PPO with composite distribution crash before giving the warning on how to fix it.

[siegelaaron94]

This block

rl/torchrl/objectives/ppo.py

Lines 601 to 602 in 795e362

	if log_prob.batch_size != adv_shape:
	log_prob.batch_size = adv_shape

causes
AttributeError: 'Tensor' object has no attribute 'batch_size'

Before, the warning on how to fix it is shown.

rl/torchrl/objectives/ppo.py

Lines 603 to 614 in 795e362

	if not is_tensor_collection(prev_log_prob):
	# this isn't great: in general, multi-head actions should have a composite log-prob too
	warnings.warn(
	"You are using a composite distribution, yet your log-probability is a tensor. "
	"Make sure you have called tensordict.nn.set_composite_lp_aggregate(False).set() at "
	"the beginning of your script to get a proper composite log-prob.",
	category=UserWarning,
	)
	if is_tensor_collection(log_prob):
	log_prob = _sum_td_features(log_prob)
	log_prob.view_as(prev_log_prob)

The order of the 2 blocks needs to be swapped