Performance decrease after recent commits about tensorflow-deepq HOT 6 CLOSED

I also noticed excessive memory usage which didn't happen before.

from tensorflow-deepq.

Yeah, I notice that as well. Maintaining two separate copies of the network is critical for the deep q performance, so I really want to keep it, but I am not sure what causes the leak. I will keep looking. Let me know if you find anything.

One possible thing to investigate is logging. (try passing summary writter=None

from tensorflow-deepq.

Fixed in 0d848e5

from tensorflow-deepq.

Thanks!

So the issue was caused by SummaryWriter and not by having 2 q-networks. That's good to know ) Though, graphs were beautiful ) Do you think it's a TF issue?

from tensorflow-deepq.

Nah, check out the closing commit. The issue was the fact that I create multiple tf.no_op() operations (one every training step). Apparently that's really bad. I was extremely surprised. But I guess there's some logic to it...

from tensorflow-deepq.

Aaah... I saw it, but you set calculate_summaries = False as well, so I thought that was the main reason. Now I enabled summary and see that it works as fast as before - that's great!
Agree that there could be logic - if you create thousands instances of new nodes, even if they are exact the same, it can very wasteful, if not cached somehow internally in TF.

from tensorflow-deepq.