ssc-hsr-checkpoints

SSC-HSR Released Checkpoints

This repository mirrors the released checkpoints for SSC-HSR: Unaligned RGB Guided Hyperspectral Image Super-Resolution with Spatial-Spectral Concordance by Zhang et al.

The original checkpoints were obtained from the release provided by the original authors. This Hugging Face repository is maintained to facilitate stable and convenient programmatic downloading.

Released checkpoints are provided for scale factors 4, 8, and 16.

Checkpoints

sf4/
  config_example.yaml
  ckpt/
    model-best.pth
    model-best.original-torchlight.pth

sf8/
  config_example.yaml
  ckpt/
    model-best.pth
    model-best.original-torchlight.pth

sf16/
  config_example.yaml
  ckpt/
    model-best.pth
    model-best.original-torchlight.pth

For each scale factor:

• model-best.original-torchlight.pth is the untouched checkpoint distributed by the original authors;
• model-best.pth is a Torchlight-free re-serialization intended for ordinary modern PyTorch loading.

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

Why are two checkpoint versions provided?

The original released checkpoints contain serialized historical Torchlight training metadata.

The trained SSC-HSR model tensors themselves do not depend on Torchlight, but the serialized metadata causes ordinary checkpoint loading to require the old Torchlight package.

The canonical model-best.pth files therefore retain only the model state needed for inference:

checkpoint["module"]["model"]

while the untouched artifacts remain available under the .original-torchlight.pth names for provenance and reproducibility.

No model weights were modified during this re-serialization.

For all three scale factors, the original and sanitized state dictionaries were verified to have:

• identical keys;
• identical tensor shapes;
• identical tensor dtypes;
• tensor-for-tensor equality using torch.equal.

The sanitized x4, x8, and x16 checkpoints were additionally strict-loaded into the corresponding released SSC-HSR architecture.

Reproducing the sanitized checkpoints

The repository includes sanitize_torchlight_checkpoint.py.

Scale factor 4

python sanitize_torchlight_checkpoint.py \
    sf4/ckpt/model-best.original-torchlight.pth \
    sf4/ckpt/model-best.pth

Scale factor 8

python sanitize_torchlight_checkpoint.py \
    sf8/ckpt/model-best.original-torchlight.pth \
    sf8/ckpt/model-best.pth

Scale factor 16

python sanitize_torchlight_checkpoint.py \
    sf16/ckpt/model-best.original-torchlight.pth \
    sf16/ckpt/model-best.pth

The script verifies the sanitized state dictionary against the original checkpoint after writing it.

Loading

The canonical checkpoints can be loaded using ordinary PyTorch:

checkpoint = torch.load("sf4/ckpt/model-best.pth", map_location="cpu")
state_dict = checkpoint["module"]["model"]
model.load_state_dict(state_dict, strict=True)

No Torchlight installation or compatibility loader is required for the canonical files.

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 checkpoints: released by the original authors.

This repository originally served as a mirror for convenient automated downloads and now additionally provides Torchlight-free checkpoint copies while retaining the untouched original artifacts.

All scientific credit goes to the original authors.