ChessCV Lichess FEN Square Classifier
This repository contains an MLX chess-board square classifier for FEN recognition from Lichess-style 2D board screenshots.
The model classifies each 64×64 RGB/BGR-normalized board-square crop into one of 13 classes: empty square plus the 12 standard chess pieces. A full FEN can be produced by running the classifier over all 64 square crops in board order and converting the predicted class sequence into piece-placement notation.
Model file
mlx_model_lichess_fen_64.safetensors— MLX weights forChessPieceCNN64mlx_model.py— MLX model definitionslabels.json— class-index mappingtraining_config.json— high-level training metadata
Labels
0: empty
1: white pawn
2: white rook
3: white knight
4: white bishop
5: white queen
6: white king
7: black pawn
8: black rook
9: black knight
10: black bishop
11: black queen
12: black king
Input format
- Shape:
(N, 64, 64, 3) - Value range: float32 normalized to
[0, 1] - Channel order: use the same channel order as your OpenCV crop pipeline. This model was trained from OpenCV-rendered crops.
For board-level FEN recognition, crop the detected board into 64 squares, resize each crop to 64×64, run batch inference, then map argmax outputs to FEN piece-placement symbols. For black-side board orientation, reverse the 64 square predictions before converting to canonical FEN.
Example usage
import mlx.core as mx
import numpy as np
from mlx_model import ChessPieceCNN64
model = ChessPieceCNN64()
model.load_weights("mlx_model_lichess_fen_64.safetensors")
# batch: np.ndarray shaped (64, 64, 64, 3), float32 in [0, 1]
logits = model(mx.array(batch))
mx.eval(logits)
labels = np.argmax(np.array(logits), axis=1)
Training and verification summary
The model was trained on synthetic square crops generated from Lichess default cburnett piece assets, with Lichess-like board colors, empty squares, scale changes, compression artifacts, blur, lighting/gamma variation, and small-screen resampling effects.
Final training run:
- Input size: 64×64
- Samples: 240,000
- Epochs: 14
- Empty-square samples: ~38%
- Final training accuracy: 99.98%
Rendered Lichess-board verification:
- Boards are rendered from Lichess default pieces
- Boards are split into 64 independent square crops
- Both normal and black-side-flipped orientations are tested
- Verification result: 15,872 / 15,872 squares correct across 248 board trials
Intended use
This model is intended as a component in chess computer-vision pipelines that detect a 2D chessboard on screen and reconstruct the piece-placement part of a FEN string.
License and source assets
The model was trained using Lichess piece assets from the open-source Lichess project. Lichess is distributed under AGPL-3.0-or-later; individual piece sets may have their own original author/license history upstream.
Quantized