Skip to content

Latest commit

 

History

History
69 lines (45 loc) · 2.72 KB

README.md

File metadata and controls

69 lines (45 loc) · 2.72 KB

fast-mae-lite = fml

A lightweight MAE-Lite port/implementation, compatible with Python 3.11 and torch 2+.

Written for my own use; released in case it's useful to others.

For now, this repository only supports the tiny variant, but supporting others should be straightforward. I just haven't needed them so far.

Why?

I needed:

  • To run and fine-tune a pretrained MAE-Lite under modern Python and Torch versions,
  • To recover color. In the original pretrained models, there is information loss due to patch-local normalization. The training code in this repository fine-tunes a pretrained model to recover color (and also predict patches that are seen as input - a rather trivial operation, but convenient for some applications). Few gradient descent steps are needed to do so.

This is released as-is, it is likely that there will be breakages, incompatibilities, etc. Please open an issue - or better yet, a pull request - if you encounter any.

Is it fast?

Not terrible, not insanely fast yet.

Current peak speed on a laptop with 16GB RAM, RTX 4060M (8GB VRAM), and a Ryzen 9 8945H CPU is about 2k img/s, seems to be be bottlenecked by data loading / preprocessing.

Will see how fast we can get this to train on ImageNet on a single GPU.

Prerequisites

  • Ensure you have uv installed.
  • For inference/fine-tuning from pretrained: download the original MAE-Tiny checkpoint to ckpt/mae_tiny_400e.pth.tar.
  • To try out fine-tuning on color recovery: download imagenette to ~/datasets/imagenette (or elsewhere, in which case, override the corresponding path in the config).

Test inference from pretrained

uv run pytest --verbose

The image that will be saved should look odd, but have recognizable and meaningful shapes. The odd appearance is because the original pretrained MAE-Lite was trained to predict locally-normalized patches, and the loss was masked out of patches that are fed as input (since predicting them is trivial).

Single training run

uv run -m train.main

Recommendations:

  • Set compile=false for quick iteration.
  • Be mindful of available system RAM when setting workers.
  • Likely to be CPU- or disk-bound, not GPU-bound.
  • You probably don't need warmup for this task.

Hyperparameter sweep

uv run python -m train.main --config-name sweep --multirun

Attribution

This project contains code derived from the original MAE-Lite repo, licensed under the Apache License 2.0. We are grateful to the authors for their work, and for releasing the pretrained checkpoints that this repository is designed to work with.