WiSER 1.1 — joint100 RadioMap and CIR

This release provides the selector-chosen E070 endpoint of the corrected 100-scene, 100-epoch joint RadioMap/CIR run. The complete run performed 25,400 optimizer updates; its registered scene-disjoint selector chose epoch 70.

Code and executable examples use corrected metric-world TX/RX coordinates, 10 cm spatial CIR consolidation, and continuous-delay path targets. RadioMap uses delta-distance query features; CIR uses the full-geometry additive conditioner. Both heads share a jointly trained scene encoder.

Checkpoint identity

Field	Value
Scientific experiment	REVSHARED-EADVT82JOINT1001, v01, seed s0
Endpoint	main/SELECTED_P00_E070__REVSHARED-EADVT82JOINT1001__v01__run002__s0.pt
Bytes	642003603
SHA-256	37701c0ae1aed035305a8ae62bfd7a43caf389784172e69314274e898a6d3542
Full run	100 scenes, 100 epochs, 25,400 updates
Selection	Epoch 70, registered selector; no test-driven selection
Serialization	PyTorch payload with model and config.candidate

Use the public scripts/infer_example.py or wiser.inference.load_joint_checkpoint for strict model construction and loading. The endpoint is for inference; it has no full optimizer/RNG state and cannot resume training. Load only a trusted checkpoint and verify the SHA above.

Quick start

In the independent WiSER Conda environment described in the code repository:

hf download Jingqiao-ucsc/WiSER \
  main/SELECTED_P00_E070__REVSHARED-EADVT82JOINT1001__v01__run002__s0.pt \
  --local-dir /data/wiser-model
python scripts/infer_example.py --example-root example \
  --checkpoint /data/wiser-model/main/SELECTED_P00_E070__REVSHARED-EADVT82JOINT1001__v01__run002__s0.pt \
  --expected-sha256 37701c0ae1aed035305a8ae62bfd7a43caf389784172e69314274e898a6d3542 \
  --device cuda:0 --out-json outputs/example_predictions.json

The tested inference/training runtime is Python 3.12, PyTorch 2.10.0+cu128, xformers 0.0.35, with the none sparse-convolution backend. The public code contains real joint100 training, exact train/selector input preparation, and full-state resume support for checkpoints produced by that training entrypoint.

Data and reproducibility

The original processed source dataset is unchanged. The public preparation command derives spatially consolidated CIR records from raw multipath H5 files, retains continuous delays, verifies TX geometry, and records source hashes. It does not supply original ScanNet++ RGB/depth/raw-mesh assets. Training preparation uses train/selector identities only; the bundled compact example is development evidence, not a benchmark.

This current main contains the joint100 endpoint and provenance/config files. Earlier P9 and ablation artifacts are historical releases; they were preserved before this ordinary main commit and are not interchangeable with this endpoint. No repository history was rewritten. Model/data code and metric contracts are linked in the pinned GitHub release; no new paper metrics are claimed here.