def __init__(self,
*,
stddev_mode='predicted',
bounded_transform='tanh',
name=None,
action_spec=None,
input_spec=None):
assert action_spec.type == 'float'
parameters_spec = TensorsSpec(
mean=TensorSpec(type='float', shape=action_spec.shape),
stddev=TensorSpec(type='float', shape=action_spec.shape),
log_stddev=TensorSpec(type='float', shape=action_spec.shape))
conditions_spec = TensorsSpec()
super().__init__(name=name,
action_spec=action_spec,
input_spec=input_spec,
parameters_spec=parameters_spec,
conditions_spec=conditions_spec)
self.stddev_mode = stddev_mode
if bounded_transform is None:
bounded_transform = 'tanh'
if bounded_transform not in ('clipping', 'tanh'):
raise TensorforceError.value(name='Gaussian',
argument='bounded_transform',
value=bounded_transform,
hint='not in {clipping,tanh}')
elif bounded_transform == 'tanh' and (
(self.action_spec.min_value
is not None) is not (self.action_spec.max_value is not None)):
raise TensorforceError.value(
name='Gaussian',
argument='bounded_transform',
value=bounded_transform,
condition='one-sided bounded action space')
elif self.action_spec.min_value is None and self.action_spec.max_value is None:
bounded_transform = None
self.bounded_transform = bounded_transform
if self.input_spec.rank == 1:
# Single embedding
self.mean = self.submodule(name='mean',
module='linear',
modules=layer_modules,
size=self.action_spec.size,
initialization_scale=0.01,
input_spec=self.input_spec)
if self.stddev_mode == 'predicted':
self.stddev = self.submodule(name='stddev',
module='linear',
modules=layer_modules,
size=self.action_spec.size,
initialization_scale=0.01,
input_spec=self.input_spec)
else:
# Embedding per action
if self.input_spec.rank < 1 or self.input_spec.rank > 3:
raise TensorforceError.value(name=name,
argument='input_spec.shape',
value=self.embedding_shape,
hint='invalid rank')
elif self.input_spec.shape[:-1] == self.action_spec.shape[:-1]:
size = self.action_spec.shape[-1]
elif self.input_spec.shape[:-1] == self.action_spec.shape:
size = 0
else:
raise TensorforceError.value(
name=name,
argument='input_spec.shape',
value=self.input_spec.shape,
hint='not flattened and incompatible with action shape')
self.mean = self.submodule(name='mean',
module='linear',
modules=layer_modules,
size=size,
initialization_scale=0.01,
input_spec=self.input_spec)
if self.stddev_mode == 'predicted':
self.stddev = self.submodule(name='stddev',
module='linear',
modules=layer_modules,
size=size,
initialization_scale=0.01,
input_spec=self.input_spec)