The Custom Containers feature lets you pull your own image from a container registry eg Docker Hub. This article will help you prepare a custom Docker container and show you how to bring that Container into Gradient by creating either a Notebook or an Experiment with your custom container. We recommend using Docker to get the container image from your system to Gradient.
Create a Dockerfile Host on GitHub or a local file. Example on GitHub Example: https://github.com/Paperspace/tf-jupyter-dockerfile/blob/master/Dockerfile
Run a Job to build the container from the Dockerfile and publish to a container registry Example:
paperspace jobs create \--apiKey XXXXXXXXXXXXXXXXXXx \--workspace /path/to/repo \--useDockerfile true \--buildOnly true \--registryTarget my-registry/name:tag \--registryTargetUsername my-username \--registryTargetPassword XXXXXXXXXXXXX
From that machine, you'll need to be logged into your Docker Hub account
docker login -u <username> -p <password>
You can make your own file (see Requirements below) or use one like this example: https://github.com/Paperspace/tensorflow-python
docker build -t <name of image>
for the example file above, you would enter:
docker build -t test-container
Verify with docker images. It should look something like this:
Tag the image so that is can be added to a repo with the image id, your Docker Hub username, and a name for the image :
docker tag <image id> <dockerhub username>/test-container:latest
docker push <username>/test-container:latest
Jupyter and all of Jupyter dependencies must be installed:
conda install -c conda-forge jupyterlab
If you don't specify a user, your container user will be 'root'
After you've pushed your custom container to Docker or you found a public container that is already there, it's time to pull it into Gradient!
Select Notebooks from your left-hand navigation menu. Click the Custom Containers tab in the Base Containers section.
Container Name = Path and tags of image eg ufoym/deepo:all-jupyter-py36
Username = your Docker Hub username, can be left blank for public images
Password = your Docker Hub password, can be left blank for public images
Default Entrypoint = must be Jupyter compatible, defaults to 'jupyter notebook' if left blank
Container user = optional user, defaults to 'root' if left blank
Follow the rest of the steps to create your Notebook by selecting your machine type, naming your notebook, and clicking Create.
Just pass in the container when creating your Experiment
paperspace-python experiment createAndStart ... --container ufoym/deepo:all-py36 ...
If your container is hosted in a password protected repository (you have set your Docker Image to 'Private'), you will need to enter your Docker Hub password. Otherwise, leave this field blank. This is not your Paperspace username and password.
Privacy settings for a given container might look something like this in Docker Hub:
if a Notebook image is private and the password is missing or invalid, the Notebook won't run :(