Stockfish

Stockfish copied to clipboard

What is the master neural network referred to in SF source code near SF12 (first NNue)

1

I am trying to fill in some gaps in the documentation (including source code comments),and I have seen in that file change or addition intro blurb with refereences to transformer...

Dboingue

fstream won't create file under some OS if used with in|out|binary

1

this affects https://github.com/nodchip/Stockfish/blob/master/src/extra/nnue_data_binpack_format.h#L6718

tryingsomestuff

Sopel97

master and tools branches...

4

Since we now have the pytorch trainer and the architecture is becoming more complicated and less CPU friendly for training this trainer is less useful. It started causing issues when...

Sopel97

Shorter write_maxply and inaccurate result

2

When a user sets the write_maxply to 99 for example to terminate the game more quickly than the default 400 or has an intention to generate training positions at specific...

fsmosca

Measure the spread in Elo performance by random seed

1

In https://github.com/nodchip/Stockfish/issues/152, we have measured Elo performance between nets generated by different commits and random seeds. Random seeds should affect the spread in Elo performance. We should measure the spread...

nodchip

Invalid EP results in bad binpack conversion

6

The attached .plain file is converted to a bad .binpack file but passes validation. Conversion to .bin works, but again converting the .bin to .binpack again results in a bad...

borg323

[WIP] Fetch new validation data for each calc_loss

1

This PR makes `calc_loss` use a new validation set each time it's called (one can now limit the amount of validation data used per `calc_loss` call, so this makes sense)....

Sopel97

CI for Reinforcement Learning

3

Currently the CI for compilation and Supervised Learning were setup. We could also need the CI for Reinforcement Learning. i.e. CI for training data generation with an existing net file,...

nodchip

Compression of Feature transofrmer to 1/3 of size

4

I figured out how to efficiently compress FeatureTransformer weights in halfkp_256x2-32-32. It is exploiting fact that weights for pieces and for kings on neighbouring squares has similar weights. Then it...

hxim