def __init__(self,
env_spec,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.relu,
action_merge_layer=-2,
output_nonlinearity=None,
bn=False,
dropout=.05):
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,
nonlinearity=hidden_nonlinearity,
name="h%d" % (idx + 1))
l_hidden = L.DropoutLayer(l_hidden, dropout)
if action_merge_layer == n_layers:
l_hidden = L.ConcatLayer([l_hidden, l_action])
l_output = L.DenseLayer(l_hidden,
num_units=1,
nonlinearity=output_nonlinearity,
name="output")
output_var = L.get_output(l_output, deterministic=True)
output_var_drop = L.get_output(l_output, deterministic=False)
self._f_qval = tensor_utils.compile_function(
[l_obs.input_var, l_action.input_var], output_var)
self._f_qval_drop = tensor_utils.compile_function(
[l_obs.input_var, l_action.input_var], output_var_drop)
self._output_layer = l_output
self._obs_layer = l_obs
self._action_layer = l_action
self._output_nonlinearity = output_nonlinearity
LayersPowered.__init__(self, [l_output])
def __init__(
self,
name,
output_dim,
hidden_sizes,
hidden_nonlinearity,
output_nonlinearity,
hidden_W_init=L.XavierUniformInitializer(),
hidden_b_init=tf.zeros_initializer(),
output_W_init=L.XavierUniformInitializer(),
output_b_init=tf.zeros_initializer(),
input_var=None,
input_layer=None,
input_shape=None,
batch_normalization=False,
weight_normalization=False,
):
Serializable.quick_init(self, locals())
with tf.variable_scope(name):
if input_layer is None:
l_in = L.InputLayer(shape=(None, ) + input_shape,
input_var=input_var,
name="input")
else:
l_in = input_layer
self._layers = [l_in]
l_hid = l_in
if batch_normalization:
l_hid = L.batch_norm(l_hid)
for idx, hidden_size in enumerate(hidden_sizes):
l_hid = L.DenseLayer(l_hid,
num_units=hidden_size,
nonlinearity=hidden_nonlinearity,
name="hidden_%d" % idx,
W=hidden_W_init,
b=hidden_b_init,
weight_normalization=weight_normalization)
if batch_normalization:
l_hid = L.batch_norm(l_hid)
self._layers.append(l_hid)
l_out = L.DenseLayer(l_hid,
num_units=output_dim,
nonlinearity=output_nonlinearity,
name="output",
W=output_W_init,
b=output_b_init,
weight_normalization=weight_normalization)
if batch_normalization:
l_out = L.batch_norm(l_out)
self._layers.append(l_out)
self._l_in = l_in
self._l_out = l_out
# self._input_var = l_in.input_var
self._output = L.get_output(l_out)
LayersPowered.__init__(self, l_out)
def __init__(
self,
env_spec,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.relu,
action_merge_layer=-2,
output_nonlinearity=None,
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,
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,
nonlinearity=output_nonlinearity,
name="output"
)
output_var = L.get_output(l_output, deterministic=True)
self._f_qval = tensor_utils.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
LayersPowered.__init__(self, [l_output])
def __init__(self,
name,
env_spec,
oracle_policy,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh,
output_nonlinearity_binary=tf.nn.softmax,
output_dim_binary=2,
prob_network=None,
bn=False):
Serializable.quick_init(self, locals())
with tf.variable_scope(name):
if prob_network is None:
prob_network = SharedMLP(
input_shape=(env_spec.observation_space.flat_dim, ),
output_dim=env_spec.action_space.flat_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
output_nonlinearity_binary=output_nonlinearity_binary,
output_dim_binary=output_dim_binary,
# batch_normalization=True,
name="prob_network",
)
self.oracle_policy = oracle_policy
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],
L.get_output(prob_network.output_layer, deterministic=True))
self._f_prob_binary = tensor_utils.compile_function(
[prob_network.input_layer.input_var],
L.get_output(prob_network.output_layer_binary,
deterministic=True))
self.output_layer_binary = prob_network.output_layer_binary
self.binary_output = L.get_output(prob_network.output_layer_binary,
deterministic=True)
self.prob_network = prob_network
# 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.
# TODO: this doesn't currently work properly in the tf version so we leave out batch_norm
super(SharedDeterministicMLPPolicy, self).__init__(env_spec)
LayersPowered.__init__(
self,
[prob_network.output_layer, prob_network.output_layer_binary])
def __init__(self, *, network_cls, network_args):
Serializable.quick_init(self, locals())
logger.log('EmbeddingMutualInfo : {}'.format(locals()))
network = network_cls(**network_args)
self._network = network
output_layers = [self._network.output_layer]
self.output_layers = output_layers
LayersPowered.__init__(self, output_layers)
def __init__(
self,
name,
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,
):
"""
: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
# import pdb; pdb.set_trace()
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="prob_network",
)
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],
L.get_output(prob_network.output_layer))
self._dist = Categorical(env_spec.action_space.n)
super(CategoricalConvPolicy, self).__init__(env_spec)
LayersPowered.__init__(self, [prob_network.output_layer])
def __init__(
self,
name,
env_spec,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=tf.nn.tanh,
mean_network=None,
):
"""
:param env_spec:
:param hidden_sizes: list of sizes for the fully-connected hidden layers
: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
:return:
"""
Serializable.quick_init(self, locals())
assert isinstance(env_spec.action_space, Box)
with tf.variable_scope(name):
obs_dim = env_spec.observation_space.flat_dim
action_dim = env_spec.action_space.flat_dim
# create network
if mean_network is None:
mean_network = MLP(
name="mean_network",
input_shape=(obs_dim, ),
output_dim=action_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
self._l_mean = l_mean
action_var = L.get_output(self._l_mean, deterministic=True)
LayersPowered.__init__(self, [l_mean])
super(DeterministicMLPPolicy, self).__init__(env_spec)
self._f_actions = tensor_utils.compile_function(
inputs=[obs_var],
outputs=action_var,
)
def __init__(self,
name,
env_spec,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
gating_network=None,
input_layer=None,
num_options=4,
conv_filters=None,
conv_filter_sizes=None,
conv_strides=None,
conv_pads=None,
input_shape=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())
self.num_options = num_options
assert isinstance(env_spec.action_space, Discrete)
with tf.variable_scope(name):
input_layer, output_layer = self.make_network(
(env_spec.observation_space.flat_dim, ),
env_spec.action_space.n,
hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
gating_network=gating_network,
l_in=input_layer,
conv_filters=conv_filters,
conv_filter_sizes=conv_filter_sizes,
conv_strides=conv_strides,
conv_pads=conv_pads,
input_shape=input_shape)
self._l_prob = output_layer
self._l_obs = input_layer
self._f_prob = tensor_utils.compile_function(
[input_layer.input_var], L.get_output(output_layer))
self._dist = Categorical(env_spec.action_space.n)
super(CategoricalDecomposedPolicy, self).__init__(env_spec)
LayersPowered.__init__(self, [output_layer])
def __init__(
self,
name,
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,
):
"""
: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
# import pdb; pdb.set_trace()
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="prob_network",
)
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],
L.get_output(prob_network.output_layer)
)
self._dist = Categorical(env_spec.action_space.n)
super(CategoricalConvPolicy, self).__init__(env_spec)
LayersPowered.__init__(self, [prob_network.output_layer])
def __init__(self, name, output_dim, hidden_sizes, hidden_nonlinearity,
output_nonlinearity, hidden_W_init=L.XavierUniformInitializer(), hidden_b_init=tf.zeros_initializer,
output_W_init=L.XavierUniformInitializer(), output_b_init=tf.zeros_initializer,
input_var=None, input_layer=None, input_shape=None, batch_normalization=False, weight_normalization=False,
):
Serializable.quick_init(self, locals())
with tf.variable_scope(name):
if input_layer is None:
l_in = L.InputLayer(shape=(None,) + input_shape, input_var=input_var, name="input")
else:
l_in = input_layer
self._layers = [l_in]
l_hid = l_in
if batch_normalization:
l_hid = L.batch_norm(l_hid)
for idx, hidden_size in enumerate(hidden_sizes):
l_hid = L.DenseLayer(
l_hid,
num_units=hidden_size,
nonlinearity=hidden_nonlinearity,
name="hidden_%d" % idx,
W=hidden_W_init,
b=hidden_b_init,
weight_normalization=weight_normalization
)
if batch_normalization:
l_hid = L.batch_norm(l_hid)
self._layers.append(l_hid)
l_out = L.DenseLayer(
l_hid,
num_units=output_dim,
nonlinearity=output_nonlinearity,
name="output",
W=output_W_init,
b=output_b_init,
weight_normalization=weight_normalization
)
if batch_normalization:
l_out = L.batch_norm(l_out)
self._layers.append(l_out)
self._l_in = l_in
self._l_out = l_out
# self._input_var = l_in.input_var
self._output = L.get_output(l_out)
LayersPowered.__init__(self, l_out)
def __init__(
self,
name,
env_spec,
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:
"""
Serializable.quick_init(self, locals())
assert isinstance(env_spec.action_space, Discrete)
with tf.variable_scope(name):
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="prob_network",
)
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],
L.get_output(prob_network.output_layer)
)
self._dist = Categorical(env_spec.action_space.n)
super(CategoricalMLPPolicy, self).__init__(env_spec)
LayersPowered.__init__(self, [prob_network.output_layer])
def __init__(
self,
name,
env_spec,
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:
"""
Serializable.quick_init(self, locals())
assert isinstance(env_spec.action_space, Discrete)
with tf.variable_scope(name):
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="prob_network",
)
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],
L.get_output(prob_network.output_layer)
)
self._dist = Categorical(env_spec.action_space.n)
super(CategoricalMLPPolicy, self).__init__(env_spec)
LayersPowered.__init__(self, [prob_network.output_layer])
def __init__(self,
weight_manager: PropNetworkWeights,
problem_meta: ProblemMeta,
dropout: float=0.0,
norm_response: bool=False) -> None:
Serializable.quick_init(self, locals())
