ObjectNet Challenge

This page provides in-depth instructions on encapsulating your predictive model and code into a Docker image for submission to the ObjectNet Challenge. Containerising your model and code in a Docker image provides a self-contained execution environment that allows the challenge platform to run your model and evaluate your results.

If you are not familiar with docker instructions are available on how to install docker, along with a quick start guide.

Follow the steps outlined below to create a Dockerfile and build and test your docker image. Upon successful completion of locally testing your docker container, you may then continue to submit your model to the challenge.

Select your instruction set

Challenge participants should use one of the following instructions that matches the framework used to create a docker image:

Docker internet access, memory and execution time constraints

The following restrictions will be placed on docker images submitted for evaluation to the ObjectNet Challenge:

Network access will be disabled for Docker containers submitted to the challenge, i.e., your code and models will not be able to access the internet. Ensure that any code and models packaged into the docker image are self-contained and do not require access to the internet.
Docker storage usage is limited to 15GB.
An execution time limit of 2 hours has been set for submissions. Submissions exceeding this time limit will be interrupted and the submission will fail. To avoid exceeding the time limit use optimisation techniques such as batching and parallel data loading to improve inference efficiency.
Submissions are required to use GPUs for inferencing. Each GPU has 16GB of memory allocated. Please ensure your models are able to operate within the 16GB memory limit on the GPUs.

Managing Docker Artifacts

When you build and run docker images locally, and when you submit images for evaluation using the evalai push command, docker images and containers are stored on your local hard drive. If you do not housekeep these containers and images, they can quickly consume a significant amount of space on your hard drive.

Housekeeping Docker images

To list the images stored locally use:

# docker images -a

To remove an individual image use:

# docker rm <image> <image>

To remove dangling images. Dangling images are layers that have no relationship to any tagged images. They no longer serve a purpose and consume disk space.

# docker image prune <image> <image>

To list images matching a certain pattern:

# docker images -a |  grep "pattern"

To remove images matching a certain pattern:

# docker images -a | grep "pattern" | awk '{print $3}' | xargs docker rmi

Housekeeping Docker containers

A docker container is essentially an instantiation of a docker image at runtime. To list all containers:

# docker ps -a

To remove containers:

# docker rm <ID_or_Name> <ID_or_Name>

To remove all exited containers:

# docker rm $(docker ps -a -f status=exited -q)

To list containers matching a pattern:

# docker ps -a |  grep "pattern”

To remove containers matching a pattern:

# docker ps -a | grep "pattern" | awk '{print $1}' | xargs docker rm

Additional Information

For more information on housekeeping your docker environment see How To Remove Docker Images, Containers, and Volumes

Other Docker resources

Docker base images

Whenever possible, use current Official Repositories as the basis to build your customised image:

Other information

For more info about using Caffe to build deep learning models.

ObjectNet

Challenge Portal

ObjectNet Support