Read the paper (NeurIPS 2026 Interp4Discovery Workshop)
All code needed for replication is available in the dpnt/ folder. CFMT, CFMT-AUS, and CFMT+ are not publically available datasets and must be requested access from the respective authors.
results/results_cfmt.csv holds the results of the 53 models for the CFMT. Most models work out of box, but a few need to be downloaded (info below)
Most models download automatically. Six need a checkpoint downloaded once by hand; the sweep skips them if no path is given.
Put the files anywhere (e.g. checkpoints/) and pass their paths via model_checkpoints:
local_model_checkpoints = {
"arcface_r50_webface600k": "checkpoints/w600k_r50.onnx",
"arcface_r50_ms1mv3": "checkpoints/ms1mv3_arcface_r50/backbone.pth",
"resnet50_vggface2_official": "checkpoints/resnet50_ft_weight.pkl",
"resnet50_robust_l2_eps3": "checkpoints/resnet50_l2_eps3.ckpt",
"resnet50_toponet_tau30": "checkpoints/",
"resnet18_toponets_tau10": "checkpoints/",
}| Model | Registry name | Download |
|---|---|---|
| ArcFace R50, WebFace600K | arcface_r50_webface600k |
InsightFace |
| ArcFace R50, MS1MV3 | arcface_r50_ms1mv3 |
arcface_torch model zoo |
| VGGFace2 ResNet-50 | resnet50_vggface2_official |
VGGFace2-pytorch |
| Robust ResNet-50, L2 ε=3 | resnet50_robust_l2_eps3 |
Madry Lab robust ImageNet models |
for toponets, function auto-downloads. just do
local_model_checkpoints[resnet50_toponet_tau30] = "checkpoints/resnet18_tau_30.pt"
local_model_checkpoints[resnet18_toponets_tau10] = "checkpoints/resnet18_tau_10.pt"however you will need to pip install topoloss to run these models:
pip install git+https://github.com/toponets/toponets.git