def _build_net(self, reuse=None, custom_getter=None, trainable=None):
"""
Set up q network based on class attributes. This function uses layers
defined in rllab.tf.
Args:
reuse: A bool indicates whether reuse variables in the same scope.
custom_getter: A customized getter object used to get variables.
trainable: A bool indicates whether variables are trainable.
"""
with tf.variable_scope(self.name,
reuse=reuse,
custom_getter=custom_getter):
l_obs = L.InputLayer(shape=(None, self._obs_dim), name="obs")
l_action = L.InputLayer(shape=(None, self._action_dim),
name="actions")
n_layers = len(self._hidden_sizes) + 1
if n_layers > 1:
action_merge_layer = \
(self._action_merge_layer % n_layers + n_layers) % n_layers
else:
action_merge_layer = 1
l_hidden = l_obs
for idx, size in enumerate(self._hidden_sizes):
if self._batch_norm:
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=self._hidden_nonlinearity,
trainable=trainable,
name="hidden_%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=self._output_nonlinearity,
trainable=trainable,
name="output")
output_var = L.get_output(l_output)
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
LayersPowered.__init__(self, [l_output])
def build_net(self, trainable=True, name=None):
"""
Set up q network based on class attributes. This function uses layers
defined in garage.tf.
Args:
reuse: A bool indicates whether reuse variables in the same scope.
trainable: A bool indicates whether variables are trainable.
"""
with tf.variable_scope(name):
l_obs = L.InputLayer(shape=(None, self._obs_dim), name="obs")
l_action = L.InputLayer(shape=(None, self._action_dim),
name="actions")
n_layers = len(self._hidden_sizes) + 1
if n_layers > 1:
action_merge_layer = \
(self._action_merge_layer % n_layers + n_layers) % n_layers
else:
action_merge_layer = 1
l_hidden = l_obs
for idx, size in enumerate(self._hidden_sizes):
if self._batch_norm:
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=self._hidden_nonlinearity,
trainable=trainable,
name="hidden_%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=self._output_nonlinearity,
trainable=trainable,
name="output")
output_var = L.get_output(l_output)
f_qval = tensor_utils.compile_function(
[l_obs.input_var, l_action.input_var], output_var)
output_layer = l_output
obs_layer = l_obs
action_layer = l_action
return f_qval, output_layer, obs_layer, action_layer
def build_net(self, trainable=True, name=None):
"""
Set up q network based on class attributes. This function uses layers
defined in garage.tf.
Args:
reuse: A bool indicates whether reuse variables in the same scope.
trainable: A bool indicates whether variables are trainable.
"""
input_shape = self._env_spec.observation_space.shape
assert len(input_shape) in [2, 3]
if len(input_shape) == 2:
input_shape = (1, ) + input_shape
with tf.variable_scope(name):
l_in = layers.InputLayer(shape=(None, self._obs_dim), name="obs")
l_hid = layers.reshape(l_in, ([0], ) + input_shape,
name="reshape_input")
if self._batch_norm:
l_hid = layers.batch_norm(l_hid)
for idx, conv_filter, filter_size, stride, pad in zip(
range(len(self._conv_filters)),
self._conv_filters,
self._conv_filter_sizes,
self._conv_strides,
self._conv_pads,
):
l_hid = layers.Conv2DLayer(
l_hid,
num_filters=conv_filter,
filter_size=filter_size,
stride=(stride, stride),
pad=pad,
nonlinearity=self._hidden_nonlinearity,
name="conv_hidden_%d" % idx,
weight_normalization=self._weight_normalization,
trainable=trainable,
)
if self._pooling:
l_hid = layers.Pool2DLayer(l_hid,
pool_size=self._pool_size)
if self._batch_norm:
l_hid = layers.batch_norm(l_hid)
l_hid = layers.flatten(l_hid, name="conv_flatten")
l_action = layers.InputLayer(shape=(None, self._action_dim),
name="actions")
n_layers = len(self._hidden_sizes) + 1
if n_layers > 1:
action_merge_layer = \
(self._action_merge_layer % n_layers + n_layers) % n_layers
else:
action_merge_layer = 1
for idx, size in enumerate(self._hidden_sizes):
if self._batch_norm:
l_hid = batch_norm(l_hid)
if idx == action_merge_layer:
l_hid = layers.ConcatLayer([l_hid, l_action])
l_hid = layers.DenseLayer(
l_hid,
num_units=size,
nonlinearity=self._hidden_nonlinearity,
trainable=trainable,
name="hidden_%d" % (idx + 1))
if action_merge_layer == n_layers:
l_hid = layers.ConcatLayer([l_hid, l_action])
l_output = layers.DenseLayer(
l_hid,
num_units=1,
nonlinearity=self._output_nonlinearity,
trainable=trainable,
name="output")
output_var = layers.get_output(l_output)
f_qval = tensor_utils.compile_function(
[l_in.input_var, l_action.input_var], output_var)
output_layer = l_output
obs_layer = l_in
action_layer = l_action
return f_qval, output_layer, obs_layer, action_layer