In the coming minutes, we'll walk you through setting up your account, linking your code and data, and training your model.
To keep things simple, we'll show you how to use Clusterone using a ready-to-run demo of a self-driving car simulation.
Clusterone is deep learning platform that allows you to train your models on distributed GPUs and CPUs without setup or maintenance. Think of it as the operating system for deep learning. Clusterone runs in the cloud, in on-premise installations, or even a combination of the two. We offer a SaaS platform as well as dedicated enterprise installations.
Before we begin, make sure you have your gear ready:
A Clusterone account. Join the waitlist if you don't have one yet.
A GitHub account. You can register here.
Python 2.7 or 3.5+
The Clusterone Python package. Install it with
pip install clusterone.
The Clusterone command line interface, called
just, is installed automatically with the Clusterone Python package. Clusterone also provides a graphical web interface, the Matrix.
Linking your GitHub account allows you to access GitHub repositories from within Clusterone. To do this, you need to create a GitHub access token and add it to your Clusterone account.
Log into your GitHub account and navigate to the Personal Access Tokens page in the developer settings. Generate a new token and grant it the
Copy the token when it's created.
Log into your Clusterone account and open the Matrix. On the Account page, select the Keys tab. Click the Add GitHub OAuth Token button and paste the access token you created above. Click Save to store the token.
Perfect, you have successfully linked your GitHub account to Clusterone.
For more information on linking GitHub to Clusterone, see here.
Log into the Matrix and toggle the switch on the left to show your projects. Click the Add Project button and select Link GitHub Repository in the wizard:
On the next screen, type
clusterone/self-driving-demo to find the repository. Click the button at the bottom right to create the project.
To learn more about other ways to create a project, see here.
For the self-driving car example, you don't have to worry about creating a dataset. We've already uploaded the data for you.
To learn more about how to use data with Clusterone, see here.
Open a command line and log into your Clusterone account:
If this command fails or
just isn't recognized by your command line, make sure the Clusterone Python package is installed and has been added to your PATH.
Next, create a job:
just create job distributed --project self-driving-demo --module main_tf \--datasets tensorbot/self-driving-demo-data --ps-type c4.2xlarge \--worker-type c4.2xlarge --name first-job
Let's go over the parameters:
Here we are creating a distributed job, meaning that we're using multiple GPUs in parallel. If you'd rather run your code on a single machine, use
just create job single ... instead.
--project parameter determines the project you want to run code from. You can only run code from one project per job.
--dataset parameter can accept multiple datasets if your code uses them. You can also omit it if you don't want to use any dataset.
--module parameter is used to define which Python file Clusterone should execute. If this parameter is not provided, Clusterone assumes the file is called
main.py. In the self-driving car example, our module is called
main_tf.py, so we have to set
Clusterone offers a variety of different machines to run jobs on. The type of machine is defined by the
--worker-type parameters for parameter servers and worker machines respectively. In case you want to run on a single machine, use the
--instance-type parameter. See here for a list of available instance types.
--name parameter is used to give the job a name. Use this name to refer to the job in the
just start job command below.
Finally, all that's left to do is starting the job:
just start job -p self-driving-demo/first-job
-p parameter determines which job to start.
As soon as the job is started, it will gather the necessary resources and run once all resources are available.
You can follow the progress of your job on the Matrix. Click the "See Details" button under the name of your job to see how it's doing.
The "Events" tab provides a graphical representation of the startup progress of the job. Four circles allow you to see at a glance if your job has gathered all the resources it needs, or what is still missing.
The Creation Status tells you if the job has been created.
The Computational Requirements circle lists all required workers and parameter servers. It also contains information if the workers are running or if your job is still waiting for workers to become available.
The Code Cloning circle tell you if the repository code has been successfully cloned onto the worker machines.
The Process Start-Up circle represents the overall status of the job. Once the job has started, it will say "Running".
The "Outputs" tab contains a list of all raw output files that are generated while running the job. Here you can find the log files for each worker, event logs, and more. Click on each file to open it, or follow the download link on the right to download the file.
Clusterone provides direct access to TensorBoard, TensorFlow's suite of visualization tools.
To add your running job to TensorBoard, click the "Add to TensorBoard" button. Your job is now available on TensorBoard.
To access TensorBoard, click the TensorBoard button on the top bar. You can observe how well the model trains using the
Validation_Loss curves on the "Scalars" page of TensorBoard.
You can further examine a graph representation of the model on the "Graph" page.
In this guide, you have learned how to set up your first project on Clusterone, how to run it, and how you can examine its results. What's next?
If you're looking for another use case example, you can follow our DCGAN tutorial. In this more complex example, we run a Deep Convolutional GAN and generate artificial celebrity faces based on the celebA dataset.
If you want to learn more about a specific part of Clusterone, check out our Documentation Homepage with articles on all the details of running state-of-the-art distributed machine learning models on Clusterone.
Or jump right in and run your own project. If you have any comments, questions, or concerns, please don't hesitate to contact us, we'd love to hear from you!
Join our Slack to get support and tips from the community.