Closed tatsubori closed 5 years ago
Hello,
Thank you for your interest.
I like the idea. Please don't forget to update the readme with the new instructions ;)
The CI is failing because the docker image does not have minigrid installed... please add an exception in the test script so your trained models are not tested.
The CI is failing because the docker image does not have minigrid installed... please add an exception in the test script so your trained models are not tested.
Added --gym-packages gym_minigrid if "-MiniGrid-" in trained_model. pip install gym-minigrid for docker scripts (cpu & gpu) as well.
Good, I'll test that on my local machine before pushing a new docker image (so the CI can pass)
I got qt segmentation fault on my machine trying to render the scene :/ Anyway, it seems ok otherwise.
I would also use the flat obs wrapper from minigrid: https://github.com/maximecb/gym-minigrid/blob/master/gym_minigrid/wrappers.py#L167
Better to exclude the cases from the test case for now?
Better to exclude the cases from the test case for now?
yep! (that was my first remark ;))
@tatsubori is that ok if I merge now and in another PR, you'll include a better support for minigrid? (using FlatObsWrapper for instance)
Yes, please proceed to merge. I have CNN impl with it at hand but it won’t converge. I am trying CNN+Lstm but it may take time.
Will continue enhancing it.
StableBaselines to support (not to fail) Dict is better to handle mission text, I believe, though.
StableBaselines to support (not to fail)
I agree, but the support is far from trivial ;) (however we would appreciate PR to solve that issue)
Congratulations of this very very handy work making everyone happy. :->
Just want to start discussing this quick hack enabling to use MiniGrid gyms. https://github.com/maximecb/gym-minigrid but not limited to this.
Instead of hard-coding imports, I have provided a command-line option to dynamically import specified gym packages for train.py and enjoy.py.
Adaptation of observation space (Dict) is hard-coded.
I am not sure if this breaks others, while I tried to be careful not to.
Cheers, Mich