cryovit 0.1.2

Deep learning automated segmentation models using high-level features from foundation ViT models.

CryoVIT: Efficient Segmentation of Cryo-electron Tomograms with Vision Foundation Models

Installation Instructions

CryoVIT uses Mamba to manage python packages and dependencies and can be downloaded here. You should also be able to use Conda instead of Mamba but setting up the environment may take an unreasonably long time.

1. Clone the CryoVIT github repository: git clone https://github.com/VivianDLi/CryoVIT.git
2. Setup the mamba environment: mamba env create -f CryoVIT/env.yml
3. Activate the mamba environment: mamba activate cryovit_env

Usage Instructions

For more details, check out the documentation.

Replicating Experiments

All experiments are managed using Hydra, with their configurations stored in src/cryovit/configs.

Before running experiments, make sure to change the relevant entries in src/cryovit/configs/paths/default.yaml.

You can run a specific experiment using the scripts in the cryovit.training module.

For example:

$ mamba activate cryovit_env
$ python -m cryovit.training.dino_features
$ python -m cryovit.training.train_model +experiment=<experiment_config_name>
$ python -m cryovit.training.eval_model +experiment=<experiment_config_name>

These commands assume your data is setup in the following format:

dataset/
├── sample_1/
│   ├── tomogram1.hdf
│   ├── tomogram2.hdf
│   └── ...
├── sample_2/
│   ├── tomogram1.hdf
│   ├── tomogram2.hdf
│   └── ...
├── sample_3/
│   └── ...
├── sample_4/
│   └── ...
│   ...

where each tomogram has a data key, with labels saved in labels/<label_name> keys.

All figures in the paper were produced using the cryovit.training.visualize_results command.