Create a docker multistage build script

[dkuegler]

Create a docker multistage build script that is compatible with FastSurfer interface rework.

All images are built with the "core" Dockerfile at Docker/Dockerfile. Also update some additional folders to not be copied automatically by the build (.dockerignore).

TODOs:

• Testing
• check for the updates to AMD (rocm/pytorch is now python 3.8 as well)

[dkuegler]

@AhmedFaisal95 Please have a look and see what still needs to be done.
I have already added some documentation in Docker/README.md

At a minimum, this drastically reduces the doubling of code.

It seems that DOCKER_BUILDKIT also supports the script now. Which speeds up builds a lot, since it builds stages that are independent in parallel.

For example

DOCKER_BUILDKIT=1 docker build --rm=true --target runtime --build-arg DEVICE=cuda --build-arg FREESURFER=pruned -t fastsurfer:gpu -f ./Docker/Dockerfile .

[af-a]

This seems to run fine except when:

1. build arg DEVICE is set to cpu (any target)
2. the target is runtime_surf_only and DEVICE is set to none.

Case 1

Command:

DOCKER_BUILDKIT=1 docker build --rm=true --target runtime --build-arg DEVICE=cpu -t IMAGE_ID -f ./Docker/Dockerfile .

Result:

[+] Building 0.6s (11/17)                                                                                                                                                                                                                                                                                                                                                           
 => [internal] load build definition from Dockerfile                                                                                                                                                                                                                                                                                                                           0.3s
 => => transferring dockerfile: 5.62kB                                                                                                                                                                                                                                                                                                                                         0.0s
 => [internal] load .dockerignore                                                                                                                                                                                                                                                                                                                                              0.4s
 => => transferring context: 136B                                                                                                                                                                                                                                                                                                                                              0.0s
 => [internal] load metadata for docker.io/library/ubuntu:20.04                                                                                                                                                                                                                                                                                                                0.0s
 => CANCELED [internal] load build context                                                                                                                                                                                                                                                                                                                                     0.0s
 => [build_conda_base 1/3] FROM docker.io/library/ubuntu:20.04                                                                                                                                                                                                                                                                                                                 0.0s
 => CACHED [build_conda_base 2/3] RUN apt-get update && apt-get install -y --no-install-recommends       wget       git       ca-certificates       upx       file &&     apt clean &&     rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*                                                                                                                                       0.0s
 => CACHED [build_conda_base 3/3] RUN wget --no-check-certificate -qO ~/miniconda.sh https://repo.continuum.io/miniconda/Miniconda3-py38_4.11.0-Linux-x86_64.sh  &&      chmod +x ~/miniconda.sh &&      ~/miniconda.sh -b -p /opt/conda &&      rm ~/miniconda.sh                                                                                                             0.0s
 => CACHED [build_conda_cpu 1/1] RUN conda install --channel pytorch cpuonly=1.0                                                                                                                                                                                                                                                                                               0.0s
 => CACHED [runtime 1/7] RUN apt-get update && apt-get install -y --no-install-recommends       tcsh       time       bc       gawk       libgomp1       libquadmath0 &&     apt clean &&     rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*                                                                                                                                    0.0s
 => CACHED [runtime 2/7] RUN echo "source /venv/bin/activate" >> ~/.bashrc                                                                                                                                                                                                                                                                                                     0.0s
 => ERROR [runtime 3/7] COPY --from=build_conda /venv /venv                                                                                                                                                                                                                                                                                                                    0.0s
------
 > [runtime 3/7] COPY --from=build_conda /venv /venv:
------
failed to compute cache key: "/venv" not found: not found

Case 2

Command:

DOCKER_BUILDKIT=1 docker build --rm=true --target runtime_surf_only --build-arg DEVICE=none --build-arg FREESURFER=pruned -t IMAGE_ID -f ./Docker/Dockerfile .

Result:

[+] Building 1.6s (4/4) FINISHED                                                           
 => [internal] load build definition from Dockerfile                                  0.3s
 => => transferring dockerfile: 140B                                                  0.0s
 => [internal] load .dockerignore                                                     0.5s
 => => transferring context: 136B                                                     0.0s
 => ERROR [internal] load metadata for docker.io/library/base_none:latest             1.0s
 => [internal] load metadata for docker.io/library/ubuntu:20.04                       0.0s
------
 > [internal] load metadata for docker.io/library/base_none:latest:
------

[af-a]

The issues mentioned above should be fixed in this PR: dkuegler#2

[dkuegler]

@AhmedFaisal95 Can you confirm this works now?

We need a "basic test" of all images :).

[dkuegler]

@m-reuter The Apple M1 device is missing from the docker multistage built. The reason is that it was not in dev at all.

For the AMD build, the base image got updated and I am not really sure if the updated image works. Maybe, someone can check this as well?

[af-a]

@dkuegler

@AhmedFaisal95 Can you confirm this works now?

We need a "basic test" of all images :).

I already tested the fixes before submitting the PR, specifically all images except for the AMD case (no hardware).

[dkuegler]

The build script should also compile the byte code https://docs.python.org/3/library/py_compile.html .

[dkuegler]

I think what is left for this PR to be merged is proper testing.
Especially for the AMD build.

[m-reuter]

The Apple M1 device is missing from the docker multistage built. The reason is that it was not in dev at all.

For M1 we do not use a docker, only native install on Mac. Inside Docker it was much slower for yet unknown reasons, see #109

[dkuegler]

@agirodi can you test the AMD docker script?

[agirodi]

@agirodi can you test the AMD docker script?

I just did and it is working correctly. The build went without issues and running produced a correct segmentation.

[dkuegler]

Then, this PR should be ready for merging.

[m-reuter]

A few points for documentation:

• due to the minimal larger size when including FreeSurfer binaries, we will stop publishing docker images without FreeSurfer
• pytorch gpu is still much larger than cpu version so we will continue to publish both versions
• AMD is experimental and can break without notice, which is why we will provide build files, but will not publish the container for now.
• As dependencies (versions) remain relatively stable, we will have a base release image without pytorch (including all other dependencies like python packages and FreeSurfer binaries), and then two versions from that with pytorch GPU and CPU.
• FastSurfer is then added on top of these three as the final layer. The non-pyorch entry point remains recon-surf, for the other two it will be run_fastsurfer.sh.. Users who want to directly run_prediction, need to modify the docker run command and overwrite the entry point or add another layer.

The advantage of this approach is, that for future docker builds we can pull the base release images and use them (in most cases they probably exist locally anyway). Adding FastSurfer is then extremely fast :-)