WARNING This will only work in the FAIR cluster since that's where all our datasets are, some paths are even hardcoded.
- Clone the git repo with all the submodules included (only if you are in the private LSeg and Detic forks.)
git clone --recursive git@github.com:aszlam/geometric_database.git
- Create a new conda environment, install
pytorch
and habitat.
a. Important to keep the python version <= 3.8.
b. Keep python version <= 1.8.2 b. Important to install pytorch from conda, my pytorch installation from PIP would freeze and crash sometimes. - Install the pip requirements,
pip install -r requirements.txt
- Log in to wandb, optionally, if you want to track the trainings.
- Create a
.cache
directory for caching dataloader (speeds up subsequent runs). - Run
python train_surface_model.py
to train the implicit models. Seeconfigs/train.yaml
to figure out the configs.