self._weight_manager = weight_manager
self._prob_meta = problem_meta
# ReLU dies before convergence on TTW-1
# self.nonlinearity = tf.nn.relu
# tanh learns better, but didn't fully converge on TTW-1 w/ one layer
# and saturated on deeper networks
# self.nonlinearity = tf.nn.tanh
# softplus learns even better on TTW-1 w/ more layers, but results in
# HUGE activations eventually, and ultimately NaNs (only after solving
# the problem fully, though)
# self.nonlinearity = tf.nn.softplus
# Leaky ReLU doesn't make activations explode, and does seem to result
# in some change in activation distribution. It doesn't learn quickly,
# however, and seems to suffer from vanishing gradients in earlier
# layers. Vanishing problem resolved after 60+ iterations, but gave way
# to exploding activation problem (at least when pooling with
# reduce_sum).
# self.nonlinearity = make_leaky_relu(0.1)
# elu takes a bit longer to converge than softplus (in wall time), but
# a bit less than the leaky relu above. Seems to lead to even less
# activation-explosion.
self.nonlinearity = tf.nn.elu
self.dropout = dropout
self.norm_response = norm_response
self._make_mlp()
LayersPowered.__init__(self, [self._l_out])
def __init__(
self,
name,
env_spec,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh,
prob_network=None,
bn=False):
Serializable.quick_init(self, locals())
with tf.variable_scope(name):
if prob_network is None:
prob_network = MLP(
input_shape=(env_spec.observation_space.flat_dim,),
output_dim=env_spec.action_space.flat_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
# batch_normalization=True,
name="prob_network",
)
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],
L.get_output(prob_network.output_layer, deterministic=True)
)
self.prob_network = prob_network
# 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.
# TODO: this doesn't currently work properly in the tf version so we leave out batch_norm
super(DeterministicMLPPolicy, self).__init__(env_spec)
LayersPowered.__init__(self, [prob_network.output_layer])
def __init__(
self,
name,
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=tf.nn.tanh,
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
mean_network=None,
std_network=None,
std_parametrization='exp',
# added arguments
w_auxiliary=False,
auxliary_classes=0.,
):
"""
:param env_spec:
:param hidden_sizes: list of sizes 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: list of sizes 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
:param std_parametrization: how the std should be parametrized. There are a few options:
- exp: the logarithm of the std will be stored, and applied a exponential transformation
- softplus: the std will be computed as log(1+exp(x))
:return:
"""
Serializable.quick_init(self, locals())
assert isinstance(env_spec.action_space, Box)
with tf.variable_scope(name):
obs_dim = env_spec.observation_space.flat_dim
action_dim = env_spec.action_space.flat_dim
# create network
if mean_network is None:
mean_network = MLP(
name="mean_network",
input_shape=(obs_dim,),
output_dim=action_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
w_auxiliary=w_auxiliary,
auxliary_classes=auxliary_classes,
)
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_std_param = std_network.output_layer
else:
if adaptive_std:
std_network = MLP(
name="std_network",
input_shape=(obs_dim,),
input_layer=mean_network.input_layer,
output_dim=action_dim,
hidden_sizes=std_hidden_sizes,
hidden_nonlinearity=std_hidden_nonlinearity,
output_nonlinearity=None,
)
l_std_param = std_network.output_layer
else:
if std_parametrization == 'exp':
init_std_param = np.log(init_std)
elif std_parametrization == 'softplus':
init_std_param = np.log(np.exp(init_std) - 1)
else:
raise NotImplementedError
l_std_param = L.ParamLayer(
mean_network.input_layer,
num_units=action_dim,
param=tf.constant_initializer(init_std_param),
name="output_std_param",
trainable=learn_std,
)
self.std_parametrization = std_parametrization
if std_parametrization == 'exp':
min_std_param = np.log(min_std)
elif std_parametrization == 'softplus':
min_std_param = np.log(np.exp(min_std) - 1)
else:
raise NotImplementedError
self.min_std_param = min_std_param
# mean_var, log_std_var = L.get_output([l_mean, l_std_param])
#
# if self.min_std_param is not None:
# log_std_var = tf.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_std_param = l_std_param
self._dist = DiagonalGaussian(action_dim)
outputs = [l_mean, l_std_param]
if w_auxiliary:
print('network.py: Using auxiliary model')
l_aux_pred = mean_network.aux_layer
self._l_aux_pred = l_aux_pred
aux_pred_var = self.auxiliary_pred_sym(mean_network.input_layer.input_var, dict())
self._f_aux_pred = tensor_utils.compile_function(
inputs=[obs_var],
outputs=aux_pred_var
)
outputs += [l_aux_pred]
LayersPowered.__init__(self, outputs)
super(GaussianMLPPolicy, self).__init__(env_spec)
dist_info_sym = self.dist_info_sym(mean_network.input_layer.input_var, dict())
mean_var = dist_info_sym["mean"]
log_std_var = dist_info_sym["log_std"]
self._f_dist = tensor_utils.compile_function(
inputs=[obs_var],
outputs=[mean_var, log_std_var],
)
def __init__(
self,
name,
env_spec,
hidden_dim=32,
feature_network=None,
state_include_action=True,
hidden_nonlinearity=tf.tanh,
learn_std=True,
init_std=1.0,
output_nonlinearity=None,
lstm_layer_cls=L.LSTMLayer,
):
"""
: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:
"""
with tf.variable_scope(name):
Serializable.quick_init(self, locals())
super(GaussianLSTMPolicy, 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 = L.OpLayer(
l_flat_feature,
extras=[l_input],
name="reshape_feature",
op=lambda flat_feature, input: tf.reshape(
flat_feature,
tf.stack([
tf.shape(input)[0],
tf.shape(input)[1], feature_dim
])),
shape_op=lambda _, input_shape:
(input_shape[0], input_shape[1], feature_dim))
mean_network = LSTMNetwork(input_shape=(feature_dim, ),
input_layer=l_feature,
output_dim=action_dim,
hidden_dim=hidden_dim,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
lstm_layer_cls=lstm_layer_cls,
name="mean_network")
l_log_std = L.ParamLayer(
mean_network.input_layer,
num_units=action_dim,
param=tf.constant_initializer(np.log(init_std)),
name="output_log_std",
trainable=learn_std,
)
l_step_log_std = L.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.feature_network = feature_network
self.l_input = l_input
self.state_include_action = state_include_action
flat_input_var = tf.placeholder(dtype=tf.float32,
shape=(None, input_dim),
name="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_mean_std = tensor_utils.compile_function(
[
flat_input_var,
mean_network.step_prev_state_layer.input_var,
],
L.get_output([
mean_network.step_output_layer, l_step_log_std,
mean_network.step_hidden_layer,
mean_network.step_cell_layer
], {mean_network.step_input_layer: feature_var}))
self.l_log_std = l_log_std
self.input_dim = input_dim
self.action_dim = action_dim
self.hidden_dim = hidden_dim
self.prev_actions = None
self.prev_hiddens = None
self.prev_cells = None
self.dist = RecurrentDiagonalGaussian(action_dim)
out_layers = [mean_network.output_layer, l_log_std]
if feature_network is not None:
out_layers.append(feature_network.output_layer)
LayersPowered.__init__(self, out_layers)
def set_param_values(self, flattened_params, **tags):
return LayersPowered.set_param_values(self, flattened_params, **tags)
def __init__(self, name, input_shape, output_dim,
conv_filters, conv_filter_sizes, conv_strides, conv_pads,
hidden_sizes, hidden_nonlinearity, output_nonlinearity,
hidden_W_init=L.XavierUniformInitializer(), hidden_b_init=tf.zeros_initializer,
output_W_init=L.XavierUniformInitializer(), output_b_init=tf.zeros_initializer,
input_var=None, input_layer=None, batch_normalization=False, weight_normalization=False):
Serializable.quick_init(self, locals())
"""
A network composed of several convolution layers followed by some fc layers.
input_shape: (width,height,channel)
HOWEVER, network inputs are assumed flattened. This network will first unflatten the inputs and then apply the standard convolutions and so on.
