KataGo icon indicating copy to clipboard operation
KataGo copied to clipboard

Published 20 hours ago • verified publisher icon lightvector

Not tfrecord but npz

Open Harder-Run opened this issue 1 year ago • 3 comments

When I ran shuffle.sh, the result was: Number of rows by output file: [('D:/KataGo/python/shuffleddata/20240120-093419/train\data0.npz', 75008), ('D:/KataGo/python/shuffleddata/20240120-093419/train\data1.npz', 74752), ('D:/KataGo/python/shuffleddata/20240120-093419/train\data2.npz', 74752), ('D:/KataGo/python/shuffleddata/20240120-093419/train\data3.npz', 74752)] The output files are't tfrecord. How could I resolve it?

Harder-Run avatar Jan 20 '24 01:01 Harder-Run

Tensorflow training is no longer implemented or supported. If you need to train with Tensorflow, you need to use a much older version of the repo. Pytorch training uses the npz files.

lightvector avatar Jan 20 '24 06:01 lightvector

Thanks. But it seemed to be wrong. When I exported the weight model.bin.gz, the weight size was only 12KB(b5c192nbt) and then the load failed.

Harder-Run avatar Jan 26 '24 11:01 Harder-Run

model.bin.gz error

Harder-Run avatar Jan 26 '24 13:01 Harder-Run