search

matpalm / cartpoleplusplus

3d cartpole gym env using bullet physics trained from pixels with tensorflow LRPG, DDPG & NAF
http://matpalm.com/blog/cartpole_plus_plus/
MIT License
57 stars 14 forks source link

Fix problematic case with opts==None and therefore no opts.use_dropout. #3

Closed benelot closed 7 years ago

benelot commented 7 years ago

I found some problem(s) while reading through your code and running it with different settings. I think it would be helpful if I contributed some fix(es) back while I do this. For now, there is only one fix in the request but maybe there are more to come.

benelot commented 7 years ago

Oh ok, I will check later if I can add this quickly to the setup.py if you do not have time to fix it quickly. It seems to work with the other one as well btw, but I did not look into the details, I just ran it and it trained. So I can not say if it actually works or if the code just "runs".

matpalm commented 7 years ago

yeah the import would have worked but then it would have failed if you'd being using any type of event logging (i.e. --event-log=...) i use the logging of runs for debugging, prepopulating replay memories etc, but it's not a requirement, you can train / run without it (which is why i think i'll move the import to just before it's being used) thanks for the sanity checks!

benelot commented 7 years ago

Thanks, nice to know!