Step by Step Guide

1. Check that your GPU supports CUDA

https://developer.nvidia.com/cuda-gpus

2. Install CUDA

https://developer.nvidia.com/cuda-downloads

3. Download training data generator

https://github.com/fairy-stockfish/variant-nnue-tools/releases

• .exe for Windows, others for Linux
• -largeboard are for variants with boards >8x8
• Depending on your CPU use x86-64-bmi2, x86-64-modern, or `x86-64. In doubt use the latter.

If you aren't sure, use fairy-stockfish-tools-largeboard_x86-64.exe.

4. Generate training data

https://github.com/fairy-stockfish/variant-nnue-pytorch/wiki/Training-data-generation#training-data-generation-using-fairy-stockfish

5. Install git (optional)

6. Clone trainer

If you installed git, download the training code with

git clone https://github.com/fairy-stockfish/variant-nnue-pytorch.git

Otherwise you can download the code as a ZIP and extract it.

7. Install dependencies

https://github.com/fairy-stockfish/variant-nnue-pytorch#setup

Now you should be able to run python3 train.py -h. If there are problems with the installed dependencies it might report errors here, so you need to address these before continuing.

8. Apply code changes

https://github.com/fairy-stockfish/variant-nnue-pytorch/wiki/NNUE-training#code-changes

9. Compile training data loader

compile_data_loader.bat

10. Convert existing network (optional)

https://github.com/fairy-stockfish/variant-nnue-pytorch/wiki/NNUE-training#resume-from-existing-nnue-net

11. Run train.py

https://github.com/fairy-stockfish/variant-nnue-pytorch/wiki/NNUE-training#training-example

12. Convert checkpoint to NNUE network

https://github.com/fairy-stockfish/variant-nnue-pytorch/wiki/NNUE-training#converting-a-training-checkpoint-to-nnue-file

Now you finally generated your first own NNUE network.

13. Validate (optional)

If you are doing the training for the first time, you likely want to first validate if the generated NNUE file really works. For that, e.g., load it at https://fairy-stockfish-nnue-wasm.vercel.app/ under the nnue file. If it reports an ERROR please check again if your definition in variant.py fits to the variant you wanted to train.

14. Test playing strength

Depending on the variant there are different options how to compare playing strength to the previous best NNUE network. One option that works for all variants is https://github.com/ianfab/variantfishtest.

15. Upload

If testing shows that your NNUE network performs well, please upload it at https://forms.gle/8Am9LTqXQJo43ps79 in order to share it with others. It will automatically be made available at https://drive.google.com/drive/folders/1m5PpiI3Kjzk_ow7F5RkwKnbO0Td-qb9J?usp=sharing then. The file name will be derived from the variant name and the value you added for the SHA256.