conv_filters: a list of numbers of convolution kernel
conv_filter_sizes: a list of sizes (int) of the convolution kernels
conv_strides: a list of strides (int) of the conv kernels
conv_pads: a list of pad formats (either 'SAME' or 'VALID')
hidden_nonlinearity: a nonlinearity from tf.nn, shared by all conv and fc layers
hidden_sizes: a list of numbers of hidden units for all fc layers
"""
with tf.variable_scope(name):
if input_layer is not None:
l_in = input_layer
l_hid = l_in
elif len(input_shape) == 3:
l_in = L.InputLayer(shape=(None, np.prod(input_shape)), input_var=input_var, name="input")
l_hid = L.reshape(l_in, ([0],) + input_shape, name="reshape_input")
elif len(input_shape) == 2:
l_in = L.InputLayer(shape=(None, np.prod(input_shape)), input_var=input_var, name="input")
input_shape = (1,) + input_shape
l_hid = L.reshape(l_in, ([0],) + input_shape, name="reshape_input")
else:
l_in = L.InputLayer(shape=(None,) + input_shape, input_var=input_var, name="input")
l_hid = l_in
if batch_normalization:
l_hid = L.batch_norm(l_hid)
for idx, conv_filter, filter_size, stride, pad in zip(
range(len(conv_filters)),
conv_filters,
conv_filter_sizes,
conv_strides,
conv_pads,
):
l_hid = L.Conv2DLayer(
l_hid,
num_filters=conv_filter,
filter_size=filter_size,
stride=(stride, stride),
pad=pad,
nonlinearity=hidden_nonlinearity,
name="conv_hidden_%d" % idx,
weight_normalization=weight_normalization,
)
if batch_normalization:
l_hid = L.batch_norm(l_hid)
if output_nonlinearity == L.spatial_expected_softmax:
assert len(hidden_sizes) == 0
assert output_dim == conv_filters[-1] * 2
l_hid.nonlinearity = tf.identity
l_out = L.SpatialExpectedSoftmaxLayer(l_hid)
else:
l_hid = L.flatten(l_hid, name="conv_flatten")
for idx, hidden_size in enumerate(hidden_sizes):
l_hid = L.DenseLayer(
l_hid,
num_units=hidden_size,
nonlinearity=hidden_nonlinearity,
name="hidden_%d" % idx,
W=hidden_W_init,
b=hidden_b_init,
weight_normalization=weight_normalization,
)
if batch_normalization:
l_hid = L.batch_norm(l_hid)
l_out = L.DenseLayer(
l_hid,
num_units=output_dim,
nonlinearity=output_nonlinearity,
name="output",
W=output_W_init,
b=output_b_init,
weight_normalization=weight_normalization,
)
if batch_normalization:
l_out = L.batch_norm(l_out)
self._l_in = l_in
self._l_out = l_out
# self._input_var = l_in.input_var
LayersPowered.__init__(self, l_out)
def __init__(self,
env_spec,
name='qnet',
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.relu,
action_merge_layer=-2,
output_nonlinearity=None,
eqf_use_full_qf=False,
eqf_sample_size=1,
mqprop=False,
bn=False):
Serializable.quick_init(self, locals())
assert not env_spec.action_space.is_discrete
self._env_spec = env_spec
with tf.variable_scope(name):
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,
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,
nonlinearity=output_nonlinearity,
name="output")
output_var = L.get_output(l_output, deterministic=True)
output_var = tf.reshape(output_var, (-1, ))
self._f_qval = tensor_utils.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
self.eqf_use_full_qf = eqf_use_full_qf
self.eqf_sample_size = eqf_sample_size
self.mqprop = mqprop
LayersPowered.__init__(self, [l_output])
def __init__(self,
name,
input_shape,
extra_input_shape,
output_dim,
hidden_sizes,
conv_filters,
conv_filter_sizes,
conv_strides,
conv_pads,
extra_hidden_sizes=None,
hidden_W_init=L.XavierUniformInitializer(),
hidden_b_init=tf.zeros_initializer(),
output_W_init=L.XavierUniformInitializer(),
output_b_init=tf.zeros_initializer(),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=None,
input_var=None,
input_layer=None):
Serializable.quick_init(self, locals())
if extra_hidden_sizes is None:
extra_hidden_sizes = []
with tf.variable_scope(name):
input_flat_dim = np.prod(input_shape)
extra_input_flat_dim = np.prod(extra_input_shape)
total_input_flat_dim = input_flat_dim + extra_input_flat_dim
if input_layer is None:
l_in = L.InputLayer(shape=(None, total_input_flat_dim),
input_var=input_var,
name="input")
else:
l_in = input_layer
l_conv_in = L.reshape(L.SliceLayer(l_in,
indices=slice(input_flat_dim),
name="conv_slice"),
([0], ) + input_shape,
name="conv_reshaped")
l_extra_in = L.reshape(L.SliceLayer(l_in,
indices=slice(
input_flat_dim, None),
name="extra_slice"),
([0], ) + extra_input_shape,
name="extra_reshaped")
l_conv_hid = l_conv_in
for idx, conv_filter, filter_size, stride, pad in zip(
range(len(conv_filters)),
conv_filters,
conv_filter_sizes,
conv_strides,
conv_pads,
):
l_conv_hid = L.Conv2DLayer(
l_conv_hid,
num_filters=conv_filter,
filter_size=filter_size,
stride=(stride, stride),
pad=pad,
nonlinearity=hidden_nonlinearity,
name="conv_hidden_%d" % idx,
)
l_extra_hid = l_extra_in
for idx, hidden_size in enumerate(extra_hidden_sizes):
l_extra_hid = L.DenseLayer(
l_extra_hid,
num_units=hidden_size,
nonlinearity=hidden_nonlinearity,
name="extra_hidden_%d" % idx,
W=hidden_W_init,
b=hidden_b_init,
)
l_joint_hid = L.concat(
[L.flatten(l_conv_hid, name="conv_hidden_flat"), l_extra_hid],
name="joint_hidden")
for idx, hidden_size in enumerate(hidden_sizes):
l_joint_hid = L.DenseLayer(
l_joint_hid,
num_units=hidden_size,
nonlinearity=hidden_nonlinearity,
name="joint_hidden_%d" % idx,
W=hidden_W_init,
b=hidden_b_init,
)
l_out = L.DenseLayer(
l_joint_hid,
num_units=output_dim,
nonlinearity=output_nonlinearity,
name="output",
W=output_W_init,
b=output_b_init,
)
self._l_in = l_in
self._l_out = l_out
LayersPowered.__init__(self, [l_out], input_layers=[l_in])
def __init__(
self,
name,
input_shape,
output_dim,
prob_network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
optimizer=None,
tr_optimizer=None,
use_trust_region=True,
step_size=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 step_size: KL divergence constraint for each iteration
"""
Serializable.quick_init(self, locals())
with tf.variable_scope(name):
if optimizer is None:
optimizer = LbfgsOptimizer(name="optimizer")
if tr_optimizer is None:
tr_optimizer = ConjugateGradientOptimizer()
self.output_dim = output_dim
self.optimizer = optimizer
self.tr_optimizer = tr_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=tf.nn.softmax,
name="prob_network"
)
l_prob = prob_network.output_layer
LayersPowered.__init__(self, [l_prob])
xs_var = prob_network.input_layer.input_var
ys_var = tf.placeholder(dtype=tf.float32, shape=[None, output_dim], name="ys")
old_prob_var = tf.placeholder(dtype=tf.float32, shape=[None, output_dim], name="old_prob")
x_mean_var = tf.get_variable(
name="x_mean",
shape=(1,) + input_shape,
initializer=tf.constant_initializer(0., dtype=tf.float32)
)
x_std_var = tf.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
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 = tensor_utils.to_onehot_sym(tf.argmax(prob_var, dimension=1), 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, step_size)
)
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
def get_param_values(self, **tags):
return LayersPowered.get_param_values(self, **tags)
def __init__(self, name, input_shape, output_dim,
conv_filters, conv_filter_sizes, conv_strides, conv_pads,
hidden_sizes, hidden_nonlinearity, output_nonlinearity,
hidden_W_init=L.XavierUniformInitializer(), hidden_b_init=tf.zeros_initializer(),
output_W_init=L.XavierUniformInitializer(), output_b_init=tf.zeros_initializer(),
input_var=None, input_layer=None, batch_normalization=False, weight_normalization=False):
Serializable.quick_init(self, locals())
"""
A network composed of several convolution layers followed by some fc layers.
input_shape: (width,height,channel)
HOWEVER, network inputs are assumed flattened. This network will first unflatten the inputs and then apply the standard convolutions and so on.
