F. Huizinga

Confirmed.

Hi, I'm experiencing the same problems and am able to reproduce it consistently with the following steps: 1. wait until laptop locks for being idle 2. close the laptop lid...

Agreed, I've got a system here with an _Intel(R) Core(TM) i7-6800K CPU @ 3.40GHz_ and two 1080Ti's. Running `run.py self` is not saturating GPU:0 nor is it saturating the CPU...

Hi @mokemokechicken, Thanks for your response. I don't see how multiple self instances follow different trajectories. Can you elaborate? (s, a) Bookkeeping isn't shared across instances right?

Right, I see. I'm also still confused as to what exactly defines a training example (_s_t_, **pi_t**, _z_t_). Is _z_t_ in {-1,0,+1}? Is _s_t_ then any node in the game...

Ok so a single game produces a dataset with a single outcome _z_ from a player's perspective. Each game move is obtained by MCTS trajectories. Each node in each trajectory...

Ahhh this makes things click for me. Thank you. To get back on topic. I think the Python GIL is presenting a bottleneck here. Ideally you'd want to start multiple...

gcc-6 does work.

I disabled all tensorflow parts as cudnn is superior anyway. Just disable everything in the meson build that is related to tensorflow. Also I'm working on a cmake :)

This patch should work for Ubuntu 16.04 with cudnn 7.1.2, cuda 9.0, clang-6.0 and gcc-6 ```diff diff --git a/lc0/build.sh b/lc0/build.sh index 00c52ff..656088c 100755 --- a/lc0/build.sh +++ b/lc0/build.sh @@ -1,7 +1,7...