),
output_b_init=lasagne.init.Uniform(-3e-3, 3e-3),
bn=False):
Serializable.quick_init(self, locals())
l_obs = L.InputLayer(shape=(None, env_spec.observation_space.flat_dim),
name="obs")
l_action = L.InputLayer(shape=(None, env_spec.action_space.flat_dim),
name="actions")
n_layers = len(hidden_sizes) + 1
if n_layers > 1:
action_merge_layer = \
(action_merge_layer % n_layers + n_layers) % n_layers
else:
action_merge_layer = 1
l_hidden = l_obs
for idx, size in enumerate(hidden_sizes):
if bn:
l_hidden = batch_norm(l_hidden)
if idx == action_merge_layer:
l_hidden = L.ConcatLayer([l_hidden, l_action])
l_hidden = L.DenseLayer(l_hidden,
num_units=size,
W=hidden_w_init,
b=hidden_b_init,
nonlinearity=hidden_nonlinearity,
name="h%d" % (idx + 1))
if action_merge_layer == n_layers:
l_hidden = L.ConcatLayer([l_hidden, l_action])
l_output = L.DenseLayer(l_hidden,
num_units=1,
W=output_w_init,
b=output_b_init,
nonlinearity=output_nonlinearity,
name="output")
output_var = L.get_output(l_output, deterministic=True).flatten()
self._f_qval = ext.compile_function(
[l_obs.input_var, l_action.input_var], output_var)
self._output_layer = l_output
self._obs_layer = l_obs
self._action_layer = l_action
self._output_nonlinearity = output_nonlinearity
LasagnePowered.__init__(self, [l_output])