Error loading FastSAM checkpoint on Windows 10 with Python 3.12.8 – UnpicklingError: Unsupported global ultralytics.nn.tasks.SegmentationModel

[MauroAndretta]

I'm trying to run FastSAM on my Windows 10 machine using VS Code with a conda environment named segmentation_test (Python 3.12.8).

I followed the instructions from the FastSAM_example.ipynb notebook and did the following:

1. Cloned the repository:
   git clone https://github.com/CASIA-IVA-Lab/FastSAM.git

2. Downloaded the weight file from:
   https://huggingface.co/spaces/An-619/FastSAM/resolve/main/weights/FastSAM.pt

3. Installed the required packages:

!pip install -r FastSAM/requirements.txt
!pip install git+https://github.com/openai/CLIP.git

Then, I attempted to run the inference command from the terminal:

python FastSAM/Inference.py --model_path ./weights/FastSAM.pt --img_path ./images_for_models_comparison/png/image_tech_398_006657a8.png --imgsz 512 --device cpu

However, I encountered the following error:

Traceback (most recent call last):
  File "C:\Users\mauoa\Desktop\dl-image-segmentation\notebooks\SegmentationTest\FastSAM\Inference.py", line 122, in <module>
    main(args)
  File "C:\Users\mauoa\Desktop\dl-image-segmentation\notebooks\SegmentationTest\FastSAM\Inference.py", line 76, in main
    model = FastSAM(args.model_path)
            ^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\mauoa\Desktop\dl-image-segmentation\notebooks\SegmentationTest\FastSAM\ultralytics\yolo\engine\model.py", line 107, in __init__
    self._load(model, task)
  File "C:\Users\mauoa\Desktop\dl-image-segmentation\notebooks\SegmentationTest\FastSAM\ultralytics\yolo\engine\model.py", line 156, in _load
    self.model, self.ckpt = attempt_load_one_weight(weights)
                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\mauoa\Desktop\dl-image-segmentation\notebooks\SegmentationTest\FastSAM\ultralytics\nn\tasks.py", line 578, in attempt_load_one_weight
    ckpt, weight = torch_safe_load(weight)  # load ckpt
                   ^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\mauoa\Desktop\dl-image-segmentation\notebooks\SegmentationTest\FastSAM\ultralytics\nn\tasks.py", line 518, in torch_safe_load
    return torch.load(file, map_location='cpu'), file  # load
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\mauoa\miniconda3\envs\segmentation_test\Lib\site-packages\torch\serialization.py", line 1470, in load
    raise pickle.UnpicklingError(_get_wo_message(str(e))) from None
_pickle.UnpicklingError: Weights only load failed. This file can still be loaded, to do so you have two options, do those steps only if you trust the source of the checkpoint.
        (1) In PyTorch 2.6, we changed the default value of the `weights_only` argument in `torch.load` from `False` to `True`. Re-running `torch.load` with `weights_only` set to `False` will likely succeed, 
but it can result in arbitrary code execution. Do it only if you got the file from a trusted source.
        (2) Alternatively, to load with `weights_only=True` please check the recommended steps in the following error message.
        WeightsUnpickler error: Unsupported global: GLOBAL ultralytics.nn.tasks.SegmentationModel was not an allowed global by default. Please use `torch.serialization.add_safe_globals([SegmentationModel])` or the `torch.serialization.safe_globals([SegmentationModel])` context manager to allowlist this global if you trust this class/function.

What I've Tried:

I followed the instructions from the Colab notebook and installed the requirements.
I'm running on a Windows 10 PC in a conda environment with Python 3.12.8.
I've forced inference on the CPU using the --device cpu flag.
According to the error message, one solution would be to allowlist the SegmentationModel global by using torch.serialization.add_safe_globals([SegmentationModel]).

However, I'm unsure where and how to integrate this in the FastSAM code.

Questions:

What is the recommended approach for resolving this unpickling error with ultralytics.nn.tasks.SegmentationModel?
Should I modify the repository code to use torch.serialization.add_safe_globals or use weights_only=False when calling torch.load?
Are there any other known issues with running FastSAM on Windows 10 with Python 3.12.8?
Any guidance would be appreciated. Thank you!

[ericiamarino]

Edit: I modified the requirements.txt and was able to get functionality. The primary things I changed are the following:

Edits:
torch>=1.7,<2.6
torchvision>=0.8,<0.14

Additions:
numpy<2.0

[MauroAndretta]

Sorry @ericiamarino which python version did you use?

[ericiamarino]

Python 3.9.21