docker-tensorflow-armv7l

Guide to build + run container with tensorflow for ARMv7l

DockerHub Images

Repo with tf2.3 on debian buster Link to Docker Hub
Versions

• v1.0 -> base variant
• v1.1 -> added packages to download modules.
• v1.2 -> more packages to download modules, HTTPS functionality.
• v1.3 -> jupyter fully functional.
• v1.4 -> SKlearn and pandas included.

Cross compiling using buildx to generate armv7l tensorflow2.3 docker images

Set Dockerfile file to version we want

Requires linux with docker installed.
I'am using wsl2 with Ubuntu with docker installed.

Multi Architecture Builds Commands

Run the following commands to setup envrionment to cross compile docker containers.

[Enable experimental features]

• export DOCKER_CLI_EXPERIMENTAL=enabled

[List what architecture current builder can handle]

• docker buildx ls

[QEMU cross compilation framework]

• docker run --rm --privileged multiarch/qemu-user-static --reset -p yes

[Create new builder that can use QEMU]

• docker buildx create --name mybuilder --driver docker-container --use

[Display buildable platforms]

• docker buildx inspect --bootstrap

[Building, is the same, just with --platform]

• docker buildx build --platform <Target_Platforms> -t <tag> . --push
• docker buildx build --platform linux/amd64,linux/arm/v7,linux/arm64 -t 16fb/tf2.3:v1.4 . --push
  <tag> = <namespace>/<name>:<version>
  In this case its named tf2.3 in my personal 16fb namespace for version 1.4
  Replace 16fb/tf2.3:v1.4 with your respective image tag.

List of common platforms:

• linux/amd64
• linux/386
• linux/arm64
• linux/arm/v7
• linux/arm/v6
• linux/riscv64
• linux/ppc64le
• linux/s390x

[Test run docker container]

• docker run 16fb/tf2.3:v1.4

Run docker interactively

Pull Docker image to current directory

• docker pull 16fb/tf2.3:v1.4

Run in shell

• docker container run -it [docker_image] /bin/bash
• docker container run -it 16fb/tf2.3:v1.4 /bin/bash

Test Tensorflow version

• python3
• import tensorflow as tf
• print(tf.___version__)

Run docker Jupyter notebook

Runs Jupyter notebook server, exposing port 8899 on host machine, password auth is dependant on hash. How to compute Hash

export image="16fb/tf2.3:v1.4"
export Hash='<Hash here>'
export NOTEBOOKPORT=8899

docker run --env NOTEBOOKPORT=${NOTEBOOKPORT} --env Hash=${Hash} -p ${NOTEBOOKPORT}:${NOTEBOOKPORT}/tcp --rm -v $PWD:/data -w /data -u 0 --name JupyterNotebook -i $image /bin/bash << EOF	
export PATH=$PATH:~/.local/bin

echo Running Jupyter Notebook Server:
jupyter notebook --NotebookApp.password='$Hash' --no-browser --port=$NOTEBOOKPORT --ip=0.0.0.0 --allow-root

EOF

Making changes in running container

Opens bash shell on container docker exec -it JupyterNotebook /bin/bash

Run using Singularity

File size can be quite big.

Singularity pull

Singularity pull docker uri.

• singularity pull docker://16fb/tf2.3:v1.4

install squashfstools

squashfstools is required by machine to convert docker into singularity containers

• sudo apt-get install squashfs-tools

run in shell / interactively

• singularity shell 16fb/tf2.3:v1.4

Writable Version of Singularity Container (Dev purposes)

Obtain writable(can install modules) version of container by using a sandbox: \

Build a Sandbox to enable writable version of container

• sudo singularity build --sandbox <sand_box_name> <image_uri>
• sudo singularity build --sandbox tf2.3_sb_1.4 docker://16fb/tf2.3:v1.4

OR Convert SIF to Sandbox, be in same directory as SIF file (doesnt seem to work for me)

• sudo singularity build --sandbox <sand_box_name> <sif_file_path>
• sudo singularity build --sandbox tf2.3_sb_1.4 tf2.3_v1.4.sif

Run as shell, mount mounts, allow writes so can install, Ensure you are root.

• singularity shell --bind /mnt:/mnt --writable tf2.3_sb_1.4 /bin/bash

Export to tar file.

Export and Import docker image to file.

Save to tar file

• docker save --output tf2.3_v1.4 16fb/tf2.3:v1.4

Load from tar file

• docker load --input tf2.3_v1.4

Credits to these great ppl for guides

fgervais who did something similar.
this blog that guides using buildx.