conv_filters: a list of numbers of convolution kernel
conv_filter_sizes: a list of sizes (int) of the convolution kernels
conv_strides: a list of strides (int) of the conv kernels
conv_pads: a list of pad formats (either 'SAME' or 'VALID')
hidden_nonlinearity: a nonlinearity from tf.nn, shared by all conv and fc layers
hidden_sizes: a list of numbers of hidden units for all fc layers
"""
with tf.variable_scope(name):
if input_layer is not None:
l_in = input_layer
l_hid = l_in
elif len(input_shape) == 3:
l_in = L.InputLayer(shape=(None, np.prod(input_shape)), input_var=input_var, name="input")
l_hid = L.reshape(l_in, ([0],) + input_shape, name="reshape_input")
elif len(input_shape) == 2:
l_in = L.InputLayer(shape=(None, np.prod(input_shape)), input_var=input_var, name="input")
input_shape = (1,) + input_shape
l_hid = L.reshape(l_in, ([0],) + input_shape, name="reshape_input")
else:
l_in = L.InputLayer(shape=(None,) + input_shape, input_var=input_var, name="input")
l_hid = l_in
if batch_normalization:
l_hid = L.batch_norm(l_hid)
critical_size = hidden_sizes[0]
for idx, conv_filter, filter_size, stride, pad in zip(range(len(conv_filters)),
conv_filters,
conv_filter_sizes,
conv_strides,
conv_pads,
):
l_hid = L.Conv2DLayer(
l_hid,
num_filters=conv_filter,
filter_size=filter_size,
stride=(stride, stride),
pad=pad,
nonlinearity=hidden_nonlinearity,
name="SL_conv_hidden_%d" % idx,
weight_normalization=weight_normalization,
)
if batch_normalization:
l_hid = L.batch_norm(l_hid)
l_hid = L.flatten(l_hid, name="conv_flatten")
critical_layer = L.DenseLayer(
l_hid,
num_units=hidden_sizes[0],
nonlinearity=None,
name="SL_fc",
W=hidden_W_init,
b=hidden_b_init,
weight_normalization=weight_normalization,
)
#critical_layer = L.flatten(critical_layer)
# if output_nonlinearity == L.spatial_expected_softmax:
# assert len(hidden_sizes) == 0
# assert output_dim == conv_filters[-1] * 2
# l_hid.nonlinearity = tf.identity
# l_out = L.SpatialExpectedSoftmaxLayer(l_hid)
self.actValues = L.get_output(critical_layer)
#list_rem = hidden_sizes[1:]
#####Forward pass block#################################
with tf.variable_scope("PG"):
# fcFor = L.DenseLayer(
# critical_layer,
# num_units = hidden_sizes[1],
# nonlinearity=hidden_nonlinearity,
# name="pgLayer_init",
# W=hidden_W_init,
# b=hidden_b_init,
# weight_normalization=weight_normalization,
# )
fc_1 = L.DenseLayer(
critical_layer,
num_units=hidden_sizes[1],
nonlinearity=hidden_nonlinearity,
name="pgLayer_1",
W=hidden_W_init,
b=hidden_b_init,
weight_normalization=weight_normalization,
)
fc_2 = L.DenseLayer(
fc_1,
num_units=hidden_sizes[2],
nonlinearity=hidden_nonlinearity,
name="pgLayer_2" ,
W=hidden_W_init,
b=hidden_b_init,
weight_normalization=weight_normalization,
)
if batch_normalization:
fc_2 = L.batch_norm(fcFor)
fcOut = L.DenseLayer(
fc_2,
num_units=output_dim,
nonlinearity=output_nonlinearity,
name="output",
W=output_W_init,
b=output_b_init,
weight_normalization=weight_normalization,
)
if batch_normalization:
fcOut = L.batch_norm(fcOut)
###################################################
self.actVariable = tf.Variable(initial_value = tf.zeros([ 10000, 32], dtype = tf.float32),name = "act_var1", trainable = True)
bcOut = fcOut.get_output_for(fc_2.get_output_for(fc_1.get_output_for(self.actVariable)))
self.bcOut = bcOut
backOutLayer = L.InputLayer(shape = (), input_var= bcOut , name="OutputLayer")
# shape is (actVariable[0] , 2)
self._l_in = l_in
self.forwardOutLayer = fcOut
self.backOutLayer = backOutLayer
outLayers = [fcOut, backOutLayer]
# self._input_var = l_in.input_var
LayersPowered.__init__(self, outLayers)
def __init__(
self,
name,
input_shape,
output_dim,
network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
optimizer=None,
normalize_inputs=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.
"""
Serializable.quick_init(self, locals())
with tf.variable_scope(name):
if optimizer is None:
optimizer = LbfgsOptimizer(name="optimizer")
self.output_dim = output_dim
self.optimizer = optimizer
if network is None:
network = MLP(
input_shape=input_shape,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
name="network"
)
l_out = network.output_layer
LayersPowered.__init__(self, [l_out])
xs_var = network.input_layer.input_var
ys_var = tf.placeholder(dtype=tf.float32, shape=[None, output_dim], name="ys")
x_mean_var = tf.get_variable(
name="x_mean",
shape=(1,) + input_shape,
initializer=tf.constant_initializer(0., dtype=tf.float32)
)
x_std_var = tf.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
fit_ys_var = L.get_output(l_out, {network.input_layer: normalized_xs_var})
loss = - tf.reduce_mean(tf.square(fit_ys_var - ys_var))
self.f_predict = tensor_utils.compile_function([xs_var], fit_ys_var)
optimizer_args = dict(
loss=loss,
target=self,
network_outputs=[fit_ys_var],
)
optimizer_args["inputs"] = [xs_var, ys_var]
self.optimizer.update_opt(**optimizer_args)
self.name = name
self.l_out = l_out
self.normalize_inputs = normalize_inputs
self.x_mean_var = x_mean_var
self.x_std_var = x_std_var
def __init__(self,
*,
state_network_cls,
state_network_args,
action_network_cls,
action_network_args,
env_spec,
reconciler_cls=None,
reconciler_args=dict(),
mutualinfo_model_cls=None,
mutualinfo_action_model_args=None,
mutualinfo_obs_model_args=None):
"""Use given state embedding network and one FC for action embedding."""
Serializable.quick_init(self, locals())
logger.log('GeneralEmbedding: {}'.format(locals()))
self.embedding_dim = state_network_args['output_dim']
self.env_spec = env_spec
state_network = state_network_cls(**state_network_args)
self._state_network = state_network
self._l_state = state_network.input_layer
self._l_phi = state_network.output_layer
action_network = action_network_cls(**action_network_args)
self.action_network = action_network
self._l_action = action_network.input_layer
self._l_psi = action_network.output_layer
output_layers = [self._l_phi, self._l_psi]
if reconciler_cls is not None:
reconciler_state_input_dim = env_spec.observation_space.flat_dim
reconciler_args['state_input_dim'] = reconciler_state_input_dim
reconciler_args['common_network_args']['input_shape'] = (
reconciler_state_input_dim + env_spec.action_space.flat_dim, )
reconciler_args['env_spec'] = env_spec
self.reconciler = reconciler_cls(**reconciler_args)
output_layers.extend(self.reconciler.output_layers)
else:
self.reconciler = None
if mutualinfo_model_cls is not None:
if mutualinfo_action_model_args is not None:
self.mutualinfo_action_model = mutualinfo_model_cls(
**mutualinfo_action_model_args)
output_layers.extend(
self.mutualinfo_action_model.output_layers)
else:
self.mutualinfo_action_model = None
if mutualinfo_obs_model_args is not None:
self.mutualinfo_obs_model = mutualinfo_model_cls(
**mutualinfo_obs_model_args)
output_layers.extend(self.mutualinfo_obs_model.output_layers)
else:
self.mutualinfo_obs_model = None
LayersPowered.__init__(self, output_layers)
phi_output = L.get_output(self._l_phi)
psi_output = L.get_output(self._l_psi)
self._obs_to_phi = tensor_utils.compile_function(
inputs=[self._l_state.input_var], outputs=phi_output)
self._action_to_psi = tensor_utils.compile_function(
inputs=[self._l_action.input_var], outputs=psi_output)
if self.reconciler is not None:
self._obs_action_to_reconciler = tensor_utils.compile_function(
inputs=[
self.reconciler.state_input_layer.input_var,
self.reconciler.action_input_layer.input_var
],
outputs=L.get_output(self.reconciler.output_layer))
def __init__(self,
env_spec,
name='QuadraticPhinet',
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=None,
vs_form=None,
bn=False,
A=None,
init_a=1.0,
a_parameterization='exp'):
Serializable.quick_init(self, locals())
assert not env_spec.action_space.is_discrete
self._env_spec = env_spec
self.vs_form = vs_form
with tf.variable_scope(name):
obs_dim = env_spec.observation_space.flat_dim
action_dim = env_spec.action_space.flat_dim
l_act = L.InputLayer(shape=(None, action_dim), name="action")
action_var = l_act.input_var
l_obs = L.InputLayer(shape=(None, obs_dim), name="obs")
self.obs_rms = RunningMeanStd(shape=(obs_dim, ))
obz = L.NormalizeLayer(l_obs, rms=self.obs_rms)
l_hidden = l_obs
hidden_sizes += (action_dim, )
for idx, size in enumerate(hidden_sizes):
if bn:
l_hidden = batch_norm(l_hidden)
l_hidden = L.DenseLayer(l_hidden,
num_units=size,
nonlinearity=hidden_nonlinearity,
name="h%d" % (idx + 1))
obs_var = l_obs.input_var
fs = l_hidden # fs_network.output_layer
if A is not None:
l_A_param = A.output_layer
else:
if a_parameterization == 'exp':
init_a_param = np.log(init_a) - .5
elif a_parameterization == 'softplus':
init_a_param = np.log(np.exp(init_a) - 1)
else:
raise NotImplementedError
l_log_A = L.ParamLayer(
l_obs,
num_units=action_dim,
param=tf.constant_initializer(init_a_param),
name="diagonal_a_matrix",
trainable=True)
if vs_form is not None:
raise NotImplementedError
self._l_log_A = l_log_A
self.a_parameterization = a_parameterization
self.fs = fs
if vs_form is not None:
self._output_vs = vs
LayersPowered.__init__(
self, [self.fs, self._l_log_A, self._output_vs])
else:
LayersPowered.__init__(self, [self.fs, self._l_log_A])
output_var = self.get_phival_sym(obs_var, action_var)
self._f_phival = tensor_utils.compile_function(
inputs=[obs_var, action_var], outputs=output_var)
def set_param_values(self, flattened_params, **tags):
return LayersPowered.set_param_values(self, flattened_params, **tags)
def __init__(self,
env_spec,
name='MLPPhinet',
hidden_sizes=(100, 100),
hidden_nonlinearity=tf.nn.relu,
action_merge_layer=-2,
output_nonlinearity=None,
vs_form=None,
bn=False):
Serializable.quick_init(self, locals())
assert not env_spec.action_space.is_discrete
self._env_spec = env_spec
self.vs_form = vs_form
with tf.variable_scope(name):
obs_dim = env_spec.observation_space.flat_dim
action_dim = env_spec.action_space.flat_dim
l_obs = L.InputLayer(shape=(None, obs_dim), name="obs")
l_action = L.InputLayer(shape=(None, action_dim), name="action")
self.obs_rms = RunningMeanStd(shape=(obs_dim, ))
obz = L.NormalizeLayer(l_obs,
rms=self.obs_rms,
clip_min=-5.,
clip_max=5.)
obs_hidden = L.DenseLayer(obz,
num_units=hidden_sizes[0],
nonlinearity=hidden_nonlinearity,
name="obs_h%d" % (0))
print("hidden sizes...", hidden_sizes[0], hidden_sizes[1:])
act_hidden = L.DenseLayer(l_action,
num_units=hidden_sizes[0],
nonlinearity=hidden_nonlinearity,
name="act_h%d" % (0))
merge_hidden = L.OpLayer(obs_hidden,
op=lambda x, y: x + y,
shape_op=lambda x, y: y,
extras=[act_hidden])
l_hidden = merge_hidden
for idx, size in enumerate(hidden_sizes[1:]):
if bn:
l_hidden = batch_norm(l_hidden)
l_hidden = L.DenseLayer(l_hidden,
num_units=size,
nonlinearity=hidden_nonlinearity,
name="h%d" % (idx + 1))
l_output = L.DenseLayer(l_hidden,
num_units=1,
nonlinearity=output_nonlinearity,
name="output")
if vs_form is not None:
if vs_form == 'linear':
vs = L.DenseLayer(l_obs,
num_units=1,
nonlinearity=None,
name='vs')
elif vs_form == 'mlp':
vs = L.DenseLayer(l_obs,
num_units=64,
nonlinearity=tf.nn.relu,
name='hidden_vs')
vs = L.DenseLayer(vs,
num_units=1,
nonlinearity=None,
name='vs')
else:
raise NotImplementedError
output_var = L.get_output(l_output, deterministic=True) + \
L.get_output(vs, deterministic=True)
output_var = tf.reshape(output_var, (-1, ))
else:
output_var = L.get_output(l_output, deterministic=True)
output_var = tf.reshape(output_var, (-1, ))
self._f_phival = tensor_utils.compile_function(
inputs=[l_obs.input_var, l_action.input_var],
outputs=output_var)
self._output_layer = l_output
self._obs_layer = l_obs
self._action_layer = l_action
self.output_nonlinearity = output_nonlinearity
if vs_form is not None:
self._output_vs = vs
LayersPowered.__init__(self, [l_output, self._output_vs])
else:
LayersPowered.__init__(self, [l_output])
def __init__(
self,
name,
env_spec,
hidden_dim=32,
feature_network=None,
prob_network=None,
state_include_action=True,
hidden_nonlinearity=tf.tanh,
forget_bias=1.0,
use_peepholes=False):
"""
: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:
"""
with tf.variable_scope(name):
assert isinstance(env_spec.action_space, Discrete)
Serializable.quick_init(self, locals())
super(CategoricalLSTMPolicy, 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 = L.OpLayer(
l_flat_feature,
extras=[l_input],
name="reshape_feature",
op=lambda flat_feature, input: tf.reshape(
flat_feature,
tf.pack([tf.shape(input)[0], tf.shape(input)[1], feature_dim])
),
shape_op=lambda _, input_shape: (input_shape[0], input_shape[1], feature_dim)
)
if prob_network is None:
prob_network = LSTMNetwork(
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=tf.nn.softmax,
forget_bias=forget_bias,
use_peepholes=use_peepholes,
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 = tf.placeholder(dtype=tf.float32, shape=(None, input_dim), name="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 = tensor_utils.compile_function(
[
flat_input_var,
prob_network.step_prev_hidden_layer.input_var,
prob_network.step_prev_cell_layer.input_var
],
L.get_output([
prob_network.step_output_layer,
prob_network.step_hidden_layer,
prob_network.step_cell_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_actions = None
self.prev_hiddens = None
self.prev_cells = 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)
LayersPowered.__init__(self, out_layers)
def __init__(
self,
name,
input_shape,
output_dim,
mean_network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
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,
subsample_factor=1.0
):
"""
: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())
with tf.variable_scope(name):
if optimizer is None:
if use_trust_region:
optimizer = PenaltyLbfgsOptimizer("optimizer")
else:
optimizer = LbfgsOptimizer("optimizer")
self._optimizer = optimizer
self._subsample_factor = subsample_factor
if mean_network is None:
mean_network = MLP(
name="mean_network",
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(
name="log_std_network",
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 = L.ParamLayer(
mean_network.input_layer,
num_units=output_dim,
param=tf.constant_initializer(np.log(init_std)),
name="output_log_std",
trainable=learn_std,
)
LayersPowered.__init__(self, [l_mean, l_log_std])
xs_var = mean_network.input_layer.input_var
ys_var = tf.placeholder(dtype=tf.float32, name="ys", shape=(None, output_dim))
old_means_var = tf.placeholder(dtype=tf.float32, name="ys", shape=(None, output_dim))
old_log_stds_var = tf.placeholder(dtype=tf.float32, name="old_log_stds", shape=(None, output_dim))
x_mean_var = tf.Variable(
np.zeros((1,) + input_shape, dtype=np.float32),
name="x_mean",
)
x_std_var = tf.Variable(
np.ones((1,) + input_shape, dtype=np.float32),
name="x_std",
)
y_mean_var = tf.Variable(
np.zeros((1, output_dim), dtype=np.float32),
name="y_mean",
)
y_std_var = tf.Variable(
np.ones((1, output_dim), dtype=np.float32),
name="y_std",
)
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 + tf.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 - tf.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 = tf.reduce_mean(dist.kl_sym(
dict(mean=normalized_old_means_var, log_std=normalized_old_log_stds_var),
normalized_dist_info_vars,
))
loss = - tf.reduce_mean(dist.log_likelihood_sym(normalized_ys_var, normalized_dist_info_vars))
self._f_predict = tensor_utils.compile_function([xs_var], means_var)
self._f_pdists = tensor_utils.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,
name,
env_spec,
hidden_dim=32,
feature_network=None,
state_include_action=True,
hidden_nonlinearity=tf.tanh,
learn_std=True,
init_std=1.0,
output_nonlinearity=None,
):
"""
: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:
"""
with tf.variable_scope(name):
Serializable.quick_init(self, locals())
super(GaussianGRUPolicy, 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 = L.OpLayer(
l_flat_feature,
extras=[l_input],
name="reshape_feature",
op=lambda flat_feature, input: tf.reshape(
flat_feature,
tf.pack([tf.shape(input)[0], tf.shape(input)[1], feature_dim])
),
shape_op=lambda _, input_shape: (input_shape[0], input_shape[1], feature_dim)
)
mean_network = GRUNetwork(
input_shape=(feature_dim,),
input_layer=l_feature,
output_dim=action_dim,
hidden_dim=hidden_dim,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
name="mean_network"
)
l_log_std = L.ParamLayer(
mean_network.input_layer,
num_units=action_dim,
param=tf.constant_initializer(np.log(init_std)),
name="output_log_std",
trainable=learn_std,
)
l_step_log_std = L.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.feature_network = feature_network
self.l_input = l_input
self.state_include_action = state_include_action
flat_input_var = tf.placeholder(dtype=tf.float32, shape=(None, input_dim), name="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_mean_std = tensor_utils.compile_function(
[
flat_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,
], {mean_network.step_input_layer: feature_var})
)
self.l_log_std = l_log_std
self.input_dim = input_dim
self.action_dim = action_dim
self.hidden_dim = hidden_dim
self.prev_actions = None
self.prev_hiddens = None
self.dist = RecurrentDiagonalGaussian(action_dim)
out_layers = [mean_network.output_layer, l_log_std, l_step_log_std]
