WriteRHand-MNIST

Four handwritten-digit classifiers whose mathematical core was implemented from scratch (NumPy; PyTorch only as a tensor library for the MLP, no autograd) for a university Pattern Recognition course. They are educational models, not production OCR. Code, tests and report: https://github.com/TranTrinhNhan1/WriteRHand. Live demo (runs the models in the browser): see the repository README.

• Gaussian Naive Bayes - per-class, per-pixel Gaussian likelihoods with MLE parameters and log-space inference.
• Softmax regression - multiclass logistic regression, softmax + cross-entropy, mini-batch gradient descent, L2.
• Multilayer perceptron - fully connected ReLU network, hand-written backpropagation, momentum SGD, early stopping.
• AdaBoost (SAMME) - multiclass AdaBoost (SAMME) over weighted histogram decision trees.

Intended use and limitations

• Recognise a single, centred, 28x28 handwritten digit (white on black), MNIST style.
• Not trained on other scripts, multiple digits, noisy photographs or non-centred input; the demo therefore crops and centres drawings with the MNIST recipe, but robustness is limited.
• Accuracy numbers below are for MNIST only. Naive Bayes is weak by design (independence assumption). AdaBoost scores are weighted vote shares, not calibrated probabilities.

Data and preprocessing (mnist-global-standardize-v2-stratified)

• MNIST: 60 000 training + 10 000 test images. Training set split with a seeded stratified split into 50000 train / 10000 validation (split fingerprint 52cc8206ff96f56d, seed 42); the official 10 000 test images are used only for final reporting.
• Pixels / 255, then x = (p - mean) / (std + eps) with mean = 0.130635, std = 0.308079, eps = 1e-8 (computed on the training portion only). The ONNX graphs contain this step: feed raw pixels in [0, 1], shape [N, 784], read probabilities [N, 10].

Results (measured, CPU)

Hyper-parameters were chosen on the validation split only. Time columns: wall-clock fit time on the full 50 000-image train split and median latency of a single-image predict (CPU: 13th Gen Intel(R) Core(TM) i7-13620H, 16 logical cores, PyTorch threads: 8).

Model	Train acc %	Val acc %	Test acc %	Test macro-F1 %	Fit (s)	1-image latency (ms)
Gaussian NB	80.13	79.67	81.32	81.16	0.4	0.076
Softmax regression	93.62	92.15	92.52	92.42	12.7	0.010
MLP	99.87	98.18	98.24	98.23	11.0	0.042
AdaBoost-SAMME	100.00	97.59	97.60	97.58	817.9	18.705

Files

file	format	size	exported vs Python: argmax agreement / max prob. diff	test images checked
models/naive_bayes.onnx	onnx	62 KiB	100.00% / 1.8e-04	10,000
models/softmax.onnx	onnx	31 KiB	100.00% / 3.4e-06	10,000
models/mlp.onnx	onnx	2094 KiB	100.00% / 1.9e-06	10,000
models/adaboost.json	json	15607 KiB	100.00% / 0.0e+00	1,000

python/*.npz hold the original weights (BaseClassifier.load), test_vectors.json raw test digits with expected outputs, results.json all benchmark numbers used by the website.

Training procedure

See the repository: uv run python -m src prepare | tune | train | baselines | export | upload. Experiment tracking: Weights & Biases (offline fallback). Seeds are fixed; runs are reproducible on the same hardware/library versions (git commit fa51c49c8079469287591cbd9b6e9f2186d1ee63).

AI-assisted development

Parts of this code, the tests, the export pipeline and the website were produced with the help of an AI coding assistant (Claude Code) and reviewed/run by the project authors, as disclosed in the course report.

License

No license has been chosen by the authors yet (educational project). MNIST is provided by Y. LeCun, C. Cortes and C. J. C. Burges.

Citation

LeCun, Y., Bottou, L., Bengio, Y., Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324.