def get_variables(cls, all_units, vf_share_layers=False, name=None):
name = name or "keras_mlp_policy"
if "policy" in name:
last_kernel_init_value = 0.01
elif "vf" in name:
last_kernel_init_value = 1.0
else:
raise NotImplementedError
variables = {}
with tf.name_scope(name):
for i, (size_in, size_out) in \
enumerate(zip(all_units, all_units[1:])):
name = f"dense_{i}"
variables[name] = \
Dense.get_variables(
size_in,
size_out,
name=name,
kernel_initializer=(
normc_initializer(1.0) if i < len(all_units) - 2
else normc_initializer(last_kernel_init_value)))
if vf_share_layers:
name = f"dense_vf"
variables[name] = \
Dense.get_variables(
size_in,
1,
name=name,
kernel_initializer=normc_initializer(1.0))
# tricky to remove the name count of the dummy instance
# since the default name is used here
# graph = tf.get_default_graph()
# K.PER_GRAPH_LAYER_NAME_UIDS[graph][
# ("", dummy_instance.name)] -= 1
return variables
def get_variables(
cls,
latent_units,
projection_kernel_init_value,
# mlp_input_units,
# mlp_units,
# vf_share_layers=False
name=None):
name = name or "keras_maesn_policy"
variables = {}
with tf.name_scope(name):
variables["latent_z"] = tf.Variable(
initial_value=normc_initializer(projection_kernel_init_value)(
shape=(latent_units, )),
trainable=True,
dtype=tf.float32,
name="latent_z")
# variables["cell"] = GRUCell.get_variables(
# rnn_input_units, rnn_units)
# variables["projection"] = Dense.get_variables(
# rnn_units,
# rnn_output_units,
# name="dense_projection",
# kernel_initializer=normc_initializer(projection_kernel_init_value))
# variables["mlp"] = KerasMLP.get_dummy_variables(
# [mlp_input_units + rnn_output_units] + mlp_units,
# vf_share_layers=vf_share_layers, **mlp_kwargs)
# graph = tf.get_default_graph()
# K.PER_GRAPH_LAYER_NAME_UIDS[graph][
# ("", dummy_instance.name)] -= 1
return variables
def get_variables(cls,
rnn_input_units,
rnn_units,
rnn_output_units,
projection_kernel_init_value,
# mlp_input_units,
# mlp_units,
# vf_share_layers=False
name=None):
name = name or "keras_tesp_policy"
variables = {}
with tf.name_scope(name):
variables["cell"] = GRUCell.get_variables(
rnn_input_units, rnn_units)
variables["projection"] = Dense.get_variables(
rnn_units,
rnn_output_units,
name="dense_projection",
kernel_initializer=normc_initializer(projection_kernel_init_value))
# variables["mlp"] = KerasMLP.get_dummy_variables(
# [mlp_input_units + rnn_output_units] + mlp_units,
# vf_share_layers=vf_share_layers, **mlp_kwargs)
# graph = tf.get_default_graph()
# K.PER_GRAPH_LAYER_NAME_UIDS[graph][
# ("", dummy_instance.name)] -= 1
return variables
def __init__(self,
layer_units_exclude_first,
activation,
custom_params=None,
vf_share_layers=False,
name=None):
"""
layer_units: list, a list of the number of units of all layers
except the input layer
"""
name = name or "keras_mlp_policy"
if "policy" in name:
last_kernel_init_value = 0.01
elif "vf" in name:
last_kernel_init_value = 0.01
else:
raise NotImplementedError
keras_models.Model.__init__(self, name=name)
custom_params = custom_params or {}
for i, size in enumerate(layer_units_exclude_first):
name = f"dense_{i}"
layer = Dense(
size,
custom_params=custom_params.get(name),
activation=(activation if
i < len(layer_units_exclude_first) - 1 else None),
kernel_initializer=(normc_initializer(1.0) if
i < len(layer_units_exclude_first) - 1 else
normc_initializer(last_kernel_init_value)),
name=name)
setattr(self, name, layer)
if vf_share_layers:
name = f"dense_vf"
layer = Dense(1,
custom_params=custom_params.get(name),
activation=None,
kernel_initializer=normc_initializer(1.0),
name=name)
setattr(self, name, layer)
self._vf_share_layers = vf_share_layers