if feature_network is not None:
out_layers.append(feature_network.output_layer)
LayersPowered.__init__(self, out_layers)
def __init__(
self,
name,
env_spec,
num_ensembles=5,
num_models_per_ensemble=3,
hidden_sizes=(512, 512),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=None,
batch_size=500,
step_size=0.001,
weight_normalization=False,
normalize_input=True,
):
Serializable.quick_init(self, locals())
self.normalization = None
self.normalize_input = normalize_input
self.batch_size = batch_size
self.step_size = step_size
self.num_ensembles = num_ensembles
self.num_models_per_ensemble = num_models_per_ensemble
self.num_models = num_ensembles * num_models_per_ensemble
# determine dimensionality of state and action space
self.obs_space_dims = obs_space_dims = env_spec.observation_space.shape[
0]
self.action_space_dims = action_space_dims = env_spec.action_space.shape[
0]
# set model - ensemble assignment
self.model_ensemble_assignment = [
list(
range(i * self.num_models_per_ensemble,
(i + 1) * self.num_models_per_ensemble))
for i in range(self.num_ensembles)
]
with tf.variable_scope(name):
# placeholders
self.obs_ph = tf.placeholder(tf.float32,
shape=(None, obs_space_dims))
self.act_ph = tf.placeholder(tf.float32,
shape=(None, action_space_dims))
self.delta_ph = tf.placeholder(tf.float32,
shape=(None, obs_space_dims))
# concatenate action and observation --> NN input
self.nn_input = tf.concat([self.obs_ph, self.act_ph], axis=1)
# create MLP
mlps = []
delta_preds = []
self.obs_next_pred = []
for i in range(self.num_models):
with tf.variable_scope('model_{}'.format(i)):
mlp = MLP(name,
obs_space_dims,
hidden_sizes,
hidden_nonlinearity,
output_nonlinearity,
input_var=self.nn_input,
input_shape=(obs_space_dims +
action_space_dims, ),
weight_normalization=weight_normalization)
mlps.append(mlp)
delta_preds.append(mlp.output)
self.delta_pred = tf.stack(
delta_preds, axis=2) # shape: (batch_size, ndim_obs, n_models)
# define loss and train_op
self.loss = tf.reduce_mean(
(self.delta_ph[:, :, None] - self.delta_pred)**2)
self.optimizer = tf.train.AdamOptimizer(self.step_size)
self.train_op = self.optimizer.minimize(self.loss)
# tensor_utils
self.f_delta_pred = tensor_utils.compile_function(
[self.obs_ph, self.act_ph], self.delta_pred)
LayersPowered.__init__(self, [mlp.output_layer for mlp in mlps])
def __init__(self,
name,
input_shape,
output_dim,
conv_filters,
conv_filter_sizes,
conv_strides,
conv_pads,
hidden_sizes,
hidden_nonlinearity,
output_nonlinearity,
hidden_W_init=L.XavierUniformInitializer(),
hidden_b_init=tf.zeros_initializer(),
output_W_init=L.XavierUniformInitializer(),
output_b_init=tf.zeros_initializer(),
input_var=None,
input_layer=None,
batch_normalization=False,
weight_normalization=False):
Serializable.quick_init(self, locals())
"""
A network composed of several convolution layers followed by some fc layers.
input_shape: (width,height,channel)
HOWEVER, network inputs are assumed flattened. This network will first unflatten the inputs and then apply the standard convolutions and so on.
conv_filters: a list of numbers of convolution kernel
conv_filter_sizes: a list of sizes (int) of the convolution kernels
conv_strides: a list of strides (int) of the conv kernels
conv_pads: a list of pad formats (either 'SAME' or 'VALID')
hidden_nonlinearity: a nonlinearity from tf.nn, shared by all conv and fc layers
hidden_sizes: a list of numbers of hidden units for all fc layers
"""
with tf.variable_scope(name):
if input_layer is not None:
l_in = input_layer
l_hid = l_in
elif len(input_shape) == 3:
l_in = L.InputLayer(shape=(None, np.prod(input_shape)),
input_var=input_var,
name="input")
l_hid = L.reshape(l_in, ([0], ) + input_shape,
name="reshape_input")
elif len(input_shape) == 2:
l_in = L.InputLayer(shape=(None, np.prod(input_shape)),
input_var=input_var,
name="input")
input_shape = (1, ) + input_shape
l_hid = L.reshape(l_in, ([0], ) + input_shape,
name="reshape_input")
else:
l_in = L.InputLayer(shape=(None, ) + input_shape,
input_var=input_var,
name="input")
l_hid = l_in
if batch_normalization:
l_hid = L.batch_norm(l_hid)
for idx, conv_filter, filter_size, stride, pad in zip(
range(len(conv_filters)),
conv_filters,
conv_filter_sizes,
conv_strides,
conv_pads,
):
l_hid = L.Conv2DLayer(
l_hid,
num_filters=conv_filter,
filter_size=filter_size,
stride=(stride, stride),
pad=pad,
nonlinearity=hidden_nonlinearity,
name="conv_hidden_%d" % idx,
weight_normalization=weight_normalization,
)
if batch_normalization:
l_hid = L.batch_norm(l_hid)
if output_nonlinearity == L.spatial_expected_softmax:
assert len(hidden_sizes) == 0
assert output_dim == conv_filters[-1] * 2
l_hid.nonlinearity = tf.identity
l_out = L.SpatialExpectedSoftmaxLayer(l_hid)
else:
l_hid = L.flatten(l_hid, name="conv_flatten")
for idx, hidden_size in enumerate(hidden_sizes):
l_hid = L.DenseLayer(
l_hid,
num_units=hidden_size,
nonlinearity=hidden_nonlinearity,
name="hidden_%d" % idx,
W=hidden_W_init,
b=hidden_b_init,
weight_normalization=weight_normalization,
)
if batch_normalization:
l_hid = L.batch_norm(l_hid)
l_out = L.DenseLayer(
l_hid,
num_units=output_dim,
nonlinearity=output_nonlinearity,
name="output",
W=output_W_init,
b=output_b_init,
weight_normalization=weight_normalization,
)
if batch_normalization:
l_out = L.batch_norm(l_out)
self._l_in = l_in
self._l_out = l_out
# self._input_var = l_in.input_var
LayersPowered.__init__(self, l_out)
def __getstate__(self):
state = LayersPowered.__getstate__(self)
state['normalization'] = self.normalization
state['model_ensemble_assignment'] = self.model_ensemble_assignment
return state
def __init__(self,
name,
output_dim,
hidden_sizes,
hidden_nonlinearity,
output_nonlinearity,
hidden_W_init=L.XavierUniformInitializer(),
hidden_b_init=tf.zeros_initializer(),
output_W_init=L.XavierUniformInitializer(),
output_b_init=tf.zeros_initializer(),
input_var=None,
input_layer=None,
input_shape=None,
batch_normalization=False,
weight_normalization=False,
latent_dim=0,
latent_shape=None,
obs_shape=None):
Serializable.quick_init(self, locals())
with tf.variable_scope(name):
if input_layer is None:
l_in = L.InputLayer(shape=(None, ) + input_shape,
input_var=input_var,
name="input")
else:
l_in = input_layer
# latent_in = L.InputLayer(shape=(None,) + latent_shape, input_var=l_in.input_var[:, -latent_dim:], name='latent')
# obs_in = L.InputLayer(shape=(None,) + obs_shape, input_var=l_in.input_var[:, :-latent_dim], name='obs_input')
latent_in = L.SliceLayer(l_in,
slice(-latent_dim, None, None),
axis=-1)
obs_in = L.SliceLayer(l_in, slice(0, -latent_dim, None), axis=-1)
self._layers = [obs_in]
l_hid = obs_in
if batch_normalization:
l_hid = L.batch_norm(l_hid)
for idx, hidden_size in enumerate(hidden_sizes):
l_hid = L.DenseLayer(l_hid,
num_units=hidden_size,
nonlinearity=hidden_nonlinearity,
name="hidden_%d" % idx,
W=hidden_W_init,
b=hidden_b_init,
weight_normalization=weight_normalization)
if batch_normalization:
l_hid = L.batch_norm(l_hid)
self._layers.append(l_hid)
l_latent_out = L.DenseLayer(
latent_in,
num_units=hidden_size,
