Some problems when using this code in discrete action gym environment

[Felixvillas]

At first, Thank you very much for sharing this code!!!

Now I'm applying this open source code to discrete action environment of gym, but I've come across the following problems.

(1) From value of the info['logp_altz'] of agent.train_loop(), I find that for the same (s, s'), different skills get very similar log_probability, which seems that skill (i.e z) has very very little influence on q (s' | s, z). Do you have any guidance for this problem?

(2) And from the variation tendency of info['logp_altz']，the log_probability is always converge a terrible value in (-2, -1.8). So the probability is always converge a value located in (0.13, 0.17). Surely, the situation is different when I use different gym environment，but all of them converge a not good value.

(3) How can I record the loss value when I train skill dynamics? I'd like to know the trend of skill dynamics training.

(4) Is the skill type related to the action type (discrete or continue) of the environment? What I mean is can I use 'cont_uniform' of skill_type in a discrete action environment of gym. From reading of this code, I think the skill_type has nothing to do with action type.

Thank you for reading my questions. And I would appreciate it if you could answer some of them！！！