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 __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 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,
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,
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 __init__(self,
input_shape,
extra_input_shape,
output_dim,
hidden_sizes,
conv_filters,
conv_filter_sizes,
conv_strides,
conv_pads,
name=None,
extra_hidden_sizes=None,
hidden_w_init=ly.XavierUniformInitializer(),
hidden_b_init=tf.zeros_initializer(),
output_w_init=ly.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, 'ConvMergeNetwork'):
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 = ly.InputLayer(shape=(None, total_input_flat_dim),
input_var=input_var,
name='input')
else:
l_in = input_layer
l_conv_in = ly.reshape(ly.SliceLayer(l_in,
indices=slice(input_flat_dim),
name='conv_slice'),
([0], ) + input_shape,
name='conv_reshaped')
l_extra_in = ly.reshape(ly.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 = ly.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 = ly.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 = ly.concat(
[ly.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 = ly.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 = ly.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,
input_shape,
output_dim,
hidden_dim,
name=None,
hidden_nonlinearity=tf.nn.relu,
output_w_init=ly.XavierUniformInitializer(),
recurrent_nonlinearity=tf.nn.sigmoid,
recurrent_w_x_init=ly.XavierUniformInitializer(),
recurrent_w_h_init=ly.OrthogonalInitializer(),
lstm_layer_cls=ly.LSTMLayer,
output_nonlinearity=None,
input_var=None,
input_layer=None,
forget_bias=1.0,
use_peepholes=False,
layer_args=None):
with tf.variable_scope(name, 'LSTMNetwork'):
if input_layer is None:
l_in = ly.InputLayer(shape=(None, None) + input_shape,
input_var=input_var,
name='input')
else:
l_in = input_layer
l_step_input = ly.InputLayer(shape=(None, ) + input_shape,
name='step_input')
# contains previous hidden and cell state
l_step_prev_state = ly.InputLayer(shape=(None, hidden_dim * 2),
name='step_prev_state')
if layer_args is None:
layer_args = dict()
l_lstm = lstm_layer_cls(l_in,
num_units=hidden_dim,
hidden_nonlinearity=hidden_nonlinearity,
gate_nonlinearity=recurrent_nonlinearity,
hidden_init_trainable=False,
name='lstm_layer',
forget_bias=forget_bias,
cell_init_trainable=False,
w_x_init=recurrent_w_x_init,
w_h_init=recurrent_w_h_init,
use_peepholes=use_peepholes,
**layer_args)
l_lstm_flat = ly.ReshapeLayer(l_lstm,
shape=(-1, hidden_dim),
name='lstm_flat')
l_output_flat = ly.DenseLayer(l_lstm_flat,
num_units=output_dim,
nonlinearity=output_nonlinearity,
w=output_w_init,
name='output_flat')
l_output = ly.OpLayer(
l_output_flat,
op=lambda flat_output, l_input: tf.reshape(
flat_output,
tf.stack(
(tf.shape(l_input)[0], tf.shape(l_input)[1], -1))),
shape_op=lambda flat_output_shape, l_input_shape:
(l_input_shape[0], l_input_shape[1], flat_output_shape[-1]),
extras=[l_in],
name='output')
l_step_state = l_lstm.get_step_layer(l_step_input,
l_step_prev_state,
