Ceibo 1.0 Released

[federicorojo]

Essentially it is Ceibo 0.9 , but have replaced the Evaluation with the tiniest of neural networks (768x128x1).

The training data was created based on Ceibo 0.9 self play at 4k nodes only, for a total of 64MM unique samples.

As for the trainer, I created several from scratch to train the net embedded into the executable. First few versions of the trainer were created in python, then C++ and later added SIMD instructions to speed up the training sessions. However, although would create working nets, it was still super slow, and was unable to experiment with different hyperparameter values due to the extremely long training sessions.

Then I ventured into CUDA and quickly gave up. So i decided to create a new version of the trainer, but this time using libtorch (C++ pytorch) where I didnt need to worry about learning CUDA at all. With this last trainer it would need about 15hrs to process 64MM training samples during an average of 350 epochs. Probably still very slow , but there's another reason for that, and still much faster than previous versions.

In terms of GPU hardware, I was limited in terms of processing power and VRAM as I only had at my disposal an old nVidia GTX970 to train the net used on Ceibo 1.0.


Here is the link -> https://sites.google.com/site/ceibochess/downloads



Happy New Year !!

[Graham Banks]

Thanks.

I needed to request permission to download.

How much stronger roughly?

[federicorojo]

Apologies! The file should be available to anyone now.

ELO wise, about 100.

Thanks