def __init__(
self,
name,
env_spec,
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:
"""
with tf.variable_scope(name):
assert isinstance(env_spec.action_space, Discrete)
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.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 = 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="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
],
L.get_output([
prob_network.step_output_layer,
prob_network.step_hidden_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.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,
hidden_dim,
hidden_nonlinearity=tf.nn.relu,
lstm_layer_cls=L.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):
if input_layer is None:
l_in = L.InputLayer(shape=(None, None) + input_shape,
input_var=input_var,
name="input")
else:
l_in = input_layer
l_step_input = L.InputLayer(shape=(None, ) + input_shape,
name="step_input")
# contains previous hidden and cell state
l_step_prev_state = L.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",
forget_bias=forget_bias,
cell_init_trainable=False,
use_peepholes=use_peepholes,
**layer_args)
l_lstm_flat = L.ReshapeLayer(l_lstm,
shape=(-1, hidden_dim),
name="lstm_flat")
l_output_flat = L.DenseLayer(l_lstm_flat,
num_units=output_dim,
nonlinearity=output_nonlinearity,
name="output_flat")
l_output = L.OpLayer(
l_output_flat,
op=lambda flat_output, l_input: tf.reshape(
flat_output,
tf.pack((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 = L.SliceLayer(l_step_state,
indices=slice(hidden_dim),
name="step_hidden")
l_step_cell = L.SliceLayer(l_step_state,
indices=slice(hidden_dim, None),
name="step_cell")
l_step_output = L.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,
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.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 = 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_hidden_layer.input_var,
mean_network.step_prev_cell_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 __init__(self,
name,
input_shape,
output_dim,
hidden_dim,
hidden_nonlinearity=tf.nn.relu,
gru_layer_cls=L.GRULayer,
output_nonlinearity=None,
input_var=None,
input_layer=None,
layer_args=None):
with tf.variable_scope(name):
if input_layer is None:
l_in = L.InputLayer(shape=(None, None) + input_shape,
input_var=input_var,
name="input")
else:
l_in = input_layer
l_step_input = L.InputLayer(shape=(None, ) + input_shape,
name="step_input")
l_step_prev_state = L.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 = L.ReshapeLayer(l_gru,
shape=(-1, hidden_dim),
name="gru_flat")
l_output_flat = L.DenseLayer(l_gru_flat,
num_units=output_dim,
nonlinearity=output_nonlinearity,
name="output_flat")
l_output = L.OpLayer(
l_output_flat,
op=lambda flat_output, l_input: tf.reshape(
flat_output,
tf.pack((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 = L.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