vision-foundation-mae

Production-minded, config-driven Vision Transformer + Masked Autoencoder pipeline for self-supervised pretraining and downstream transfer.

Features

• MAE pretraining on open datasets (CIFAR-10 default; STL-10 optional)
• Linear probe and full finetuning on CIFAR-10
• Encoder export and folder inference CLI
• Hydra config system with CLI overrides
• JSONL + TensorBoard logging, optional W&B
• Reproducibility manifest (resolved config, env, git hash)
• Pytest coverage for key invariants

Quickstart

pip install -e .
python scripts/train_mae.py
python scripts/eval_linear_probe.py pretrained_ckpt=outputs/latest/checkpoints/last.pt
python scripts/train_finetune.py pretrained_ckpt=outputs/latest/checkpoints/last.pt
python scripts/export_encoder.py pretrained_ckpt=outputs/latest/checkpoints/last.pt
python scripts/infer.py classifier_ckpt=outputs/latest/checkpoints/best.pt image_dir=./some_images

Structure

See configs/, src/minimae_vit/, scripts/, and tests/.

Notes

• Default image size is 64x64 (CIFAR-10 upsampled from 32x32).
• MAE reconstruction visualization uses pixel-space unnormalization from stored per-patch statistics.