[Bug] Converting BiRefNet-HRSOD.pth failed

[xiazhesheng]

Hello, this is a great job to help me using quantitized birefnet easier.

In my using cases, the BiRefNet-HRSOD produces better results than general BiRefNet models. So I tried to convert BiRefNet-HRSOD.pth into .gguf model using command "uv run scripts/convert.py birefnet BiRefNet-HRSOD_DHU-epoch_115.pth". However, it failed to convert.

This is the error report.

~/vision.cpp$ uv run scripts/convert.py birefnet BiRefNet-HRSOD_DHU-epoch_115.pth
Converting birefnet
* input: BiRefNet-HRSOD_DHU-epoch_115.pth
* output: models/BiRefNet-HRSOD_DHU-epoch_115.gguf
* birefnet.tensor_data_layout = whcn
Error: Error while deserializing header: header too large

Does it mean that the .pth file size is too large to convert? And what should I do to convert it successfully? Thanks a lot if you could help me.

[Acly]

I think issue is that convert.py for BiRefNet only handles .safetensors files, not .pth.

You can either adapt the script, something like this

    # instead of this:
    weights = safetensors.safe_open(input_filepath, "pt")
    model: dict[str, Tensor] = {k: weights.get_tensor(k) for k in weights.keys()}

    # do this:
    model: dict[str, Tensor] = torch.load(input_filepath, map_location="cpu", weights_only=True)

or try .safetensors version, like https://huggingface.co/ZhengPeng7/BiRefNet-HRSOD/tree/main

[xiazhesheng]

Thank you! Both the two methods succeeded!