nets

Pytorch networks used by r-pad.

Excludes pytorch-geometric-based networks, (which can be found in r-pad/pyg_libs).

List of networks implemented: *

Installation

All of the following should be done inside your venv or conda environment.

Strongly recommend using a conda env, because it makes installing flash attention MUCH easier.

Prerequisites

DiPTv3

1. Choose a CUDA version you'll want to use! Should be compatible with everything... CUDA 12.4 is good, but so is 11.8.

2. Install CUDA toolkit inside your conda env, so that nvcc is available. For instance:

   conda install -c "nvidia/label/cuda-12.4.0" cuda-toolkit

3. Choose a torch version (including GPU) and install it. See here for instructions. For instance:

    conda install pytorch torchvision torchaudio pytorch-cuda=12.4 -c pytorch -c nvidia

4. Install an appropriate torch_scatter version. See here For instance:

       pip install torch_scatter -f https://data.pyg.org/whl/torch-2.5.0+cu124.html

5. Install spconv.

       # spconv (SparseUNet)
       # refer https://github.com/traveller59/spconv
       pip install spconv-cu124  # choose version match your local cuda version

6. Install flash_attention.

       pip install flash-attn --no-build-isolation

Option 1: No changes needed (simplest).

If you want to use the architecutres as-is (or their building blocks), you can just do the following:

export NETS_MODULE=diptv3
pip install 'git+https://github.com/r-pad/nets.git#egg=nets[$NETS_MODULE]@main'

You can replace main with a specific branch or tag.

Option 2: Need to be able to modify the code.

If you want to be able to modify the code, but don't want to contribute back, you can do the following:

export NETS_MODULE=diptv3
cd $CODE_DIRECTORY
git clone https://github.com/r-pad/nets.git
cd nets
pip install -e '.[$NETS_MODULE]'

You can then modify the code in $CODE_DIRECTORY/nets and the changes will be reflected in your environment. And if you want to contribute back, you can create a branch and PR.

Usage