search

glinscott / leela-chess

**MOVED TO https://github.com/LeelaChessZero/leela-chess ** A chess adaption of GCP's Leela Zero
http://lczero.org
GNU General Public License v3.0
760 stars 301 forks source link

Normalize the 50 move rule input in training #691

Closed Tilps closed 6 years ago

Tilps commented 6 years ago

This is kind of a merge of #605 and #655 to fix #602. Unlike #655 this one puts the weight manipulation in the training scripts themselves rather than in separate awk scripts.

Note: if this PR is accepted, the correct flow to perform the transition is, stop training, run net_to_model.py with the most recent training output, then resume training.

Basic testing still in progress.

Tilps commented 6 years ago

The load and save functions appear to roundtrip with 0 diff successfully - so that is something.