def batch_tensor(self, name):
""" A buffer of a given value in a 'flat' (minibatch-indexed) format """
if name in self.transition_tensors:
return tensor_util.merge_first_two_dims(
self.transition_tensors[name])
else:
return self.rollout_tensors[name]
def to_transitions(self) -> 'Transitions':
""" Convert given rollout to Transitions """
# No need to propagate 'rollout_tensors' as they won't mean anything
return Transitions(
size=self.num_steps * self.num_envs,
environment_information=
[ei for l in self.environment_information for ei in l]
if self.environment_information is not None else None,
transition_tensors={
name: tensor_util.merge_first_two_dims(t) for name, t in self.transition_tensors.items()
},
extra_data=self.extra_data
)
def to_transitions(self) -> 'Transitions':
""" Convert given rollout to Transitions """
# No need to propagate 'rollout_tensors' as they won't mean anything
transition_tensor = {}
for name, t in self.transition_tensors.items():
if isinstance(t, dict):
transition_tensor[name] = {
k: tensor_util.merge_first_two_dims(v)
for k, v in t.items()
}
else:
transition_tensor[name] = tensor_util.merge_first_two_dims(t)
return Transitions(
size=self.num_steps * self.num_envs,
environment_information=[
ei for l in self.environment_information for ei in l
],
transition_tensors=transition_tensor,
# {
# name: tensor_util.merge_first_two_dims(t) for name, t in self.transition_tensors.items()
# },
extra_data=self.extra_data)
def flatten_tensor(self, tensor):
""" Merge first two dims of a tensor """
return tensor_util.merge_first_two_dims(tensor)