Train

Train your own Policy Network and Value Network.

Utilities for training is availabel at https://github.com/zakki/rn-tools

Requires

• Windows
• Cygwin
• Perl
• Java

Step

1. Convert KGS/GoGod/Tygem sgf files to gtp files which interleave custom gtp commands: _store/_dump/_clear with sgfvar2gtp2.sh Filtering by komi and rank are hard coded in sgfvar2gtp2.pl.
2. Run ray with converted gtp file to generate 'data.txt'. This generates 100GB~2TB data file.
3. Shuffle and split data.txt using jshuf.
4. Run cntk. Model definition is https://github.com/zakki/Ray/blob/nn/cntk/. Since CNTK can't handle over 100GBs data on my PC, I stop cntk every some epochs and I switch data.txt using loop-v20.sh

Step 4 is needed by limitation of CNTK 1. If you use CNTK 2, I think there are more sophisticated way that uses Python API.

Version

model3.bin

"exp-gnugo-value" branch (e.g. Rn.4.9.1) outputs data.txt for model3.bin that is trained by ResNetV97.bs and modified TrainV50.cntk.

safety = [
  dim = 2888
  format = "sparse"
]

model2.bin

"nn" branch (e.g. Rn.4.20) outputs data.txt for model2.bin that is trained by TrainV50.cntk and ResNetV91b.bs.

model.bin

Rn.3.x outputs data.txt for model.bin that is trained by ResNetV25.ndl TrainValueResNetV20.cntk