Variant NNUE training data generator

This is a variant NNUE training data generator based on Fairy-Stockfish to generate data for the variant NNUE trainer. See the wiki of the training code for more information.

This is a fork of variant-nnue-tools

This fork would only work for the xiangqi variant.

To compile

    cd src
    sh m.sh

To run

    time ./stockfish

This would generate 10000000 (10 millions) positions.

This may take some time. For example, it took 43 minutes to complete in my laptop.

Result is stored in xiangqi.bin

The file size is 720000000 (arround 720Mb). Each position occupies 72 bytes.

If you run the program again, make sure to remove the existing xiangqi.bin first

Default Settings

Inside the function getline() in uci.cpp

    void getline(string &cmd) {
      const int size=9;
      static string list[size]={
       "", 
       "ucci",
       "setoption Use_NNUE pure",
       "setoption EvalFile xiangqi-aa162e1771e5.nnue",
       "setoption Threads 8",
       "setoption Hash 512",
       "setoption UCI_Variant xiangqi",
       "generate_training_data depth 2 count 10000000"
       " random_multi_pv 4 random_multi_pv_diff 100"
       " random_move_count 8 random_move_max_ply 20"
       " write_min_ply 5 eval_limit 10000"
       " set_recommended_uci_options data_format bin"
       " output_file_name xiangqi.bin",
       "quit"
      };

For example, to change the number of generated positions to 5 million, modify the above function:

       "generate_training_data depth 2 count 5000000"

To use another neural network, modify this line:

       "setoption EvalFile xiangqi-aa162e1771e5.nnue",

The default settings works well as a quick start for training nnue network.

To actually train the network, run the trainer

The trainer is a fork of variant-nnue-pytorch.

The fork can be used to train xiangqi network with the data generated above (xiangqi.bin)

time python train.py --gpus 1 --max_epochs 10 xiangqi.bin xiangqi.bin

The trainer took 90 minutes to complete in my laptop

Setting up the trainer with all dependencies is a complex task for beginners, and it require a GPU capable of running CUDA.

If you have difficulties setting up a local training environment, you may follow these instructions to run a trainer online