Setting up

1. Clone GFM.

2. Create uv venv by running

   uv init

3. download dependencies:

   uv add torch pyyaml scipy termcolor timm yacs torchmetrics rasterio torchgeo opencv-python

4. [OPTIONAL: only for speeding up, you can ommit the dependency to apex by using the parameter --amp-opt-level O0 when running scripts]

   Clone apex and run commands:

   git clone https://github.com/NVIDIA/apex

   cd apex

   rm pyproject.toml

   uv run setup.py install

   uv run pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation --global-option="--cpp_ext" --global-option="--cuda_ext" ./

5. Download weights from imagenet

   wget https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_base_patch4_window7_224_22k.pth
   

   and place it under the following folder structure.

   output
   |- simmim_finetune
       |- swin_base_patch4_window7_224_22k.pth
   

6. Download GeoPile from NAS-3: "\\NAS-3\Imagery\ai-internship-2025\geopile\GeoPile.zip", place the file in data and unzip it to obtain the GeoPileV0 folder (might take several minutes to unzip).

7. To train their fundation model, you can run the main_teacher.py script as follows:

   uv run -m torch.distributed.run --nproc_per_node 1 main_teacher.py --cfg configs/simmim_pretrain__swin_base__img192_window6__100ep.yaml --batch-size 1 --data-path data/GeoPileV0 --tag gfm --pretrained output/simmim_finetune/swin_base_patch4_window7_224_22k.pth --amp-opt-level O0

   This is slightly different from GFM instruction because the torch.distributed library evolved. Note that we use a single GPU and a batch size that is much smaller because of current RAM issues. Depending on your set, adapt those parameters.