Getting actions probabilities instead of an unique prediction in Stable Baselines 3 SAC?

Question

I try to understand how getting an actions probability table instead of an unique prediction in stable baselines 3 SAC in order to override 'predict' method to filter invalid actions. I guess the good way to do that is to remove invalid actions from table and resize probabilities.

Models predictions can be deterministic or not but i can't find any function that returns action probabilities table (such function was available in SB2 if i remember well...)

I read these topics but they are not about SAC :

https://github.com/DLR-RM/stable-baselines3/issues/1347

https://github.com/DLR-RM/stable-baselines3/issues/568

https://github.com/DLR-RM/stable-baselines3/issues/385

Best regards.

Answer 1

Convert your observation to a tensor of the right shape if it is not already

obs_tensor = torch.tensor(obs.reshape((1,-1)), dtype=torch.float32)

with torch.no_grad():
    distribution = model.policy.get_distribution(obs_tensor)

The distribution object has logits and probs attributes:

print(distribution.distribution.probs)