name='step_state')
l_step_hidden = ly.SliceLayer(l_step_state,
indices=slice(hidden_dim),
name='step_hidden')
l_step_cell = ly.SliceLayer(l_step_state,
indices=slice(hidden_dim, None),
name='step_cell')
l_step_output = ly.DenseLayer(l_step_hidden,
num_units=output_dim,
nonlinearity=output_nonlinearity,
w=l_output_flat.w,
b=l_output_flat.b,
name='step_output')
self._l_in = l_in
self._hid_init_param = l_lstm.h0
self._cell_init_param = l_lstm.c0
self._l_lstm = l_lstm
self._l_out = l_output
self._l_step_input = l_step_input
self._l_step_prev_state = l_step_prev_state
self._l_step_hidden = l_step_hidden
self._l_step_cell = l_step_cell
self._l_step_state = l_step_state
self._l_step_output = l_step_output
self._hidden_dim = hidden_dim
def __init__(self,
input_shape,
output_dim,
hidden_dim,
name=None,
hidden_nonlinearity=tf.nn.relu,
output_w_init=ly.XavierUniformInitializer(),
recurrent_nonlinearity=tf.nn.sigmoid,
recurrent_w_x_init=ly.XavierUniformInitializer(),
recurrent_w_h_init=ly.OrthogonalInitializer(),
gru_layer_cls=ly.GRULayer,
output_nonlinearity=None,
input_var=None,
input_layer=None,
layer_args=None):
with tf.variable_scope(name, 'GRUNetwork'):
if input_layer is None:
l_in = ly.InputLayer(shape=(None, None) + input_shape,
input_var=input_var,
name='input')
else:
l_in = input_layer
l_step_input = ly.InputLayer(shape=(None, ) + input_shape,
name='step_input')
l_step_prev_state = ly.InputLayer(shape=(None, hidden_dim),
name='step_prev_state')
if layer_args is None:
layer_args = dict()
l_gru = gru_layer_cls(l_in,
num_units=hidden_dim,
hidden_nonlinearity=hidden_nonlinearity,
gate_nonlinearity=recurrent_nonlinearity,
hidden_init_trainable=False,
w_x_init=recurrent_w_x_init,
w_h_init=recurrent_w_h_init,
name='gru',
**layer_args)
l_gru_flat = ly.ReshapeLayer(l_gru,
shape=(-1, hidden_dim),
name='gru_flat')
l_output_flat = ly.DenseLayer(l_gru_flat,
num_units=output_dim,
nonlinearity=output_nonlinearity,
w=output_w_init,
name='output_flat')
l_output = ly.OpLayer(
l_output_flat,
op=lambda flat_output, l_input: tf.reshape(
flat_output,
tf.stack(
(tf.shape(l_input)[0], tf.shape(l_input)[1], -1))),
shape_op=lambda flat_output_shape, l_input_shape:
(l_input_shape[0], l_input_shape[1], flat_output_shape[-1]),
extras=[l_in],
name='output')
l_step_state = l_gru.get_step_layer(l_step_input,
l_step_prev_state,
name='step_state')
l_step_hidden = l_step_state
l_step_output = ly.DenseLayer(l_step_hidden,
num_units=output_dim,
nonlinearity=output_nonlinearity,
w=l_output_flat.w,
b=l_output_flat.b,
name='step_output')
self._l_in = l_in
self._hid_init_param = l_gru.h0
self._l_gru = l_gru
self._l_out = l_output
self._l_step_input = l_step_input
self._l_step_prev_state = l_step_prev_state
self._l_step_hidden = l_step_hidden
self._l_step_state = l_step_state
self._l_step_output = l_step_output
self._hidden_dim = hidden_dim
def __init__(
self,
env_spec,
name="CategoricalGRUPolicy",
hidden_dim=32,
feature_network=None,
state_include_action=True,
hidden_nonlinearity=tf.tanh,
gru_layer_cls=L.GRULayer,
):
"""
:param env_spec: A spec for the env.
