hsifn-checkpoints

HSIFN Released Checkpoint

This repository mirrors the released HSIFN checkpoint associated with Hyperspectral Image Super Resolution with Real Unaligned RGB Guidance (HSI-RefSR) by Lai, Fu, and Zhang.

The original checkpoint was obtained from the release provided by the original authors. This Hugging Face repository is maintained to provide stable and convenient programmatic access to the checkpoint.

Checkpoint

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

• 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 checkpoint goes to the original authors.

Torchlight-free checkpoint

The original released checkpoint contains serialized historical Torchlight training metadata.

The HSIFN model tensors themselves do not depend on Torchlight, but this metadata causes the original artifact to require the old Torchlight package during deserialization.

The canonical model-best.pth therefore preserves the released model state:

checkpoint["module"]["model"]

while removing the unused Torchlight serialization dependency.

The original and sanitized model states were verified to have:

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

The sanitized checkpoint was also strict-loaded into the released HSIFN / CrossNetHSI architecture.

Reproducing the sanitized checkpoint

The repository includes sanitize_torchlight_checkpoint.py.

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

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

Loading

The canonical checkpoint 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 checkpoint.

Original work

• Paper: Hyperspectral Image Super Resolution with Real Unaligned RGB Guidance, Zeqiang Lai, Ying Fu, and Jun Zhang.
• Project / implementation: HSI-RefSR, by the original authors.
• Dataset: https://huggingface.co/datasets/ZeqiangLai/HSI-RefSR
• Original checkpoint: released by the original authors.

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

All scientific credit goes to the original authors.