def __init__(
self,
output_dim,
hidden_sizes,
hidden_nonlinearity,
output_nonlinearity,
name=None,
hidden_w_init=ly.XavierUniformInitializer(),
hidden_b_init=tf.zeros_initializer(),
output_w_init=ly.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, "MLP"):
if input_layer is None:
l_in = ly.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 = ly.batch_norm(l_hid)
for idx, hidden_size in enumerate(hidden_sizes):
l_hid = ly.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 = ly.batch_norm(l_hid)
self._layers.append(l_hid)
l_out = ly.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 = ly.batch_norm(l_out)
self._layers.append(l_out)
self._l_in = l_in
self._l_out = l_out
LayersPowered.__init__(self, l_out)
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, reuse=None, custom_getter=None, trainable=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.
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_in = layers.InputLayer(shape=(None, self._obs_dim), name="obs")
l_hidden = l_in
for idx, hidden_size in enumerate(self._hidden_sizes):
if self._batch_norm:
l_hidden = batch_norm(l_hidden)
l_hidden = layers.DenseLayer(
l_hidden,
hidden_size,
nonlinearity=self._hidden_nonlinearity,
trainable=trainable,
name="hidden_%d" % idx)
l_output = layers.DenseLayer(
l_hidden,
self._action_dim,
nonlinearity=self._output_nonlinearity,
trainable=trainable,
name="output")
with tf.name_scope(self._policy_network_name):
action = layers.get_output(l_output)
scaled_action = tf.multiply(action,
self._action_bound,
name="scaled_action")
self._f_prob_online = tensor_utils.compile_function(
inputs=[l_in.input_var], outputs=scaled_action)
self._output_layer = l_output
self._obs_layer = l_in
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_in = layers.InputLayer(shape=(None, self._obs_dim), name='obs')
l_hidden = l_in
for idx, hidden_size in enumerate(self._hidden_sizes):
if self._batch_norm:
l_hidden = batch_norm(l_hidden)
l_hidden = layers.DenseLayer(
l_hidden,
hidden_size,
nonlinearity=self._hidden_nonlinearity,
trainable=trainable,
name='hidden_%d' % idx)
l_output = layers.DenseLayer(
l_hidden,
self._action_dim,
nonlinearity=self._output_nonlinearity,
trainable=trainable,
name='output')
with tf.name_scope(self._policy_network_name):
action = layers.get_output(l_output)
scaled_action = tf.multiply(action,
self._action_bound,
name='scaled_action')
f_prob_online = tensor_utils.compile_function(inputs=[l_in.input_var],
outputs=scaled_action)
output_layer = l_output
obs_layer = l_in
return f_prob_online, output_layer, obs_layer
def __init__(self,
input_shape,
output_dim,
conv_filters,
conv_filter_sizes,
conv_strides,
conv_pads,
hidden_sizes,
hidden_nonlinearity,
output_nonlinearity,
name=None,
hidden_w_init=ly.XavierUniformInitializer(),
hidden_b_init=tf.zeros_initializer(),
output_w_init=ly.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, 'ConvNetwork'):
if input_layer is not None:
l_in = input_layer
l_hid = l_in
elif len(input_shape) == 3:
l_in = ly.InputLayer(shape=(None, np.prod(input_shape)),
input_var=input_var,
name='input')
l_hid = ly.reshape(l_in, ([0], ) + input_shape,
name='reshape_input')
elif len(input_shape) == 2:
l_in = ly.InputLayer(shape=(None, np.prod(input_shape)),
input_var=input_var,
name='input')
input_shape = (1, ) + input_shape
l_hid = ly.reshape(l_in, ([0], ) + input_shape,
name='reshape_input')
else:
l_in = ly.InputLayer(shape=(None, ) + input_shape,
input_var=input_var,
name='input')
l_hid = l_in
if batch_normalization:
l_hid = ly.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 = ly.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 = ly.batch_norm(l_hid)
if output_nonlinearity == ly.spatial_expected_softmax:
assert not hidden_sizes
assert output_dim == conv_filters[-1] * 2
l_hid.nonlinearity = tf.identity
l_out = ly.SpatialExpectedSoftmaxLayer(l_hid)
else:
l_hid = ly.flatten(l_hid, name='conv_flatten')
for idx, hidden_size in enumerate(hidden_sizes):
l_hid = ly.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 = ly.batch_norm(l_hid)
l_out = ly.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 = ly.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 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
def build_net(self, trainable=True, name=None):
"""
Set up policy 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")
for idx, hidden_size in enumerate(self._hidden_sizes):
l_hid = layers.DenseLayer(
l_hid,
num_units=hidden_size,
nonlinearity=self._hidden_nonlinearity,
name="hidden_%d" % idx,
weight_normalization=self._weight_normalization,
trainable=trainable,
)
if self._batch_norm:
l_hid = layers.batch_norm(l_hid)
l_output = layers.DenseLayer(
l_hid,
num_units=self._action_dim,
nonlinearity=self._output_nonlinearity,
name="output",
weight_normalization=self._weight_normalization,
trainable=trainable,
)
with tf.name_scope(self._policy_network_name):
action = layers.get_output(l_output)
# scaled_action = tf.multiply(
# action, self._action_bound, name="scaled_action")
f_prob_online = tensor_utils.compile_function(
inputs=[l_in.input_var], outputs=action)
output_layer = l_output
obs_layer = l_in
return f_prob_online, output_layer, obs_layer
