DOCKER.md

Building the container for train.sh

train.sh runs the model inside an Apptainer (Singularity) image:

apptainer exec --nv $CONTAINER python main.py fit ...

where CONTAINER points to a .sif file. The Dockerfile defines the environment; you just need to turn it into that .sif file. Three steps:

1. Build the Docker image

From the project root (where the Dockerfile is):

docker build -t atlas:latest .

2. Convert it to a .sif file

HPC clusters use Apptainer, not Docker, so convert the image:

# Save the Docker image to a tar archive
docker save atlas:latest -o atlas.tar

# Build the .sif from that archive
apptainer build atlas.sif docker-archive://atlas.tar

(If your machine has Apptainer with Docker access, you can skip the tar step: apptainer build atlas.sif docker-daemon://atlas:latest.)

3. Point train.sh at the .sif

Copy atlas.sif to your cluster and set the path at the top of train.sh:

CONTAINER=/path/to/atlas.sif

Now train.sh (or sbatch train.sh) will run training inside the container. The --nv flag gives the container access to the GPU.