Docker image submission

Hi there, I have a question regarding the docker submission. If we are submitting for both tasks, how do we distinguish the two docker images? Should we tag them as organoids-task1 and organoids-task2? Is it permitted to just submit one short paper covering both tasks? Best, Fabian

Ah and while we're at it: will the newest pytorch container from nvidia (nvcr.io/nvidia/pytorch:22.06-py3) work on the server? It requires a cuda 11.7 compatible driver. Best, Fabian

Final question: Will there be enough space on /output for us to write temporary files? If space is limited, how much space do we have? Best, Fabian

I submitted the docker images with the same name, assuming different tasks run different servers. I confirmed to organizers that one paper is OK if you use the same method in both two tasks.

I don't know about the next two questions. I used nvcr.io/nvidia/pytorch:20.10-py3, I assumed there will be enough space and memory during inference.

Thanks! And congrats on your leaderboard position :-) It's going to be a tight race

You too! Our methods mainly depend on your proposed nnU-Net. Thanks ;)

hi, Issense, we miss this meesage and sorry for reply. For the frist question: Either form is fine, just indicate the task of docker when submitting the email. For the second questION, we have not tried this environment yet, but we believe we can easily adapt our environment to fit the submitted docker, and if there are any environmental issues during testing, we will communicate with the participating teams in a timely manner, and for the last problem, yes, we believe we can provide enough space for a "reasonable" submission algorithm (as discussed in the problem pf previous inference time).

Thanks, fantastic! If there are issues with our containers due to too little space for temporary files we can probably fix this quite easily. However, all our containers are already done and I would prefer not to do this now because I might break things and don't have enough time to validate any changes I might be doing.

All the new NGC container from Nvidia needs (I think) is a graphics driver that is compatible with CUDA 11.7. You can check that by running nvidia-smi and read the cuda version from the upper right corner. It should say 11.7 or newer. If it's older then installing a new driver will solve the problem. The docker might still run in compatibility mode if the cuda version is not satisfied but this may be at a reduced speed. In that case we might not be able to adhere to the 72h inference limit per task. (right now we estimate 60h for Task 1 and 50h for Task 2, both evaluated on an RTX 3090 using our dockers)

Best, Fabian