MoME: Forget the Data, Keep the Capability

Weights for MoME, which splits one re-identification backbone into core parameters that preserve retrieval and zero-initialized parasitic parameters that carry the full-data knowledge, trained under gradient isolation so the core never receives a forget-set gradient. One checkpoint is read two ways: with the branch off it satisfies a request to delete a person's training data, with the branch on it restores what that data taught.

Paper: https://arxiv.org/abs/2610.36858 · Code: https://github.com/Jingong-Chen/MoME

Files

File	What it is	mAP	R-1
market/mome_core.pth	MoME on Market-1501, read with --mode core	74.3	88.2
market/mome_full.pth	the consolidated payload for Market-1501	88.7	95.2
market/retrain_s42.pth	the retrained reference the tables compare against	70.5	86.9
market/original.pth	trained on everything, no deletion	88.0	94.8
veri/mome_core.pth	MoME on VeRi-776, read with --mode core	42.8	69.6
veri/original.pth	trained on everything, no deletion	82.3	97.1

All are ViT-B/16 at 256 × 128, initialized from the ImageNet ViT the TransReID recipe uses. The deletion request is identity-level: split seed 42, half the identities deleted, four images kept per retained identity. Each file reads its own number rather than the five-split mean the paper's tables report.

The VeRi-776 payload is not here as a file. It is thirty epochs from veri/mome_core.pth and veri/original.pth with scripts/full_fit.py, which is how the one in the paper was made; the command is in the repository README.

Use

from utils.parasitic_utils import set_parasitic_mode

set_parasitic_mode(model, False)   # Core: the deletion request is honored
set_parasitic_mode(model, True)    # Full: the full-data model is back

python eval_partition.py --config_file configs/Market/vit_transreid_stride_mome_v3.yml \
  --weight market/mome_core.pth --mode core

Citation

@article{chen2026mome,
  title   = {Forget the Data, Keep the Capability: Switchable Unlearning for Retrieval},
  author  = {Chen, Jingong and Gao, Kyle and Li, Guangyuan and Pack, Chulwoo and
             Liu, Sijia and Won, Kwanghee},
  journal = {arXiv preprint arXiv:2610.36858},
  year    = {2026}
}

Licence

MIT, as the code is. The datasets these were trained on are not redistributed and keep the terms of their own releases.