Create an experiment

How to create an experiment

To create an experiment you need to select few options including cluster, resource, image, start command, volume, environment variables, and termination protection.
The overall options for creating an experiment
Put the asterisk mark(*) on the required option

Runtime*

Cluster*

If you successfully added your custom cluster, then you can choose between the SavviHub-managed cluster and the custom cluster. The SavviHub-managed cluster is on the cloud vendor server, whereas the custom cluster could be the cloud server or the on-premise server.
SavviHub Cluster
Custom Cluster
For the SavviHub Cluster, you should choose the type of resource that the pod will use. Select the resource option you need for the service you are running.
SavviHub also supports the spot instances type.
SavviHub Cluster with the GPU resource type
For the Custom Cluster, you should choose the processor type and specify the resource requirements. The experiment job will be automatically assigned to an available node according to the given resource requirements.
Custom Cluster with processor and resource requirements

Image*

You can choose the Docker image that the experiment container will use. There are two types of images: the Python image and the public image. Select the Docker image that you want to run on the experiment container.
Python Image
Docker Image
Python image type is pre-pulled images by SavviHub. You can find the available image tags in SavviHub Docker Hub. Some images are pushed by SavviHub and the others are frequently used PyTorch and Tensorflow images. The detailed information is provided in the Readme in Docker Hub. To list the installed pip packages, click the Packages button on the right side of Image form.
Python image with all machine learning packages installed
Docker image type is NOT managed by SavviHub. Instead, it allows you to pull images from either Docker Hub or AWS ECR.

Public image

To pull images from the public Docker registry, you can simply pass the image URL as the below example of the official TensorFlow development GPU image for Docker Hub.
Public image from Docker Hub

Private image

To pull images from the private Docker registry, you should integrate your credentials in organization settings first. Then check the private image checkbox, fill in the image URL, and select the credentials you have just integrated. Below is an example of a private image from the AWS ECR.
AWS ECR private image with the integrated AWS access key

Start Command

You must specify the start command in the experiment container. You can put a running script with command-line arguments. Also, if you want to put multiple commands, you can write the start command as command 1 && command 2.
A start command to run python script

Volume

Users can mount the project, dataset, and files to the experiment container.
Mounted dataset(floyd/mnist-s3-latest) and project(savvihub/examples)

Environment Variables

Users can set environment variables as key-value pairs. You can also delete the variables by clicking the trash can button on the right.
Add the environment variables to the experiment

Termination Protection

You might want to access the experiment container after it runs. Termination protection allows you to do that. If you checked the checkbox, then the experiment will go idle status after finish it.
Enable termination protection with checkbox
Last modified 4mo ago