WiSER checkpoints

This repository contains the revised WiSER release checkpoint and the paper ablation endpoints. The matching public code is the canonical main branch at merge commit 1105ae03caf3757d43002b4e28b48232e2952669.

Main checkpoint

main/wiser_voxel_consolidated_small100_p9.pt is the promoted 100-scene, seed-0, Pretrain + Alternating checkpoint (R1Q9-ED8.2/v02/run001/s0). It contains the shared backbone, RadioMap head, and CIR head.

The revised CIR target applies 10 cm spatial bounce-voxel consolidation before learning. Stored path delays remain continuous; this release does not use a 128-bin temporal target. The default eight CIR queries are output slots, not fixed delay-bin classes.

SHA-256: c99c81dbbe14435f58847a6d42d77d2a592d1e19ae677cd65437b6ffa489d934.

Ablation checkpoints

The ablations/ tree contains one selected checkpoint for each of the 12 unique 50-scene, seed-0 endpoints used by the revised ablation suite. The training-scheme, RadioMap-component, and CIR-component families are kept in separate folders. joint_from_scratch also contains its epoch-100 terminal checkpoint because its selected epoch-20 checkpoint and completed training endpoint intentionally answer different questions.

These ablation checkpoints are controlled reduced-scale comparisons. They are not substitutes for the 100-scene main release checkpoint.

See checkpoint_index.json for experiment IDs, file sizes, hashes, and model roles. Each endpoint folder also has a short README.

Evaluation scope

main/eval_summary.json records two types of evidence:

• the original release-candidate validation: 80 RadioMap records and 5,000 CIR queries from 10 scene-disjoint validation scenes;
• the later balanced protocol: two 200-map RadioMap sets and three 5,000-query CIR sets.

The scene-disjoint scenes participated in historical checkpoint selection, so they are matched validation sets, not a sealed blind test.

Use with the public code

git clone https://github.com/scp10086/WiSER.git
cd WiSER
python scripts/infer_example.py \
  --example-root example \
  --checkpoint /path/to/main/wiser_voxel_consolidated_small100_p9.pt \
  --out-json outputs/example_summary.json

The bundled example scene is a seen/training-scene smoke example. It is not an unseen-scene benchmark and should not be used to reproduce paper metrics.

Dataset note

The linked Hugging Face dataset page is the companion Sionna/ScanNet++ release. The checkpoint in this repository was trained with the revised 10 cm voxel-consolidated derivative. Until that processed derivative is separately published, use the public dataset_gen/ pipeline and preserve the split and manifest identities described in the code repository.

Historical checkpoint

The former top-level wiser_sparse_scene_encoder_small100_full.pt represented the earlier 128-bin-era public code/data contract. It was removed from the current tree after this revised payload was uploaded and verified. Hugging Face Git history retains the old artifact for reproducibility.