Potential bug during training?

[liubaoryol]

Is there a reason you calculate the reward the way you do in line 69?

gail-airl-ppo.pytorch/gail_airl_ppo/algo/airl.py

Line 69 in 4e13a23

rewards = self.disc.calculate_reward(

My models were able to learn after I changed that line to

        with torch.no_grad():
            rewards = self.disc.g(states)

This gives the unshaped rewards

[Charlesyyun]

Did that work out for you? I found my actor loss unable to converge.

[liubaoryol]

Yes it did, although I was running it on discrete state and action environments. Which env are you using?

[mikhail-vlasenko]

@liubaoryol it is great to hear that you got it working with discrete action space! could you please share your code? i think it will be valuable, as multiple people here already asked about discrete action support. Thanks in advance

[liubaoryol]

Of course! Let me clean it up and I'll share it next week:)

[jagandecapri]

I'm interested to know about the implementation for discrete action support too. :)

[ChenYunan]

reward = -logsigmoid(-logits) = -log[1 - sigmoid(logits)] = -log(1 - D), which corresponds the objective of G is minimize log(1-D).