Release validation

Validated on 2026-10-27 (Asia/Shanghai).

Completed checks

10 regression tests passed: source-time interpolation, pose-gap handling, clip padding, and ordered two-clip history.
Three real EgoTouch ego-video excerpts, each 50 source frames / 25 sampled frames / four model clips, completed WiLoR → student inference → MP4 rendering.
A fold-paper excerpt detected the left hand in 21/25 sampled frames and the right hand in 20/25. Two additional gloved/occluded excerpts had very low detection coverage and exercised missing-pose behavior.
The student checkpoint loaded with strict key/shape matching. Pressure and contact outputs have shape (50, 2, 21, 21), are finite, and lie in [0, 1].
Thresholded contact excludes invalid anatomical cells. Each rendered MP4 decodes to 50 frames.
Tested on one NVIDIA H200; peak allocated GPU memory was 10,258,576,896 bytes (about 9.55 GiB). This is not a verified minimum VRAM requirement for all videos or GPU architectures.
PyTorch 2.7.1+cu126, torchvision 0.22.1, Transformers 5.3.0, Python 3.10; SDPA/BF16.
The tested WiLoR Python source exactly matches upstream commit fcb911312a38fa8badd30d9656a167485d61b8f9. Tested WiLoR weights match the pinned official HF repository by SHA256.
Source files passed Python compilation checks and were scanned for private machine paths.

See validation_results.json for per-case detection coverage, value ranges and memory measurements. Inputs and prediction videos are retained in the maintainer's private staging area; raw dataset clips are not included in the public source archive.

Scope

Testing used an isolated virtual environment with the existing server PyTorch stack plus separately installed WiLoR dependencies. A completely empty-machine installation has not been tested. Dependency downloads on this server required an offline wheel transfer because its mirror returned errors; these are infrastructure issues, not a required runtime step for users with normal PyPI access.

This verifies executable inference and output contracts, not all paper metrics, model calibration, or accuracy on arbitrary user videos. Pressure is relative to training sensors. Severe occlusion, gloves, multiple people and missed WiLoR detections remain limitations.

The trained student, matching configuration and SHA256 manifest are versioned at https://huggingface.co/RammusLeo/TouchAnnotator, tag v1.0. The source bundle, model card and inference guide are available on the repository's main branch. The anonymous downloader obtains matching model.pth, config.yaml and manifest.json and verifies the checkpoint SHA256.

Run the regression tests

python -m pip install pytest
python -m pytest tests -q

Optional 0813 variant — 2026-11-02

16 regression tests and 2 subtests passed, including variant routing and overlap-add coverage of short/tail-aligned videos.
Exported 787 state entries from saved epoch 29; every tensor equals its source tensor with the original dtype retained.
A 50-source-frame / 25-sampled-frame bimanual video completed WiLoR, strict legacy weight loading, six overlapping model windows and 50 rendered video frames. Left/right detections: 21/25 and 20/25.
Relative pressure is finite and lies in [0,1]. Contact is pressure ≥ 0.1 on valid cells. No contact-probability arrays are emitted for this checkpoint.
Peak allocated GPU memory was 3,034,049,536 bytes on an H200, not an established minimum for other hardware or inputs.
A separate 16-frame 0824 smoke test preserved finite pressure and independent contact-probability fields.
Public machine-readable evidence: validation_0813.json. This is execution/schema validation; a completely empty-machine installation and a controlled accuracy comparison between variants have not been performed.