Memory leak with DJL >0.8.0

[nwjnilsson]

Hi,

I'm facing a bit of an issue with newer versions of DJL. My project is based on a Q-learning example taken from https://github.com/kingyuluk/RL-FlappyBird, but I wanted to run on the later versions of DJL, and the problem is that by simply specifying for example version 0.15.0 in the pom, the memory consumption goes out of control. This does not occur with 0.8.0, which that project used originally.

The only changes I made when selecting a newer version (besides the pom) was changing the Tracker declaration on line 92 and the initializer on line 217 in src/main/java/com/kingyu/rlbird/ai/TrainBird.java. So instead of Tracker exploreRate = new LinearTracker.Builder()..., I have Tracker exploreRate = LinearTracker.builder()... and on line 217 I just added Parameter.Type.WEIGHT as a second argument to optInitializer(...).

I don't see why those changes would cause any issues and I can't find anything else that wouldn't be compatible with 0.15.0. I would greatly appreciate if someone could help me out here so that I'm not stuck with DJL 0.8.0 and PyTorch 1.6.0.

Edit: I should probably add that I'm running on Linux, CPU only and Java 8.

Edit 2: Looked through the history of the TicTacToe RL example in repo as well since the FlappyBird example is clearly build off of that, and I just don't get what the problem is..

Edit 3: I think I found the issue. Seems to be some incorrect usage of NDManagers so I guess some resources weren't released.

Thanks,
J