Pampaj's picture
Upload dn_mixed_topology_v1 top checkpoint
c8e42d7 verified
|
Raw
History Blame Contribute Delete
2.07 kB
---
license: other
tags:
- 3d
- mesh
- face-recognition
- diffusionnet
- topology-robustness
- pytorch
library_name: pytorch
---
# WBES FaceEmbedding DN Mixed Topology v1
This model repository contains the current top WBES-FaceEmbedding submission checkpoint for the NeurIPS 2026 research codebase.
## Checkpoint
- Model branch: `dn_mixed_topology_v1`
- Run: `mixed_xtopo_rank0p5_id0p25_bs5_best`
- Selected checkpoint: `best_by_xtopo_mesh_clean.pth`
- Local source path:
`face_embedding/gt_encdec/remeshing/intrinsic/newdata/dn_mixed_topology_v1/mixed_xtopo_rank0p5_id0p25_bs5_best/checkpoints/best_by_xtopo_mesh_clean.pth`
## Included Files
- `best_by_xtopo_mesh_clean.pth`: selected PyTorch checkpoint.
- `config.json`: training and architecture configuration copied from the run directory.
- `best_by_xtopo_mesh_clean.txt`: checkpoint-selection marker.
- `train_log.csv`: training log.
- `xtopo_mesh_log.csv`: cross-topology mesh validation log.
## Selection Marker
The stored selection marker reports:
- `best_epoch=82`
- `best_xtopo_mesh_clean=0.7483693322699105`
The same run also stores:
- `best_spearman_clean=0.8234182929370094`
- `best_auc_r=1.0041532552751473`
## Data Relationship
This checkpoint was trained against operator-enriched REMESH data in the local research workspace. The public mesh-only/no-operator dataset release is separate:
- Dataset: `Pampaj/wbes-faceembedding-noops`
- URL: <https://hf.co/datasets/Pampaj/wbes-faceembedding-noops>
That public dataset contains mesh-only `.npz` archives and does not include intrinsic operators, checkpoints, or heavy pair-level evaluation dumps.
## Integrity
- `best_by_xtopo_mesh_clean.pth` SHA256:
`0b1578a55e86ea5ed349cc59ae7921a24f5325d45e427f0582ef36f5d1088e62`
- `config.json` SHA256:
`61f428d0c6f80cb8c6af75cff36c608b8946617b90c81dc1f4222cf4b9b4cbd2`
## Intended Use
This checkpoint is intended for reproducing the topology-robust face embedding experiments in the WBES-FaceEmbedding repository, especially the FaceBench and FaceVerse evaluation paths documented there.