Open ruipengZ opened 5 months ago
My apologies that I've completely missed this issue. In case this is still relevant, from the logs I cannot determine much to be fundamentally wrong, apart from that tensorflow was not able to register a GPU (either due to a lack of one or because CUDA is not properly integrated). This slows down the process a lot. From the logs you can see that it takes ~2 minutes for a single policy update. This is normally a matter of seconds on a GPU. Moreover, it may take more than 15 epochs before the agent has learned any meaningful behavior. I suggest ensuring that CUDA is properly installed to support using a GPU and then trying again.
Hi! I have trouble reproducing the result for the single environment run_and_gun. I run
python run_single.py --scenario run_and_gun
and get the following prints: (after 15 epochs, the training success is still 0)