fixed entropy for multi-discrete action spaces

cleanrl/ppo_trxl/ppo_trxl.py

@@ -315,7 +315,7 @@ def get_action_and_value(self, x, memory, memory_mask, memory_indices, action=No
         log_probs = []
         for i, dist in enumerate(probs):
             log_probs.append(dist.log_prob(action[:, i]))
-        entropies = torch.stack([dist.entropy() for dist in probs], dim=1)
+        entropies = torch.stack([dist.entropy() for dist in probs], dim=1).sum(1).reshape(-1)
         return action, torch.stack(log_probs, dim=1), entropies, self.critic(x).flatten(), memory
 
     def reconstruct_observation(self):