nnue-interface

Python bindings for extracting Stockfish NNUE neural network activations and evaluations.

Features

Key Capabilities:

• Extract NNUE accumulator activations (hidden layer 0): 3072 dimensions (Big network) or 128 dimensions (Small network)
• Extract intermediate layer activations (layers 1-2): For deep network analysis
• Extract PSQT values: Piece-square table contributions
• Get final evaluations in centipawns
• Cross-platform: Works on Linux, macOS, and Windows
• Fast: Compiled C++ extension via pybind11
• ML-Ready: All outputs as float32 numpy arrays

Installation

From PyPI (recommended)

pip install nnue-interface

From Source

git clone https://github.com/yourusername/nnue-interface.git
cd nnue-interface
pip install -e .

Requirements:

• Python 3.8+
• C++17 compatible compiler (GCC, Clang, MSVC)
• CMake 3.15+
• NumPy 1.19+

Quick Start

import nnue_interface
import numpy as np

# Extract all NNUE activations and evaluation for a position
fen = "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1"

acc_white, acc_black, psqt, layer1, layer2, eval_final, eval_psqt = \
    nnue_interface.get_activations_and_eval(fen)

print(f"Evaluation: {eval_final:.2f} cp")
print(f"Accumulator shape: {acc_white.shape}")  # (3072,) for Big network
print(f"Layer 1 shape: {layer1.shape}")         # (30,) 
print(f"Layer 2 shape: {layer2.shape}")         # (32,)

API Reference

get_activations_and_eval(fen: str) -> tuple

Extract all NNUE activations and evaluation for a given position.

Parameters:

• fen (str): FEN notation of the chess position

Returns:

• acc_white (ndarray): White perspective accumulator, shape (3072,) or (128,)
• acc_black (ndarray): Black perspective accumulator, shape (3072,) or (128,)
• psqt (ndarray): PSQT values, shape (2, 8)
• layer1 (ndarray): First hidden layer activations, shape (30,) or (15×2)
• layer2 (ndarray): Second hidden layer activations, shape (32,)
• eval_final (float): Final evaluation in centipawns
• eval_psqt (float): PSQT-only evaluation in centipawns

get_evaluation(fen: str) -> float

Get only the final evaluation for a position (faster if you don't need activations).

Parameters:

• fen (str): FEN notation

Returns:

• Evaluation in centipawns (float)

get_network_info() -> dict

Get information about the NNUE network architecture.

Returns:

{
    'TransformedFeatureDimensionsBig': 3072,
    'TransformedFeatureDimensionsSmall': 128,
    'L2Big': 15,
    'L3Big': 32,
    'L2Small': 15,
    'L3Small': 32,
    'PSQTBuckets': 8,
}

Examples

Using Activations for Machine Learning

import nnue_interface
import numpy as np

# Collect training data
positions = [
    "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1",
    "rnbqkbnr/pppppppp/8/8/4P3/8/PPPP1PPP/RNBQKBNR b KQkq e3 0 1",
    # ... more positions
]

X_acc = []  # Accumulator features
X_layers = []  # Intermediate layer features
y = []  # Target evaluations

for fen in positions:
    acc_w, acc_b, psqt, layer1, layer2, eval_final, _ = \
        nnue_interface.get_activations_and_eval(fen)
    
    # Combine features
    features = np.concatenate([acc_w, acc_b, psqt.flatten(), layer1, layer2])
    
    X_layers.append(features)
    y.append(eval_final)

X = np.array(X_layers)
y = np.array(y)

# Now use X and y for ML training (scikit-learn, PyTorch, TensorFlow, etc.)

Analyzing Network Activations

import nnue_interface

fen = "r1bqkb1r/pppp1ppp/2n2n2/1B2p3/4P3/5N2/PPPP1PPP/RNBQK2R w KQkq - 4 4"

acc_w, acc_b, psqt, layer1, layer2, eval_final, _ = \
    nnue_interface.get_activations_and_eval(fen)

print(f"Position evaluation: {eval_final:.2f} cp")
print(f"\nAccumulator statistics (White):")
print(f"  Mean: {acc_w.mean():.2f}")
print(f"  Std:  {acc_w.std():.2f}")
print(f"  Min:  {acc_w.min():.2f}")
print(f"  Max:  {acc_w.max():.2f}")

print(f"\nLayer 1 sparsity: {(layer1 == 0).sum() / layer1.size * 100:.1f}%")
print(f"Layer 2 sparsity: {(layer2 == 0).sum() / layer2.size * 100:.1f}%")

Architecture

The NNUE network consists of:

1. Input Layer (Accumulator): 3072 or 128 dimensions

   • Efficiently updatable representation of board state
   • Updated incrementally as moves are made

2. Layer 1:

   • FC layer (sparse input) + SqrClippedReLU + ClippedReLU
   • Output: 15 dims × 2 (concatenated)

3. Layer 2:

   • FC layer + ClippedReLU
   • Output: 32 dims

4. Output Layer:

   • FC layer → single scalar evaluation
   • Also uses PSQT values for final output

All intermediate activations use int8/uint8 quantization for efficiency, converted to float32 for Python.

Performance

• Speed: ~50-200 µs per position (depending on CPU)
• Memory: ~2 MB for network weights
• No dependencies: Only NumPy required at runtime

Building from Source

Linux / macOS

git clone https://github.com/realrushil/nnue-interface.git
cd nnue-interface
pip install -e .

Windows (MSYS2 UCRT64)

git clone https://github.com/realrushil/nnue-interface.git
cd nnue-interface
pip install -e .

Build Documentation

CMake is used for cross-platform builds. To manually build:

cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build

Testing

pip install pytest numpy
pytest tests/

License

GPL-3.0-or-later (same as Stockfish)

Citation

If you use this in research, please cite Stockfish:

@software{stockfish,
  title = {Stockfish},
  url = {https://stockfishchess.org/},
  author = {Tord Romstad and Marco Costalba and Joona Kiiski and Gary Linscott},
}

Contributing

Contributions welcome! Please:

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Run tests: pytest tests/
5. Submit a pull request

Troubleshooting

Build fails on Linux with missing dependencies

# Ubuntu/Debian
sudo apt-get install build-essential cmake python3-dev

# Fedora
sudo dnf install gcc gcc-c++ cmake python3-devel

Import error on Windows

Make sure you're using the correct Python version (64-bit). The wheel must match your Python installation:

python -c "import struct; print('64-bit' if struct.calcsize('P') == 8 else '32-bit')"

Resources

• Stockfish Official
• NNUE Research
• Chess Programming Wiki - NNUE

Authors

• Created with Stockfish source code
• Python bindings by Rushil Saraf

---

Have questions? Open an issue on GitHub!