nonlinearity=hidden_nonlinearity,
name="hidden_latent_0",
W=hidden_W_init,
b=hidden_b_init,
weight_normalization=weight_normalization)
if batch_normalization:
l_latent_out = L.batch_norm(l_latent_out)
self._layers.append(l_latent_out)
l_hid = L.ElemwiseSumLayer([l_hid, l_latent_out])
# l_hid = L.OpLayer(
# l_hid,
# op=lambda l_hid, l_latent:
# l_hid + l_latent,
# shape_op=lambda l_hid_shape, l_latent_shape:
# l_hid_shape,
# extras=[l_latent_out],
# name='sum_obs_latent')
l_out = L.DenseLayer(l_hid,
num_units=output_dim,
nonlinearity=output_nonlinearity,
name="output",
W=output_W_init,
b=output_b_init,
weight_normalization=weight_normalization)
if batch_normalization:
l_out = L.batch_norm(l_out)
self._layers.append(l_out)
self._l_in = l_in
self._l_out = l_out
# self._input_var = l_in.input_var
self._output = L.get_output(l_out)
LayersPowered.__init__(self, l_out)
def __setstate__(self, state):
LayersPowered.__setstate__(self, state)
self.normalization = state['normalization']
self.model_ensemble_assignment = state['model_ensemble_assignment']
def __init__(self,
name,
env_spec,
hidden_dim=32,
feature_network=None,
prob_network=None,
state_include_action=True,
hidden_nonlinearity=tf.tanh,
forget_bias=1.0,
use_peepholes=False,
lstm_layer_cls=L.LSTMLayer):
"""
: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:
"""
with tf.variable_scope(name):
assert isinstance(env_spec.action_space, Discrete)
Serializable.quick_init(self, locals())
super(CategoricalLSTMPolicy, 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 = L.OpLayer(
l_flat_feature,
extras=[l_input],
name="reshape_feature",
op=lambda flat_feature, input: tf.reshape(
flat_feature,
tf.stack([
tf.shape(input)[0],
tf.shape(input)[1], feature_dim
])),
shape_op=lambda _, input_shape:
(input_shape[0], input_shape[1], feature_dim))
if prob_network is None:
prob_network = LSTMNetwork(
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=tf.nn.softmax,
forget_bias=forget_bias,
use_peepholes=use_peepholes,
lstm_layer_cls=lstm_layer_cls,
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 = tf.placeholder(dtype=tf.float32,
shape=(None, input_dim),
name="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 = tensor_utils.compile_function(
[
flat_input_var,
#prob_network.step_prev_hidden_layer.input_var,
#prob_network.step_prev_cell_layer.input_var
prob_network.step_prev_state_layer.input_var,
],
L.get_output([
prob_network.step_output_layer,
prob_network.step_hidden_layer,
prob_network.step_cell_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_actions = None
self.prev_hiddens = None
self.prev_cells = 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)
LayersPowered.__init__(self, out_layers)
def __init__(
self,
name,
env_spec,
qmdp_param,
feature_network=None,
state_include_action=True,
):
"""
: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:
"""
with tf.variable_scope(name):
assert isinstance(env_spec.action_space, Discrete)
Serializable.quick_init(self, locals())
super(QMDPPolicy, self).__init__(env_spec)
self.qmdp_param = qmdp_param
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 = L.OpLayer(
l_flat_feature,
extras=[l_input],
name="reshape_feature",
op=lambda flat_feature, input: tf.reshape(
flat_feature,
tf.stack([
tf.shape(input)[0],
tf.shape(input)[1], feature_dim
])),
shape_op=lambda _, input_shape:
(input_shape[0], input_shape[1], feature_dim))
prob_network = QMDPNetwork(input_shape=(feature_dim, ),
input_layer=l_feature,
output_dim=env_spec.action_space.n,
qmdp_param=qmdp_param,
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 = tf.placeholder(dtype=tf.float32,
shape=(None, input_dim),
name="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 = tensor_utils.compile_function(
[
flat_input_var,
# prob_network.step_prev_hidden_layer.input_var
prob_network.step_prev_state_layer.input_var
],
L.get_output([
prob_network.step_output_layer,
prob_network.step_hidden_layer
], {prob_network.step_input_layer: feature_var}))
self.debug = tensor_utils.compile_function(
[
flat_input_var,
# prob_network.step_prev_hidden_layer.input_var
prob_network.step_prev_state_layer.input_var
],
# [self.prob_network._l_output_flat.plannernet.printQ]
[
# self.prob_network._l_output_flat.plannernet.f_pi.fclayers.fclayers[0].w,
self.prob_network._l_output_flat.R0,
self.prob_network._l_gru.z_os
])
self.input_dim = input_dim
self.action_dim = action_dim
self.hidden_dim = qmdp_param['num_state']
self.prev_actions = None
self.prev_hiddens = 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)
LayersPowered.__init__(self, out_layers)
def __init__(self, name, input_shape, extra_input_shape, output_dim, hidden_sizes,
conv_filters, conv_filter_sizes, conv_strides, conv_pads,
extra_hidden_sizes=None,
hidden_W_init=L.XavierUniformInitializer(), hidden_b_init=tf.zeros_initializer,
output_W_init=L.XavierUniformInitializer(), output_b_init=tf.zeros_initializer,
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=None,
input_var=None, input_layer=None):
Serializable.quick_init(self, locals())
if extra_hidden_sizes is None:
extra_hidden_sizes = []
with tf.variable_scope(name):
input_flat_dim = np.prod(input_shape)
extra_input_flat_dim = np.prod(extra_input_shape)
total_input_flat_dim = input_flat_dim + extra_input_flat_dim
if input_layer is None:
l_in = L.InputLayer(shape=(None, total_input_flat_dim), input_var=input_var, name="input")
else:
l_in = input_layer
l_conv_in = L.reshape(
L.SliceLayer(
l_in,
indices=slice(input_flat_dim),
name="conv_slice"
),
([0],) + input_shape,
name="conv_reshaped"
)
l_extra_in = L.reshape(
L.SliceLayer(
l_in,
indices=slice(input_flat_dim, None),
name="extra_slice"
),
([0],) + extra_input_shape,
name="extra_reshaped"
)
l_conv_hid = l_conv_in
for idx, conv_filter, filter_size, stride, pad in zip(
range(len(conv_filters)),
conv_filters,
conv_filter_sizes,
conv_strides,
conv_pads,
):
l_conv_hid = L.Conv2DLayer(
l_conv_hid,
num_filters=conv_filter,
filter_size=filter_size,
stride=(stride, stride),
pad=pad,
nonlinearity=hidden_nonlinearity,
name="conv_hidden_%d" % idx,
)
l_extra_hid = l_extra_in
for idx, hidden_size in enumerate(extra_hidden_sizes):
l_extra_hid = L.DenseLayer(
l_extra_hid,
num_units=hidden_size,
nonlinearity=hidden_nonlinearity,
name="extra_hidden_%d" % idx,
W=hidden_W_init,
b=hidden_b_init,
)
l_joint_hid = L.concat(
[L.flatten(l_conv_hid, name="conv_hidden_flat"), l_extra_hid],
name="joint_hidden"
)
for idx, hidden_size in enumerate(hidden_sizes):
l_joint_hid = L.DenseLayer(
l_joint_hid,
num_units=hidden_size,
nonlinearity=hidden_nonlinearity,
name="joint_hidden_%d" % idx,
W=hidden_W_init,
b=hidden_b_init,
)
l_out = L.DenseLayer(
l_joint_hid,
num_units=output_dim,
nonlinearity=output_nonlinearity,
name="output",
W=output_W_init,
b=output_b_init,
)
self._l_in = l_in
self._l_out = l_out
LayersPowered.__init__(self, [l_out], input_layers=[l_in])
def __init__(
self,
name,
input_shape,
output_dim,
network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
optimizer=None,
normalize_inputs=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.