:param hidden_dim: dimension of hidden layer
:param hidden_nonlinearity: nonlinearity used for each hidden layer
:return:
"""
assert isinstance(env_spec.action_space, Discrete)
self._prob_network_name = "prob_network"
with tf.variable_scope(name, "CategoricalGRUPolicy"):
Serializable.quick_init(self, locals())
super(CategoricalGRUPolicy, 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))
prob_network = GRUNetwork(input_shape=(feature_dim, ),
input_layer=l_feature,
output_dim=env_spec.action_space.n,
hidden_dim=hidden_dim,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=tf.nn.softmax,
gru_layer_cls=gru_layer_cls,
name=self._prob_network_name)
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:
with tf.name_scope("feature_network", values=[flat_input_var]):
feature_var = L.get_output(
l_flat_feature,
{feature_network.input_layer: flat_input_var})
with tf.name_scope(self._prob_network_name, values=[feature_var]):
out_prob_step, out_prob_hidden = L.get_output(
[
prob_network.step_output_layer,
prob_network.step_hidden_layer
], {prob_network.step_input_layer: feature_var})
out_prob_step = tf.identity(out_prob_step, "prob_step_output")
out_prob_hidden = tf.identity(out_prob_hidden,
"prob_step_hidden")
self.f_step_prob = tensor_utils.compile_function(
[flat_input_var, prob_network.step_prev_state_layer.input_var],
[out_prob_step, out_prob_hidden])
self.input_dim = input_dim
self.action_dim = action_dim
self.hidden_dim = hidden_dim
self.name = name
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,
env_spec,
name='CategoricalLSTMPolicy',
hidden_dim=32,
hidden_nonlinearity=tf.nn.tanh,
recurrent_nonlinearity=tf.nn.sigmoid,
recurrent_w_x_init=L.XavierUniformInitializer(),
recurrent_w_h_init=L.OrthogonalInitializer(),
output_nonlinearity=tf.nn.softmax,
output_w_init=L.XavierUniformInitializer(),
feature_network=None,
prob_network=None,
state_include_action=True,
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:
"""
assert isinstance(env_spec.action_space, Discrete)
self._prob_network_name = 'prob_network'
with tf.variable_scope(name, 'CategoricalLSTMPolicy'):
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,
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,
forget_bias=forget_bias,
use_peepholes=use_peepholes,
lstm_layer_cls=lstm_layer_cls,
name=self._prob_network_name)
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:
with tf.name_scope('feature_network', values=[flat_input_var]):
feature_var = L.get_output(
l_flat_feature,
{feature_network.input_layer: flat_input_var})
with tf.name_scope(self._prob_network_name, values=[feature_var]):
out_prob_step, out_prob_hidden, out_step_cell = 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.f_step_prob = tensor_utils.compile_function([
flat_input_var,
prob_network.step_prev_state_layer.input_var,
], [out_prob_step, out_prob_hidden, out_step_cell])
self.input_dim = input_dim
self.action_dim = action_dim
self.hidden_dim = hidden_dim
self.name = name
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 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 __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)
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
def __init__(self,
input_shape,
output_dim,
hidden_dim,
name=None,
hidden_nonlinearity=tf.nn.relu,
lstm_layer_cls=ly.LSTMLayer,
output_nonlinearity=None,
input_var=None,
input_layer=None,
forget_bias=1.0,
use_peepholes=False,
layer_args=None):
with tf.variable_scope(name, "LSTMNetwork"):
if input_layer is None:
l_in = ly.InputLayer(
shape=(None, None) + input_shape,
input_var=input_var,
name="input")
else:
l_in = input_layer
