Visualizing chess position embeddings

[bastiansg]

Hello,

I am currently working on a project where I aim to visualize embeddings of various chess positions. My goal is to appreciate some form of semantic segmentation for these positions.

I want to use nnue-pytorch for that, and I would like to know how I can, given a chess position, input it into the PyTorch network and extract some intermediate layer representation to be used as an embedding.

Thank you in advance for your time and assistance.

[Sopel97]

input it into the PyTorch network

The simplest example would be here https://github.com/official-stockfish/nnue-pytorch/blob/master/cross_check_eval.py

and extract some intermediate layer representation

You need to make modifications to model.py to pass the intermediate layer outputs down the stack. The intermediate computations are not saved anywhere currently

[bastiansg]

input it into the PyTorch network

The simplest example would be here https://github.com/official-stockfish/nnue-pytorch/blob/master/cross_check_eval.py

and extract some intermediate layer representation

You need to make modifications to model.py to pass the intermediate layer outputs down the stack. The intermediate computations are not saved anywhere currently

Great, I'll look into it, thank you very much!

[bastiansg]

Is there any place from where I can download the already trained torch model used in Stockfish 16.1?

[Sopel97]

would have to ask @linrock

you can also convert any .nnue network back to a pytorch model (though of course it will be lossy compared to the original model).

python serialize.py --features=HalfKAv2_hm from.nnue to.pt

though you should be able to just just .nnue models

[linrock]

Is there any place from where I can download the already trained torch model used in Stockfish 16.1?

that's long gone. i stopped keeping trained .ckpt files around since they're huge and i never use them.

using .nnue files should accomplish what you want.

[bastiansg]

Is there any place from where I can download the already trained torch model used in Stockfish 16.1?

that's long gone. i stopped keeping trained .ckpt files around since they're huge and i never use them.

using .nnue files should accomplish what you want.

You mean using the serialize.py script mentioned by @Sopel97 ? what about what he mentioned about that the .nnue will be lossy compared to the original model?

[vondele]

The lossy part is very minimal, can definitely be ignored for visualization.