"""
Serializable.quick_init(self, locals())
with tf.variable_scope(name):
if optimizer is None:
optimizer = LbfgsOptimizer(name="optimizer")
self.output_dim = output_dim
self.optimizer = optimizer
if network is None:
network = MLP(input_shape=input_shape,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
name="network")
l_out = network.output_layer
LayersPowered.__init__(self, [l_out])
xs_var = network.input_layer.input_var
ys_var = tf.placeholder(dtype=tf.float32,
shape=[None, output_dim],
name="ys")
x_mean_var = tf.get_variable(name="x_mean",
shape=(1, ) + input_shape,
initializer=tf.constant_initializer(
0., dtype=tf.float32))
x_std_var = tf.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
fit_ys_var = L.get_output(l_out,
{network.input_layer: normalized_xs_var})
loss = -tf.reduce_mean(tf.square(fit_ys_var - ys_var))
self.f_predict = tensor_utils.compile_function([xs_var],
fit_ys_var)
optimizer_args = dict(
loss=loss,
target=self,
network_outputs=[fit_ys_var],
)
optimizer_args["inputs"] = [xs_var, ys_var]
self.optimizer.update_opt(**optimizer_args)
self.name = name
self.l_out = l_out
self.normalize_inputs = normalize_inputs
self.x_mean_var = x_mean_var
self.x_std_var = x_std_var
def get_param_values(self, **tags):
return LayersPowered.get_param_values(self, **tags)
def __init__(
self,
name,
input_shape,
output_dim,
prob_network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
optimizer=None,
tr_optimizer=None,
use_trust_region=True,
step_size=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 step_size: KL divergence constraint for each iteration
"""
Serializable.quick_init(self, locals())
with tf.variable_scope(name):
if optimizer is None:
optimizer = LbfgsOptimizer(name="optimizer")
if tr_optimizer is None:
tr_optimizer = ConjugateGradientOptimizer()
self.input_dim = input_shape[0]
self.observation_space = Discrete(self.input_dim)
self.action_space = Discrete(output_dim)
self.output_dim = output_dim
self.optimizer = optimizer
self.tr_optimizer = tr_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=tf.nn.softmax,
name="prob_network"
)
l_prob = prob_network.output_layer
LayersPowered.__init__(self, [l_prob])
xs_var = prob_network.input_layer.input_var
ys_var = tf.placeholder(dtype=tf.float32, shape=[None, output_dim], name="ys")
old_prob_var = tf.placeholder(dtype=tf.float32, shape=[None, output_dim], name="old_prob")
x_mean_var = tf.get_variable(
name="x_mean",
shape=(1,) + input_shape,
initializer=tf.constant_initializer(0., dtype=tf.float32)
)
x_std_var = tf.get_variable(
name="x_std",
shape=(1,) + input_shape,
initializer=tf.constant_initializer(1., dtype=tf.float32)
)
self.x_mean_var = x_mean_var
self.x_std_var = x_std_var
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 = tf.reduce_mean(dist.kl_sym(old_info_vars, info_vars))
loss = - tf.reduce_mean(dist.log_likelihood_sym(ys_var, info_vars))
predicted = tensor_utils.to_onehot_sym(tf.argmax(prob_var, axis=1), 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, step_size)
)
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
def __init__(
self,
input_shape,
output_dim,
name,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.relu,
optimizer=None,
tr_optimizer=None,
use_trust_region=True,
step_size=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 step_size: KL divergence constraint for each iteration
"""
Serializable.quick_init(self, locals())
with tf.variable_scope(name):
if optimizer is None:
optimizer = LbfgsOptimizer(name="optimizer")
if tr_optimizer is None:
tr_optimizer = ConjugateGradientOptimizer()
self.output_dim = output_dim
self.optimizer = optimizer
self.tr_optimizer = tr_optimizer
p_network = MLP(input_shape=input_shape,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=tf.nn.sigmoid,
name="p_network")
l_p = p_network.output_layer
LayersPowered.__init__(self, [l_p])
xs_var = p_network.input_layer.input_var
ys_var = tf.placeholder(dtype=tf.float32,
shape=(None, output_dim),
name="ys")
old_p_var = tf.placeholder(dtype=tf.float32,
shape=(None, output_dim),
name="old_p")
x_mean_var = tf.get_variable(name="x_mean",
initializer=tf.zeros_initializer,
shape=(1, ) + input_shape)
x_std_var = tf.get_variable(name="x_std",
initializer=tf.ones_initializer,
shape=(1, ) + input_shape)
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
p_var = L.get_output(l_p,
{p_network.input_layer: normalized_xs_var})
old_info_vars = dict(p=old_p_var)
info_vars = dict(p=p_var)
dist = self._dist = Bernoulli(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 = p_var >= 0.5
self.f_predict = tensor_utils.compile_function([xs_var], predicted)
self.f_p = tensor_utils.compile_function([xs_var], p_var)
self.l_p = l_p
self.optimizer.update_opt(loss=loss,
target=self,
network_outputs=[p_var],
inputs=[xs_var, ys_var])
self.tr_optimizer.update_opt(loss=loss,
target=self,
network_outputs=[p_var],
inputs=[xs_var, ys_var, old_p_var],
leq_constraint=(mean_kl, step_size))
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
def __init__(
self,
env_spec,
name='qnet',
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.relu,
action_merge_layer=-2,
output_nonlinearity=None,
hidden_W_init=L.XavierUniformInitializer(),
hidden_b_init=tf.zeros_initializer,
output_W_init=L.XavierUniformInitializer(),
output_b_init=tf.zeros_initializer,
bn=False):
Serializable.quick_init(self, locals())
with tf.variable_scope(name):
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 = L.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()
output_var = tf.reshape(L.get_output(l_output, deterministic=True),(-1,))
self._f_qval = tensor_utils.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
LayersPowered.__init__(self, [l_output])
def __init__(
self,
name,
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=tf.nn.tanh,
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
mean_network=None,
std_network=None,
std_parametrization='exp'
):
"""
:param env_spec:
:param hidden_sizes: list of sizes 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: list of sizes 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
:param std_parametrization: how the std should be parametrized. There are a few options:
- exp: the logarithm of the std will be stored, and applied a exponential transformation
- softplus: the std will be computed as log(1+exp(x))
:return:
"""
Serializable.quick_init(self, locals())
assert isinstance(env_spec.action_space, Box)
with tf.variable_scope(name):
obs_dim = env_spec.observation_space.flat_dim
action_dim = env_spec.action_space.flat_dim
# create network
if mean_network is None:
mean_network = MLP(
name="mean_network",
input_shape=(obs_dim,),
output_dim=action_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_std_param = std_network.output_layer
else:
if adaptive_std:
std_network = MLP(
name="std_network",
input_shape=(obs_dim,),
input_layer=mean_network.input_layer,
output_dim=action_dim,
hidden_sizes=std_hidden_sizes,
hidden_nonlinearity=std_hidden_nonlinearity,
output_nonlinearity=None,
)
l_std_param = std_network.output_layer
else:
if std_parametrization == 'exp':
init_std_param = np.log(init_std)
elif std_parametrization == 'softplus':
init_std_param = np.log(np.exp(init_std) - 1)
else:
raise NotImplementedError
l_std_param = L.ParamLayer(
mean_network.input_layer,
num_units=action_dim,
param=tf.constant_initializer(init_std_param),
name="output_std_param",
trainable=learn_std,
)
self.std_parametrization = std_parametrization
if std_parametrization == 'exp':
min_std_param = np.log(min_std)
elif std_parametrization == 'softplus':
min_std_param = np.log(np.exp(min_std) - 1)
else:
raise NotImplementedError
self.min_std_param = min_std_param
# mean_var, log_std_var = L.get_output([l_mean, l_std_param])
#
# if self.min_std_param is not None:
# log_std_var = tf.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_std_param = l_std_param
self._dist = DiagonalGaussian(action_dim)
LayersPowered.__init__(self, [l_mean, l_std_param])
super(GaussianMLPPolicy, self).__init__(env_spec)
dist_info_sym = self.dist_info_sym(mean_network.input_layer.input_var, dict())
mean_var = dist_info_sym["mean"]
log_std_var = dist_info_sym["log_std"]
self._f_dist = tensor_utils.compile_function(
inputs=[obs_var],
outputs=[mean_var, log_std_var],
)
def __init__(
self,
name,
env_spec,
hidden_dims=(32,),
feature_network=None,
state_include_action=True,
hidden_nonlinearity=tf.tanh):
"""
:param env_spec: A spec for the env.
:param hidden_dims: dimension of hidden layers
:param hidden_nonlinearity: nonlinearity used for each hidden layer
:return:
"""
with tf.variable_scope(name):
assert isinstance(env_spec.action_space, Discrete)
Serializable.quick_init(self, locals())
super(RecurrentCategoricalPolicy, 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 = L.OpLayer(
l_flat_feature,
extras=[l_input],
name="reshape_feature",
op=lambda flat_feature, input: tf.reshape(
flat_feature,
tf.pack([tf.shape(input)[0], tf.shape(input)[1], feature_dim])
),
shape_op=lambda _, input_shape: (input_shape[0], input_shape[1], feature_dim)
)
prob_network = DeepGRUNetwork(
input_shape=(feature_dim,),
input_layer=l_feature,
output_dim=env_spec.action_space.n,
hidden_dims=hidden_dims,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=tf.nn.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 = tf.placeholder(tf.float32, shape=(None, input_dim), name="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})
# Build the step feedforward function.
inputs = [flat_input_var] \
+ [prev_hidden.input_var for prev_hidden
in prob_network.step_prev_hidden_layers]
outputs = [prob_network.step_output_layer] \
+ prob_network.step_hidden_layers
outputs = L.get_output(outputs, {prob_network.step_input_layer: feature_var})
self.f_step_prob = tensor_utils.compile_function(
inputs, outputs)
# Function to fetch hidden init values
self.f_hid_inits = tensor_utils.compile_function(
[], prob_network.hid_inits)
self.input_dim = input_dim
self.action_dim = action_dim
self.hidden_dims = hidden_dims
self.prev_actions = None
self.prev_hiddens = 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)
LayersPowered.__init__(self, out_layers)