opendatacam

How to create / update a docker image for a nvidia-docker machine

Note about CUDA_ARCH_BIN

Seems this will be non needed as the opencv build we use doesn’t need CUDA support in end

Our Dockerfile is set for CUDA_ARCH_BIN=6.1 , if you GPU doesn’t support this architecture, you need to build with another CUDA_ARCH_BIN version.

Find out what is your CUDA Compute Capability depending on your GPU

If your CUDA_ARCH_BIN isn’t 6.1, change the flag in the Dockerfile

1. Pre-requisite

2. Build the image

# Go to an empty folder
mkdir docker
cd docker
# get the docker file : https://github.com/opendatacam/opendatacam/blob/master/docker/run-cloud/Dockerfile
wget https://raw.githubusercontent.com/opendatacam/opendatacam/master/docker/run-nvidia-docker/Dockerfile
# Build
# Takes a really long time the first time as it compiles opencv
sudo docker build -t opendatacam .

# If you are building a second time, use this to pull the latest opendatacam code
# TODO change this by adding the tag of the version in the Dockerfile
# Technique to rebuild the docker file from here : https://stackoverflow.com/a/49831094/1228937
# Build using date > marker && docker build .
date > marker && sudo docker build -t opendatacam .

# Test the image in interactive mode
sudo docker run --runtime=nvidia -p 8080:8080 -p 8090:8090 -p 8070:8070 -v /data/db:/data/db --rm -it opendatacam

3. Publish the docker image

# Log into the Docker Hub
sudo docker login --username=opendatacam
# Check the image ID using
sudo docker images
# You will see something like:
# REPOSITORY              TAG       IMAGE ID         CREATED           SIZE
# opendatacam             latest    023ab91c6291     3 minutes ago     1.975 GB

# Tag your image
sudo docker tag 7ef920844953 opendatacam/opendatacam:v2.1.0-nvidiadocker_cuda_archbin_6_1

# Untag image (if you made a tipo)
sudo docker rmi opendatacam/opendatacam:v2.1.0-nvidiadocker_cuda_archbin_6_1

# Push image
sudo docker push opendatacam/opendatacam:v2.1.0-nvidiadocker_cuda_archbin_6_1

NOTE: Known Improvements to make

For now the docker image is very large (12GB), need to try use the lightweight runtime of nvidia/cuda to have a build part and a run part

Would need to copy the opencv compiled file (maybe the deb as with the jetson), and the darknet compiled folder directly

Links: