search

official-stockfish / nnue-pytorch

Stockfish NNUE (Chess evaluation) trainer in Pytorch
GNU General Public License v3.0
320 stars 99 forks source link

Testing trained model #243

Open benliu961 opened 1 year ago

benliu961 commented 1 year ago

Hello,

I have a trained model. How would I get predictions and testing accuracy using a test dataset?

Thank you!

vondele commented 1 year ago

That would be the validation loss, but generally we test newly trained nets by playing games with run_games.py. easy_train.py shows how this can be done. Nets that test well locally can be tested on fishtest as well.

benliu961 commented 1 year ago

If I fine tuned the model so that it doesn't predict board evaluations anymore, how would I test it?

Blimpyway commented 1 year ago

What does it predict?

benliu961 commented 1 year ago

It predicts a value that I have defined as Value of Computation. Its just some value between 0 and 1

Sopel97 commented 1 year ago

Unless this value represents some truth there is no way to test other than play games.

benliu961 commented 1 year ago

Is there no way to test it via tensorflow?

Sopel97 commented 1 year ago

best you can do is look at loss