peal-xai-rae

Non-commercial use only (CC BY-NC 4.0).

The optional companion of peal-xai that enables PEAL's RAE generators (RAEDiffusionAutoencoder, the ImageNet and CelebA representation-autoencoder counterfactual generators). It contains a modified copy of RAEv2 by Singh, Zheng, Wu, Zhang, Shechtman and Xie, which is licensed CC BY-NC 4.0, so this package is too. peal-xai itself is LGPL-3.0-or-later and does not need this package for anything else.

Install

pip install "peal-xai[rae]"                  # from PyPI
pip install ./packages/peal-xai-rae          # from a clone of the PEAL repository
# the clip:ViT-L/14 encoder needs OpenAI CLIP, which is not on PyPI:
pip install "clip @ git+https://github.com/openai/CLIP.git"
# only for retraining stage 1 / stage 2:
pip install "peal-xai-rae[train]"

Weights

The pretrained ImageNet weights (decoder.pt 1.66 GB, stats.pt, stage2_ema.pt 4.96 GB) are too large for a wheel. They live at https://huggingface.co/sidney1505/peal-rae-clip-imagenet (also CC BY-NC 4.0) and are downloaded on first use by any generator config with weights: hf://sidney1505/peal-rae-clip-imagenet, for example configs/web_demo/imagenet_rae_clip_hf.yaml. They are cached under $PEAL_HF_WEIGHTS_DIR, default $PEAL_RUNS/hf_weights.

How PEAL finds it

PEAL looks for RAEv2 in this order: $PEAL_RAEV2_DIR, this installed package (peal_rae.raev2_dir()), the in-repository copy under packages/peal-xai-rae/peal_rae/RAEv2, and finally external/RAEv2.

Licence

LICENSE (CC BY-NC 4.0 with the RAEv2 copyright notice). What was changed relative to upstream is listed in peal_rae/RAEv2/MODIFICATIONS.md.