G-CNA - Chess Position Classifier

Neural network library from scratch applied to chess position classification.

Results

Model	Params	Best Accuracy	Time
Small	213,637	89.97%	392s
Medium	467,973	90.78%	781s
Large	566,533	92.08%	2366s

Dataset: 2.3M training (x4 augmented) / 550K test (x4 augmented)

Quick Start

# 1. Prepare dataset
python scripts/dataset/prepare.py

# 2. Augment (x4 with flip + color inversion)
python scripts/dataset/augment.py --mode both

# 3. Train
python scripts/training/train.py --arch large --epochs 20 --save model_large.nn

# 4. Demo
python demo/server.py  # Open http://localhost:8080

5-Class Classification

Class	Description
0	Nothing (normal position)
1	Check White
2	Check Black
3	Checkmate White
4	Checkmate Black

The Library: my_torch

PyTorch-like API built from scratch with NumPy:

from my_torch.nn import Linear, ReLU, Dropout, Softmax, CrossEntropyLoss
from my_torch.optim import Adam, ReduceLROnPlateau

Features:

• Tensor with gradient tracking
• Layers: Linear, Dropout, BatchNorm1d
• Activations: ReLU, LeakyReLU, Sigmoid, Softmax
• Loss: MSELoss, CrossEntropyLoss
• Optimizers: SGD (momentum), Adam
• LR Schedulers: StepLR, ExponentialLR, ReduceLROnPlateau
• Gradient Clipping: clip_grad_norm, clip_grad_value

See my_torch/README.md for full API.

Data Augmentation

python scripts/dataset/augment.py --mode both  # x4

Mode	Effect	Multiplier
flip	Horizontal mirror	x2
invert	180° rotation + color swap + label swap	x2
both	All combinations	x4

Model Architectures

Model	Architecture	Params
FenSmall	768→256→64→5	214k
FenMedium	768→512→128→64→5	468k
FenLarge	768→512→256→128→64→5	567k

Testing

pytest test/ -v                     # 140+ tests
pytest test/test_conformity.py -v   # PyTorch comparison (10 tests)

PyTorch Conformity Tests:

• Linear forward
• ReLU, Sigmoid, Softmax
• MSELoss, CrossEntropyLoss
• BatchNorm1d
• SGD, Adam optimizers

Project Structure

├── my_torch/          # Neural network library
├── scripts/
│   ├── dataset/       # prepare.py, augment.py
│   ├── training/      # train.py, search.py
│   └── benchmark/     # compare.py, visualize.py
├── demo/              # Web interface
├── test/              # 140+ unit tests
└── dataset/           # FEN data

---

Epitech Project G-CNA-500