b-mae-pretrain

Masked Autoencoder (MAE)

Masks half the image patches and reconstructs them on real handwritten digits — the MAE / VideoMAE pretraining objective; reports held-out reconstruction error.

Trained from scratch in Ropedia Academy — an interactive, bilingual course on embodied & spatial AI. Educational model: small and quick to train; the value is the method and a reproducible pipeline, not a leaderboard score. Try it live in the Ropedia demos Space.

At a glance

Dataset

• Name: Handwritten digits (UCI / scikit-learn)
• Type: real — public dataset
• Size / stats: 1,797 real 8×8 grayscale digit images; 16 patches of 2×2, 50% masked; 128/batch
• Split: 1,257 train / 540 test (held-out reconstruction)
• Source: scikit-learn load_digits (UCI Optical Recognition of Handwritten Digits)

Training config

Adam (lr 1e-3), 1200 steps, batch 128; 16 patches of 2×2, 50% masked; 2-layer transformer enc/dec.

Evaluation results

Robustness (mean ± std over 5 seeds)

Single-run numbers above are one seed; this is the distribution over independent re-trains (honest variance — no cherry-picking).

Inference example

import torch
state = torch.load("mae.pt", map_location="cpu")   # this repo's checkpoint
# Rebuild the exact module from the lab notebook (see "Reproduce"), then:
# model.load_state_dict(state); model.eval()

Limitations

Educational scale. Trained quickly on CPU on small or synthetic data, so absolute numbers are not competitive with production systems — the value is the method and a reproducible pipeline. No large-scale data, no hyperparameter sweep, and no multi-seed variance is reported. Not for production use.

Tiny model on 8×8 digits → blurry reconstructions; not a general visual encoder.

Failure cases

A high mask ratio with only 16 patches removes too much context → over-smoothed fills.

Reproduce / train your own

One click: open the notebook in Colab → Runtime → GPU → Run all, then run its Publish to the Hugging Face Hub cell.

From a shell:

git clone https://github.com/ChaoYue0307/ropedia-academy.git && cd ropedia-academy
pip install torch numpy matplotlib scikit-learn scikit-image gymnasium
jupyter nbconvert --to notebook --execute notebooks/training/B_mae_pretrain.ipynb --output run.ipynb
# optional: override training length, e.g.  STEPS=2000  (or EPISODES=600)  before running

Files

• figure.png
• mae.pt
• metrics.json
• seeds.png

License

Code & weights: MIT (this repository) — educational use encouraged.
Handwritten-digits data: UCI ML Repository via scikit-learn — CC BY 4.0.

Citation

If you use this model or the course materials, please cite:

@misc{ropedia_academy,
  title  = {Ropedia Academy: an interactive course on embodied & spatial AI},
  author = {Ropedia Academy},
  year   = {2026},
  howpublished = {\url{https://chaoyue0307.github.io/ropedia-academy/}}
}

Method / original work: He et al., Masked Autoencoders Are Scalable Vision Learners (MAE), CVPR 2022; Tong et al., VideoMAE, NeurIPS 2022.

Related assets

• 🚀 Live demos: https://huggingface.co/spaces/cy0307/ropedia-demos
• 🤗 All trained models + collection: https://huggingface.co/cy0307
• 📚 Course & all labs: https://chaoyue0307.github.io/ropedia-academy/ · Labs tab
• 💻 Source / notebooks: github.com/ChaoYue0307/ropedia-academy

---

Part of the Ropedia Academy trained-model collection. Contributions & issues welcome on GitHub.