Most installation steps are similar to ConvNeXtV2, and we provide the instructions below.
To make the installation easier, you can either choose to manually install the packages as shown below, or use the env.yml file and install it using mamba mamba env create -f env.yml.
This installation is tested for CUDA 11.8.
Creating a new conda environment
conda create -n mmearth-train python=3.9 -y
conda activate mmearth-train
pip install torch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txtInstall Minkoswki Engine (this is only required for pre-training the model from scratch, since ConvNeXt V2 uses sparse convolutions, which are implemented in the Minkowski Engine):
We use GCC 11.X for the installation.
git submodule update --init --recursive
git submodule update --recursive --remote
conda install openblas-devel -c anaconda
cd MinkowskiEngine
python setup.py install --blas_include_dirs=${CONDA_PREFIX}/include --blas=openblasPossible errors: Incase you get errors referencing the at:Tensor variable in src/spmm.cu file. Consider adding #include <ATen/core/Tensor.h> to the same file and re run the setup.
If you want to run the code using ffcv, you also need to install this:
conda config --env --set channel_priority flexible
conda install cupy pkg-config compilers libjpeg-turbo opencv numba -c conda-forge
pip install ffcvCreating a new conda environment
conda create -n mmearth-train python=3.9 -y
conda activate mmearth-train
pip install torch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txtIf you want to run the code using ffcv, you also need to install this:
conda config --env --set channel_priority flexible
conda install cupy pkg-config compilers libjpeg-turbo opencv numba -c conda-forge
pip install ffcvInstall GEO-Bench for finetuning:
pip install geobenchVisit their website for the complete installation and data downloading guide.