inoryy/tensorflow2-deep-reinforcement-learning

Agent not training with latest version of TF nightly

sibyjackgrove opened this issue · 4 comments

sibyjackgrove commented

It seems some change in the latest version of TF nightly caused the model to break. Now it is not training at all, and apparent behaving as a randomly initialized network. The same code was working for me and the agent was getting trained till last Sunday (29th). Could you please check what might have gone wrong.

Training plot:
image

TF version:
image

inoryy commented

Thanks for the report, can reproduce it locally. I'll look into it when I have a moment, but as a temporary workaround please force preview version to the one I've used in the blog, i.e.

pip install tf-nightly-2.0-preview==1.13.0-dev20190117
sibyjackgrove commented

Ok, will do that. Thanks!

inoryy commented

Fixed!

sibyjackgrove commented

Thanks, it works fine now!