Skip to content
This repository has been archived by the owner on Jan 18, 2022. It is now read-only.

Latest commit

 

History

History
88 lines (58 loc) · 4.67 KB

README.md

File metadata and controls

88 lines (58 loc) · 4.67 KB
Raw

nnUNet network architecture

2D

2D nnUNet

3D

3D nnUNet

nnUNet Docker Container (Inference)

This container takes the nnUNet code and inherits from a pytorch:pytorch Docker image. This container is supposed to run inference only, both CPU and GPU mode for VerSe 2019 challenge. To run with GPU, you need to install and configure nvidia-container-runtime.

For MICCAI 2019 VerSe challenge

Build the image

To build the docker image, you need to run this command in nnUNet's parent directory:

docker docker build . -f nnUNet/docker/Dockerfile -t deepspine

To avoid send a large context files and directories when building docker image, you should put the whole nnUNet folder to an empty folder.

Environment-Parameters

  • INPUTDIR: dir path to the input folder containing the compressed nifti files (nii.gz):
/data/
 - imageA.nii.gz
 - imageB.nii.gz

The nifti files are renamed by the container, so you don't have to rename them before.

  • OUTPUTDIR: dir path to the output folder containing all final segmentations which were calculated during the container process, which is set to /data/results by default.
  • MODELS_DIR: this is very important to make right, because here you have to mount a directory containing your model dir tree starting from the 'nnUNet'-folder Example: \opt\path\data\nnUNet is the correct path when your model lives here: \opt\path\data\nnUNet\3d_fullres\Task19_DeepSpine\nnUNetTrainer__nnUNetPlans. Actually, this is the RESULTS_FOLDER when you run nnUNet training.
  • TASK_NAME: this name should match the task name you used for training the model. The task name appears also within the RESULTS_FOLDER variable, which is set to Task19_DeepSpine.
  • THREADS: how many threads used to processing nift files, set to 16 by default, set to a smaller one if you run out of memory.
  • JOBS: how many jobs to run parallel, set to 4 by default, set to smaller one if you run out of GPU memory.

You may want to change data loading or output structuring. This shows an example for VerSe 2019 challenge.

Run it

Simply run:

docker run -v datadir:/data -v modelsdir:/models -itd --rm deepspine

If GPU is available:

docker run --runtime=nvidia -v datadir:/data -v modelsdir:/models -itd --rm deepspine

change datadir and modelsdir to corresponding dir.

Our model is available at DropBox, download the models and extract to modelsdir. By default, this docker image use entrypoint /nnUNet/docker/run-all.sh to run inference using ensemble of 2D U-Net, 3D low resolution U-Net, and 3D full resolution U-Net. To use the ensemble without 3D low resolution U-Net in case you got memory error due to numpy issue, you could specify option –-entrypoint /nnUNet/docker/run.sh.

Our Docker image is available at DockerHub. You could pull it by:

docker pull butui/deepspine

Note: the Dockerfile for VerSe challenge here might be broken, but the Docker image I push to DockerHub show work as expected.

For MICCAI 2019 StructSeg challenge

build the image

For convenient, I build a base Docker image with Dockerfile.base, which install the necessary package for nnUNet. This base image is build from pytorch/pytorch:0.4.1-cuda9-cudnn7-devel. The organizer only accept CUDA 9.0, so we use this image as base. You are free to use newer version of PyTorch and CUDA if needed. To build this base image, run:

docker build -t structseg:base -f Dockerfile.base .

We assume you have build a base image tagged structseg:base from now on. Then we could build the image for each task. You should run this command in parent directory of nnUNet:

docker build -t structseg:USERNAME_task1 -f nnUNet/docker/Dockerfile.structseg --build-arg TASK_NAME_build=Task01_HaNOAR --build-arg FOLD_build=1 --build-arg MODELS_DIR_build=structseg-models --build-arg MODEL_build=3d_fullres .

Modify the ARG and ENV according to yours. For submission, I copy the trained models into the Docker image. You might consider comment out line 16 in Dockerfile.sturctseg, and simply mount the model directory to the Docker container when you run your Docker container. The ENTRYPOINT is required by the organizer, we run script run-structseg.sh in fact. And due to time and GPU memory limitation, we don't use ensemble scheme. You might consider using ensemble to improve performance.