def __init__(
self,
env_spec,
name="GaussianGRUPolicy",
hidden_dim=32,
feature_network=None,
state_include_action=True,
hidden_nonlinearity=tf.tanh,
gru_layer_cls=L.GRULayer,
learn_std=True,
init_std=1.0,
output_nonlinearity=None,
std_share_network=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:
"""
assert isinstance(env_spec.action_space, Box)
self._mean_network_name = "mean_network"
self._std_network_name = "std_network"
with tf.variable_scope(name, "GaussianGRUPolicy"):
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.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 std_share_network:
mean_network = GRUNetwork(
input_shape=(feature_dim, ),
input_layer=l_feature,
output_dim=2 * action_dim,
hidden_dim=hidden_dim,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
gru_layer_cls=gru_layer_cls,
name="gru_mean_network")
l_mean = L.SliceLayer(mean_network.output_layer,
slice(action_dim),
name="mean_slice")
l_step_mean = L.SliceLayer(mean_network.step_output_layer,
slice(action_dim),
name="step_mean_slice")
l_log_std = L.SliceLayer(mean_network.output_layer,
slice(action_dim, 2 * action_dim),
name="log_std_slice")
l_step_log_std = L.SliceLayer(mean_network.step_output_layer,
slice(action_dim,
2 * action_dim),
name="step_log_std_slice")
else:
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,
gru_layer_cls=gru_layer_cls,
name="gru_mean_network")
l_mean = mean_network.output_layer
l_step_mean = mean_network.step_output_layer
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})
with tf.name_scope(self._mean_network_name):
out_step_mean, out_step_hidden_mean = L.get_output(
[l_step_mean, mean_network.step_hidden_layer],
{mean_network.step_input_layer: feature_var})
out_step_mean = tf.identity(out_step_mean, "step_mean")
out_step_hidden_mean = tf.identity(out_step_hidden_mean,
"step_hidden_mean")
with tf.name_scope(self._std_network_name):
out_step_log_std = L.get_output(
l_step_log_std,
{mean_network.step_input_layer: feature_var})
out_step_log_std = tf.identity(out_step_log_std,
"step_log_std")
self.f_step_mean_std = tensor_utils.compile_function([
flat_input_var,
mean_network.step_prev_state_layer.input_var,
], [out_step_mean, out_step_log_std, out_step_hidden_mean])
self.l_mean = l_mean
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)
self.name = name
out_layers = [l_mean, 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,
env_spec,
name='GaussianLSTMPolicy',
hidden_dim=32,
hidden_nonlinearity=tf.tanh,
recurrent_nonlinearity=tf.nn.sigmoid,
recurrent_w_x_init=L.XavierUniformInitializer(),
recurrent_w_h_init=L.OrthogonalInitializer(),
output_nonlinearity=None,
output_w_init=L.XavierUniformInitializer(),
feature_network=None,
state_include_action=True,
learn_std=True,
init_std=1.0,
lstm_layer_cls=L.LSTMLayer,
use_peepholes=False,
std_share_network=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:
"""
assert isinstance(env_spec.action_space, akro.Box)
self._mean_network_name = 'mean_network'
self._std_network_name = 'std_network'
with tf.variable_scope(name, 'GaussianLSTMPolicy'):
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))
if std_share_network:
mean_network = LSTMNetwork(
input_shape=(feature_dim, ),
input_layer=l_feature,
output_dim=2 * action_dim,
hidden_dim=hidden_dim,
hidden_nonlinearity=hidden_nonlinearity,
recurrent_nonlinearity=recurrent_nonlinearity,
recurrent_w_x_init=recurrent_w_x_init,
recurrent_w_h_init=recurrent_w_h_init,
output_nonlinearity=output_nonlinearity,
output_w_init=output_w_init,
lstm_layer_cls=lstm_layer_cls,
name='lstm_mean_network',
use_peepholes=use_peepholes,
)
l_mean = L.SliceLayer(
mean_network.output_layer,
slice(action_dim),
name='mean_slice',
)
l_step_mean = L.SliceLayer(
mean_network.step_output_layer,
slice(action_dim),
name='step_mean_slice',
)
l_log_std = L.SliceLayer(
mean_network.output_layer,
slice(action_dim, 2 * action_dim),
name='log_std_slice',
)
l_step_log_std = L.SliceLayer(
mean_network.step_output_layer,
slice(action_dim, 2 * action_dim),
name='step_log_std_slice',
)
else:
mean_network = LSTMNetwork(
input_shape=(feature_dim, ),
input_layer=l_feature,
output_dim=action_dim,
hidden_dim=hidden_dim,
hidden_nonlinearity=hidden_nonlinearity,
recurrent_nonlinearity=recurrent_nonlinearity,
recurrent_w_x_init=recurrent_w_x_init,
recurrent_w_h_init=recurrent_w_h_init,
output_nonlinearity=output_nonlinearity,
output_w_init=output_w_init,
lstm_layer_cls=lstm_layer_cls,
name='lstm_mean_network',
use_peepholes=use_peepholes,
)
l_mean = mean_network.output_layer
l_step_mean = mean_network.step_output_layer
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
self.name = name
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})
with tf.name_scope(self._mean_network_name, values=[feature_var]):
(out_step_mean, out_step_hidden, out_mean_cell) = L.get_output(
[
l_step_mean, mean_network.step_hidden_layer,
mean_network.step_cell_layer
], {mean_network.step_input_layer: feature_var})
out_step_mean = tf.identity(out_step_mean, 'step_mean')
out_step_hidden = tf.identity(out_step_hidden, 'step_hidden')
out_mean_cell = tf.identity(out_mean_cell, 'mean_cell')
with tf.name_scope(self._std_network_name, values=[feature_var]):
out_step_log_std = L.get_output(
l_step_log_std,
{mean_network.step_input_layer: feature_var})
out_step_log_std = tf.identity(out_step_log_std,
'step_log_std')
self.f_step_mean_std = tensor_utils.compile_function([
flat_input_var,
mean_network.step_prev_state_layer.input_var,
], [
out_step_mean, out_step_log_std, out_step_hidden, out_mean_cell
])
self.l_mean = l_mean
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 = [l_mean, l_log_std]
if feature_network is not None:
out_layers.append(feature_network.output_layer)
LayersPowered.__init__(self, out_layers)
