hanyuyuan/STSF-TPLS-weights

STSF+TPLS — pretrained checkpoints

Pretrained weights for the paper "The Lift Spectrum: How Measurement-to-Space Adaptivity Shapes Robustness in Image-Free Single-Pixel Sensing" (arXiv:2607.22077).

STSF (spatiotemporal soft-fusion network) performs image-free segmentation directly from 1D single-pixel measurement sequences (512 coded measurements at a 3.13 % sampling rate for 128×128 scenes), trained under TPLS (task-prioritized loss scheduling). This repository holds the main-model checkpoints only — the STSF+TPLS arms behind the paper's headline and real-bench results. Training/evaluation code, configs, and all baselines live in the code repository: https://github.com/Hanyuyuan6/STSF-TPLS.

Checkpoints

All files are inference-only checkpoints/<run>/best.pth exports (about 147.6 MB each; 1.48 GB total). The full training config is embedded in each checkpoint under config, but optimizer and scheduler state are intentionally omitted, so these files cannot resume training. Per-seed clean test foreground mIoU (%) as reported in the paper's provenance chain:

File	Dataset / data protocol	Seed	Clean fg-mIoU	Role
checkpoints/rev_carvana_tpls_s42/best.pth	Carvana	42	74.57	headline
checkpoints/rev_carvana_tpls_m512_s43/best.pth	Carvana	43	73.55	headline
checkpoints/rev_carvana_tpls_m512_s44/best.pth	Carvana	44	74.93	headline
checkpoints/rev_mnist_tpls_s42/best.pth	MNIST	42	56.54	headline
checkpoints/rev_mnist_tpls_m512_s43/best.pth	MNIST	43	56.06	headline
checkpoints/rev_mnist_tpls_m512_s44/best.pth	MNIST	44	56.33	headline
checkpoints/rev_wbc_tpls_s42/best.pth	WBC — paper-legacy-v1 (231/58/60)	42	52.72	headline
checkpoints/rev_wbc_tpls_s43/best.pth	WBC — paper-legacy-v1 (231/58/60)	43	52.06	headline
checkpoints/rev_wbc_tpls_s44/best.pth	WBC — paper-legacy-v1 (231/58/60)	44	52.72	headline
checkpoints/rev_mnist_tpls_m512_s42/best.pth	MNIST	42	—	real-bench checkpoint (paper Figs. 8–9)

Three-seed means match the paper's headline: 74.4 / 56.3 / 52.5 foreground mIoU on Carvana / MNIST / WBC (STSF+TPLS vs. the SPIFS baseline: +5.6 / +9.9 / +3.2 pp). Exact byte counts and Git LFS SHA-256 object IDs are in MANIFEST.json. The legacy MD5 values from the original release metadata are retained there for compatibility; they were not recomputed during this documentation update.

Usage

Clone the code repository, then point its evaluation entry at a downloaded checkpoint — the loading path, model construction (GRUUNetPP), and metrics are all handled by scripts/evaluate.py there; the checkpoint's embedded config reproduces the exact architecture. Example download:

from huggingface_hub import hf_hub_download
ckpt_path = hf_hub_download(
    repo_id="hanyuyuan/STSF-TPLS-weights",
    filename="checkpoints/rev_carvana_tpls_s42/best.pth",
    revision="22a6e8ee71212ed4574b1a35a5c27e0681219dba",
)

Load release checkpoints in fail-closed mode:

import torch

ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=True)

If weights_only=True fails, stop and verify the immutable commit, SHA-256, code revision, and PyTorch environment. Do not automatically retry the same file with weights_only=False; that enables general pickle execution. No release tag exists in either repository; the example and MANIFEST.json therefore use the real immutable Hugging Face commit 22a6e8ee71212ed4574b1a35a5c27e0681219dba and the audited code commit de7a83fc5b2ae3aa292603671d316fd604b521f4.

Intended use and limitations

• Intended use: research reproduction and extension of the experiments described in arXiv:2607.22077, using the linked repository's evaluation code and the checkpoint's embedded config.
• Not for clinical use: the WBC checkpoint is a research artifact, not a diagnostic or clinical decision-making system. None of these checkpoints is intended for safety-critical deployment.
• Fixed acquisition assumptions: natural-order Sylvester Hadamard sampling, 128x128 scenes, and M = 512 measurements (3.13 % sampling). The input construction and normalization must match the linked code and embedded config.
• Out of scope without new validation: other sensing operators or mask orderings, image resolutions, measurement counts, preprocessing pipelines, datasets, domains, or out-of-distribution inputs. Performance outside the reported settings is unverified.
• Inference only: optimizer and scheduler state were stripped; these exports cannot resume training.

Scope and provenance

• Weights only. Datasets are not redistributed here; the code repository's README documents how to obtain Carvana, MNIST, and the WBC dataset and how splits are built.
• WBC paper protocol. All three WBC checkpoints and their reported metrics use paper-legacy-v1, the deterministic 231/58/60 released split. The code repository also provides a collision-safe 400-sample full-v2 mode for future experiments; no metric in this model card is attributed to that different protocol.
• The measurement model is fixed natural-order (Sylvester) Hadamard sampling at M = 512 (3.13 % for 128²); these checkpoints assume that acquisition.
• Code license: see the code repository. Weights license: CC-BY-4.0 (attribution: cite the paper below).

Citation

Preprint: https://arxiv.org/abs/2607.22077

@article{han2026liftspectrum,
  title         = {The Lift Spectrum: How Measurement-to-Space Adaptivity Shapes Robustness
                   in Image-Free Single-Pixel Sensing},
  author        = {Han, Yuyuan and Li, Jingwei and Qiu, Long and Wang, Chong and Hao, Wenxuan
                   and Han, Jiangyu and Yao, Xinyu and He, Yuchen and Chen, Hui and Liu, Jianbin
                   and Zheng, Huaibin},
  journal       = {arXiv preprint arXiv:2607.22077},
  year          = {2026},
  eprint        = {2607.22077},
  archivePrefix = {arXiv},
  primaryClass  = {eess.IV},
}