l_step_input = ly.InputLayer(
shape=(None, ) + input_shape, name="step_input")
# contains previous hidden and cell state
l_step_prev_state = ly.InputLayer(
shape=(None, hidden_dim * 2), name="step_prev_state")
if layer_args is None:
layer_args = dict()
l_lstm = lstm_layer_cls(
l_in,
num_units=hidden_dim,
hidden_nonlinearity=hidden_nonlinearity,
hidden_init_trainable=False,
name="lstm_layer",
forget_bias=forget_bias,
cell_init_trainable=False,
use_peepholes=use_peepholes,
**layer_args)
l_lstm_flat = ly.ReshapeLayer(
l_lstm, shape=(-1, hidden_dim), name="lstm_flat")
l_output_flat = ly.DenseLayer(
l_lstm_flat,
num_units=output_dim,
nonlinearity=output_nonlinearity,
name="output_flat")
l_output = ly.OpLayer(
l_output_flat,
op=lambda flat_output, l_input: tf.reshape(
flat_output,
tf.stack((tf.shape(l_input)[0], tf.shape(l_input)[1], -1))
),
shape_op=lambda flat_output_shape, l_input_shape: (
l_input_shape[0], l_input_shape[1], flat_output_shape[-1]),
extras=[l_in],
name="output")
l_step_state = l_lstm.get_step_layer(
l_step_input, l_step_prev_state, name="step_state")
l_step_hidden = ly.SliceLayer(
l_step_state, indices=slice(hidden_dim), name="step_hidden")
l_step_cell = ly.SliceLayer(
l_step_state,
indices=slice(hidden_dim, None),
name="step_cell")
l_step_output = ly.DenseLayer(
l_step_hidden,
num_units=output_dim,
nonlinearity=output_nonlinearity,
w=l_output_flat.w,
b=l_output_flat.b,
name="step_output")
self._l_in = l_in
self._hid_init_param = l_lstm.h0
self._cell_init_param = l_lstm.c0
self._l_lstm = l_lstm
self._l_out = l_output
self._l_step_input = l_step_input
self._l_step_prev_state = l_step_prev_state
self._l_step_hidden = l_step_hidden
self._l_step_cell = l_step_cell
self._l_step_state = l_step_state
self._l_step_output = l_step_output
self._hidden_dim = hidden_dim
def __init__(self,
input_shape,
output_dim,
hidden_dim,
name=None,
hidden_nonlinearity=tf.nn.relu,
gru_layer_cls=ly.GRULayer,
output_nonlinearity=None,
input_var=None,
input_layer=None,
layer_args=None):
with tf.variable_scope(name, "GRUNetwork"):
if input_layer is None:
l_in = ly.InputLayer(
shape=(None, None) + input_shape,
input_var=input_var,
name="input")
else:
l_in = input_layer
l_step_input = ly.InputLayer(
shape=(None, ) + input_shape, name="step_input")
l_step_prev_state = ly.InputLayer(
shape=(None, hidden_dim), name="step_prev_state")
if layer_args is None:
layer_args = dict()
l_gru = gru_layer_cls(
l_in,
num_units=hidden_dim,
hidden_nonlinearity=hidden_nonlinearity,
hidden_init_trainable=False,
name="gru",
**layer_args)
l_gru_flat = ly.ReshapeLayer(
l_gru, shape=(-1, hidden_dim), name="gru_flat")
l_output_flat = ly.DenseLayer(
l_gru_flat,
num_units=output_dim,
nonlinearity=output_nonlinearity,
name="output_flat")
l_output = ly.OpLayer(
l_output_flat,
op=lambda flat_output, l_input: tf.reshape(
flat_output,
tf.stack((tf.shape(l_input)[0], tf.shape(l_input)[1], -1))
),
shape_op=lambda flat_output_shape, l_input_shape: (
l_input_shape[0], l_input_shape[1], flat_output_shape[-1]),
extras=[l_in],
name="output")
l_step_state = l_gru.get_step_layer(
l_step_input, l_step_prev_state, name="step_state")
l_step_hidden = l_step_state
l_step_output = ly.DenseLayer(
l_step_hidden,
num_units=output_dim,
nonlinearity=output_nonlinearity,
w=l_output_flat.w,
b=l_output_flat.b,
name="step_output")
self._l_in = l_in
self._hid_init_param = l_gru.h0
self._l_gru = l_gru
self._l_out = l_output
self._l_step_input = l_step_input
self._l_step_prev_state = l_step_prev_state
self._l_step_hidden = l_step_hidden
self._l_step_state = l_step_state
self._l_step_output = l_step_output
self._hidden_dim = hidden_dim