This folder contains examples for how to prepare a container submission for task 2:
- A simplistic example using a dummy prediction algorithm (
_simple) - A baseline that uses nnUNet to make predictions (
_nnunet)
Each features a definition file (.def), which is used to build the container, and a python script that performs the actual prediction (prediction_*.py) when the container is run.
Instructions on how to build and run a singularity container are given below. For a comprehensive introduction to singularity, this tutorial (external resource) is recommended. Note that singularity has to be installed (instructions). You can build and run an example from above to check that your installation is working. If you want to build the nnUNet container yourself, you have to install nnunet and download the pretrained models.
A ready-to-use nnUNet container, which was trained on the BraTS 2020 data (without taking into account the partitioning), can be downloaded from here.
A definition file is required to build a container. You can start from the example .def-files provided in this repo to create your own and then run
sudo singularity build container_simple.sif container_simple.def
where container_simple should be replaced with your corresponding file names. To avoid building with sudo, you can use singularity's --fakeroot option, but you may have to configure this first (sudo singularity config fakeroot --add <username>). Make sure that all files that are required at test-time are copied to the container when you build it. The container will not have access to your filesystem and environment when evaluated in the testing phase!
Note: Please use a bootstrap image based on CUDA 11.0 to make sure that the application can run on systems with Ampere architecture, too. Our recommendations are:
nvcr.io/nvidia/pytorch:20.08-py3(pytorch)nvcr.io/nvidia/tensorflow:20.08-tf2-py3ornvcr.io/nvidia/tensorflow:20.08-tf1-py3(tensorflow)
Tip for debugging: use the --sandbox option, as described here.
Once you have successfully built your container, it's time to test it on some data. This is the command that will be executed at test-time:
DATA_DIR=/path/to/test/data/dir
OUT_DIR=/path/to/output/dir
export SINGULARITY_BINDPATH="$DATA_DIR:/data:ro,$OUT_DIR:/out_dir:rw"
singularity run -C --writable-tmpfs --net --network=none --nv container_simple.sif -i /data -o /out_dir
SINGULARITY_BINDPATH makes sure that all necessary data files can be accessed from within the container as bind mounts (pattern src:dest:read-write-mode). Please insert the path to your data and output directory here before running the script and make sure the data is in the form described here!
Description of the singularity run options (see also singularity run --help):
--C: Restrict access to host filesytem and environment to a minimum--writable-tmpfs: Allows writing to a in-memory temporary filesystem*--net --network=none: No network access--nv: use GPU (nvidia) support
We recommend using above options when performing local tests to avoid problems during the testing phase.
Tip for debugging: singularity shell instead of singularity run starts an interactive shell inside the container.
*) Note that this filesystem is usually rather small, so we recommend to store any intermediate outputs in the output directory instead and clean up after inference is complete.