First-time PUMA Challenge Participant: Docker Submission Questions

Hi everyone!

As a first-time participant in a Docker-based competition like PUMA, I still have some questions about submitting my algorithm via Docker:

Do I need to clone ​all​ official GitHub repos (PUMA Challenge Evaluation - Track 1, PUMA Challenge Evaluation - Track 2, PUMA Challenge Baseline - Track 1, PUMA Challenge Baseline - Track 2), run their Docker containers on my server, then replace the checkpoints with my own trained models and modify the Dockerfile?

Even if my model architecture is ​different​ from the official baseline (e.g., not HoverNet/nnUNet), must I still base my submission on the official Dockerfile?

Is this workflow feasible for a Docker beginner? Any advice to simplify the process?

Thank you for your help!

I’ve trained my model and obtained the checkpoints, but I’m unsure how to proceed with submission.

Hi aia_wt,

Thank you for joining the challenge! We made our docker containers with two goals in mind: 1. To serve as an example algorithm to start with 2. To show what the specifics of the output needs to be by the docker container.

I would recommend to read the documentation on the grand challenge website on creating an algorithm: https://grand-challenge.org/documentation/create-your-own-algorithm/ as well as our submission instructions: https://puma.grand-challenge.org/submission-instructions/ https://puma.grand-challenge.org/submission/

In general, if you have created a different model just replacing checkpoints would not work due to the different model architecture.

If you have any further questions after reading this please let me know!

Kind regards, Mark

Hi Mark,

Thank you for the helpful information! I have trained my own model, and I already have the checkpoint, but it’s not based on the official HoverNet or nnUNet models.

In this case, should I continue using the official Dockerfile provided on GitHub? I plan to replace the checkpoint with mine, but I'm wondering what files exactly I need to modify. Is it enough to only modify inference.sh? Or are there other files that I need to adjust to ensure my model works correctly in the Docker container?

Looking forward to your advice!

Best regards, Tong

Dear Tong, I would recommend cloning the github repository as it will help with developing a pipeline which takes a .tif image as input generating the correct json files as an output. However, not only inference.sh need to be changed. Your container should work on an offline server so everything to make your model work should be in the container. I think it would be helpful to ask someone with whom you work with to help you with docker. It can be a bit confusing in the beginning to work with. In addition, it helps to take a look at the documentation it provided in my earlier reaction.

Good luck!

Regards, Mark