.gitignore

@@ -1,2 +1,4 @@
 /target
 TODO.txt
+nnue/batches
+nnue/venv

nnue/.gitignore

@@ -1,4 +0,0 @@
-batches/
-venv/
-train_data/
-__pycache__/

nnue/README.md

@@ -6,19 +6,15 @@ The network is trained on both self-play games, and its games on Lichess.
 Both of these sources provide games in PGN format.
 
 This folder includes the following scripts:
-- `s1_batch_pgn_data.py`: Combine and convert big PGN files into small chunked files.
+- `batch_pgn_data.py`: Combine and convert big PGN files into small chunked files.
-- `s2_process_pgn_data.py`: Convert PGN data into a format suitable for training.
+- `process_pgn_data.py`: Convert PGN data into a format suitable for training.
 
 Example training pipeline:
 ```bash
 # chunk all the PGN files in `games/`. outputs by default to `batches/batch%d.pgn`.
-./s1_batch_pgn_data.py games/*.pgn
+./batch_pgn_data.py games/*.pgn
 
-# analyze batches to turn them into training data. outputs by default to train_data/batch%d.tsv.gz.
+# analyze batches 0 to 20 to turn them into training data. outputs by default to train_data/batch%d.tsv.gz.
 # set max-workers to the number of hardware threads / cores you have.
-# this is the longest part.
+./process_pgn_data.py --engine ../target/release/chess_inator --max-workers 8 batches/batch{0..20}.pgn
-./s2_process_pgn_data.py --engine ../target/release/chess_inator --max-workers 8 batches/batch*.pgn
-
-# combine all processed data into a single training set file.
-zcat train_data/*.tsv.gz | gzip > combined_training.tsv.gz
 ```

nnue/batch_pgn_data.py

@@ -18,9 +18,11 @@ import itertools
 
 from pathlib import Path
 
+"""Games to include per file in output."""
+
 parser = argparse.ArgumentParser()
 parser.add_argument("files", nargs="+", type=Path)
-parser.add_argument("--batch-size", type=int, help="Number of games to save in each output file. Set this to two to four times the amount of concurrent workers used in the processing step.", default=8)
+parser.add_argument("--batch-size", type=int, help="Number of games to save in each output file.", default=8)
 parser.add_argument("--output-folder", type=Path, help="Folder to save batched games in.", default=Path("batches"))
 args = parser.parse_args()
 

nnue/process_pgn_data.py

@@ -160,8 +160,6 @@ async def worker(game_generator: AsyncIterator[pgn.Game]) -> None:
 
             await output_queue.put((board.fen(), tensor, int(eval_abs), wdl))
 
-    await engine.quit()
-
 
 async def analyse_games(file: Path):
     """Task that manages reading PGNs and analyzing them."""
@@ -221,7 +219,6 @@ async def main():
 
         if skipped:
             logging.info("Resuming at file '%s'.", file)
-            skipped = False
         else:
             logging.info("Reading file '%s'.", file)
 

nnue/s3_train_neural_net.py

@@ -1,75 +0,0 @@
-#!/usr/bin/env python
-
-"""Train the NNUE weights."""
-
-import torch
-import pandas as pd
-import numpy as np
-
-from torch.utils.data import Dataset, DataLoader
-from pathlib import Path
-from dataclasses import dataclass
-
-
-################################
-################################
-## Data loading / parsing
-################################
-################################
-
-
-@dataclass
-class Position:
-    """Single board position."""
-
-    fen: str
-    """Normal board representation."""
-
-    board: torch.Tensor
-    """Multi-hot board representation."""
-
-    cp_eval: np.double
-    """Centipawn evaluation (white perspective)."""
-
-    expected_points: np.double
-    """
-    Points expected to be gained for white from the game, based on centipawn evaluation.
-
-    - 0: black win
-    - 0.5: draw
-    - 1: white win
-    """
-
-
-def sigmoid(x):
-    """Calculate sigmoid of `x`, using scaling constant `K`."""
-    K = 150
-    return 1 / (1 + np.exp(-K * x / 400))
-
-
-class ChessPositionDataset(Dataset):
-    def __init__(self, data_file: Path):
-        self.data = pd.read_csv(data_file, delimiter="\t")
-
-    def __len__(self):
-        return len(self.data)
-
-    def __getitem__(self, idx):
-        row = self.data.iloc[idx]
-
-        eval = np.double(row.iloc[2])
-
-        return Position(
-            fen=row.iloc[0],
-            board=torch.as_tensor([1 if c == "1" else 0 for c in row.iloc[1]]),
-            cp_eval=eval,
-            expected_points=sigmoid(eval/100),
-        )
-
-if __name__ == "__main__":
-    full_dataset = ChessPositionDataset(Path("combined_training.tsv.gz"))
-
-    train_dataset, test_dataset = torch.utils.data.random_split(full_dataset, [0.8, 0.2])
-
-    train_dataloader = DataLoader(train_dataset, batch_size=64, shuffle=True)
-    test_dataloader = DataLoader(test_dataset, batch_size=64, shuffle=True)

src/eval.rs

@@ -397,7 +397,7 @@ pub fn eval_metrics(board: &Board) -> EvalMetrics {
     let king_distance_eval =
         -advantage * i32::try_from(king_distance).unwrap() * max(7 - phase, 0) / 100;
 
-    let eval = (pst_eval + king_distance_eval).clamp(i16::MIN.into(), i16::MAX.into());
+    let eval = pst_eval + king_distance_eval;
 
     EvalMetrics {
         pst_eval,