Help with bundle contributing

[ybagdasa]

Hi! My team working on wrapping our existing OCT RPD segmentation model into a bundle and adding it to the model zoo. I've been following the contributing guidelines and I'm at the point where I'm almost ready to do a pull request. However, I am confused about a few things:

1. Is monai or the user responsible for the environment setup for the bundle (installing "required_packages_version" in metadata.json)?
2. Is monai, the user, or the bundle code responsible for downloading the files listed in "large_files.yaml"?
3. Is there an example of where these steps are done?

Trying to test all this out before a PR.
Thank you!

[ybagdasa]

@yiheng-wang-nv would you know?

[yiheng-wang-nv]

Hi @ybagdasa , thanks for the interest in contributing!
For your questions:

1. The author should be responsible to describe the required environment for the bundle. The purpose is to ensure that users can follow up with the instructions to run (or reproduce) the training/evaluation/inference.
2. For large files (such as model weights), to avoid using git-lfs, we recommend users to upload into a host and provide the download links in large_files.yaml file. During the review stage, MONAI maintainers and our CI pipeline can use a script to grab the required large files and run tests. If the bundle is merged into the main branch, a post-merge script will be executed to collect all large files and pack up the whole bundle for users to download (Now, we use Hugging Face as the host, for example, https://huggingface.co/MONAI/lung_nodule_ct_detection/tree/0.6.9)
3. You can refer to any bundles merged in the dev branch, for example: https://github.com/Project-MONAI/model-zoo/tree/dev/models/brats_mri_segmentation

[ybagdasa]

@yiheng-wang-nv Thank you for the clear reply!
I'm having issues reproducing the required environment from scratch, since it relies on Detectron2 which has its own requirements for pytorch and cuda versions. I was wondering whether a docker file would be an acceptable provision for the environment setup, since that has been working for me.
Additionally, it looks like running pip install monai installs its own versions of pytorch and cuda. I'm wondering whether order matters here.

[yiheng-wang-nv]

HI @ybagdasa , could you submit a draft PR which includes all dependencies? Then we can work together to figure out how to prepare the accurate metadata.json. We also used a default docker image (with basic dependencies installed, like python 3.10) in our CI, and install the requirement file to run the bundle to check.
As for the pip install monai, I think you can install like pip install monai torch==2.6.0 ... to specify the expected torch vision.

[ybagdasa]

@yiheng-wang-nv I was able to find a working install prescription (with pip install) and added it to the readme and the ci/install_scripts. It works with python == 3.9, but not later version (due to a quirk of the detectron2 installation). I updated the metadata.json file as well. I submitted the draft PR.

Thanks,
Yelena