def _unwhiten_and_split(self, tensor_whitened, names):
lower, upper = self._env_spec.limits(names)
mean = 0.5 * (lower + upper)
var = 0.5 * (upper - lower)
tensor = tensor_whitened * var + mean
return AttrDict.from_dict({
k: v
for k, v in zip(
names, tf.split(tensor, self._env_spec.dims(names), axis=2))
})
def get_action(self, model, observation, goal):
assert self._session is not None
feed_dict = self._get_action_feed_dict(observation, goal)
get_action_tf = {}
for name, tensor in self._get_action_outputs.leaf_items():
get_action_tf[name] = tensor
get_action_tf_output = self._session.run(get_action_tf,
feed_dict=feed_dict)
get_action = AttrDict.from_dict(get_action_tf_output)
get_action.cost_fn = self._cost_fn
return get_action