ssc-hsr-pretrained

SSC-HSR Pretrained Auxiliary Checkpoints

This repository mirrors the pretrained auxiliary models used by SSC-HSR: Unaligned RGB Guided Hyperspectral Image Super-Resolution with Spatial-Spectral Concordance by Zhang et al.

The files were obtained from the pretrained-model release provided by the original authors. They include the RAFT, WarpNet, and feature-extraction checkpoints required by the public SSC-HSR implementation.

This Hugging Face mirror is maintained to provide stable and convenient programmatic access to these resources.

Files

The repository contains:

• raft-things.pth: base RAFT checkpoint;
• sf4/50000_raft-rgb-sf4.pth: scale-factor-4 RAFT checkpoint;
• sf8/50000_raft-rgb-sf8.pth: scale-factor-8 RAFT checkpoint;
• sf16/50000_raft-rgb-sf16.pth: scale-factor-16 RAFT checkpoint;
• feature_extraction.pth: feature-extraction checkpoint;
• sf8/warpnet-sf8.pth: Torchlight-free WarpNet checkpoint;
• sf8/warpnet-sf8.original-torchlight.pth: untouched original WarpNet checkpoint distributed by the authors.

All scientific credit for the models and original checkpoints goes to the original authors.

Torchlight-free WarpNet checkpoint

The original WarpNet checkpoint contains serialized training metadata referring to torchlight.trainer.engine.EngineConfig.

The WarpNet model tensors themselves do not depend on Torchlight, but this serialized metadata prevents the original checkpoint from being loaded with ordinary modern PyTorch when Torchlight is not installed.

For portability, this repository therefore provides two versions:

• warpnet-sf8.original-torchlight.pth: untouched original artifact;
• warpnet-sf8.pth: sanitized artifact used for normal loading.

The sanitized checkpoint preserves module.model exactly and removes the historical Torchlight serialization dependency.

The model-state keys, tensor shapes, tensor dtypes, and tensor values were verified against the original checkpoint. Every tensor was compared using torch.equal, and the sanitized checkpoint was also strict-loaded into the released WarpNet architecture.

Reproducing the sanitized checkpoint

The repository includes sanitize_torchlight_checkpoint.py, which reproduces the sanitized artifact from the untouched original:

python sanitize_torchlight_checkpoint.py \
    sf8/warpnet-sf8.original-torchlight.pth \
    sf8/warpnet-sf8.pth

The script verifies tensor-for-tensor equality after serialization.

Original work

• Paper: Unaligned RGB Guided Hyperspectral Image Super-Resolution with Spatial-Spectral Concordance, Zhang et al.
• Official implementation: https://github.com/yingkai-zhang/SSC-HSR
• Original pretrained models: released by the original authors.

This repository is a convenience mirror and additionally provides the Torchlight-free re-serialization described above. All scientific credit goes to the original authors.