Basic Setup

Run Weights and Biases on your own machines using Docker.

Installation

On any machine with Docker and Python installed, run:

pip install wandb
wandb server start

Login

If this is your first time logging in then you will need to create your local W&B server account and authorize your API key. There are several ways to control where your runs are logged to. This is particularly useful if you run wandb on multiple machines or you switch between a private instance and W&B cloud.

Follow the procedure below to send metrics to the shared private instance. Ensure you have already set up DNS:

1. Set the host flag to the address of the private instance whenever you login:

 wandb login --host=http://wandb.your-shared-local-host.com

2. Set the environment variable WANDB_BASE_URL to the address of the local instance:

export WANDB_BASE_URL = "http://wandb.your-shared-local-host.com"

In an automated environment, you can set the WANDB_API_KEY. Find your key at wandb.your-shared-local-host.com/authorize.

Set the host to api.wandb.ai to log to the public cloud instance of W&B:

wandb login --cloud

or

export WANDB_BASE_URL = "https://api.wandb.ai"

You can also switch to your cloud API key, available at https://wandb.ai/settings when you are logged in to your cloud-hosted wandb account in your browser.

Generate a free license

You need a license to complete your configuration of a W&B server. Open the Deploy Manager to generate a free license. If you do not already have a cloud account then you will need to create one to generate your free license. You can generate either a personal or team or free license:

1. Personal licenses are free forever for personal work:
2. Team trial licenses are free and last 30 days, allowing you to set up a team and connect a scalable backend:

Add a license to your Local host

1. Copy your license from your Deployment and navigate back to your W&B server's localhost:
2. Add it to your local settings by pasting it into the /system-admin page of your localhost:

Upgrades

New versions of wandb/local are pushed to DockerHub regularly. We suggest you keep your version up to date. To upgrade, copy and paste the following command into your terminal:

$ wandb server start --upgrade

Alternatively, you can upgrade your instance manually. Copy and paste the following code snippets into your terminal:

$ docker pull wandb/local
$ docker stop wandb-local
$ docker run --rm -d -v wandb:/vol -p 8080:8080 --name wandb-local wandb/local

Persistence and Scalability

• All metadata and files sent to W&B server are stored in the /vol directory. If you do not mount a persistent volume at this location all data will be lost when the docker process dies.
• This solution is not meant for production workloads.
• You can store metadata in an external MySQL database and files in an external storage bucket.
• The underlying file store should be resizable. Alerts should be put in place to let you know once minimum storage thresholds are crossed to resize the underlying file system.
• For enterprise trials, we recommend at least 100GB free space in the underlying volume for non-image/video/audio heavy workloads.

How does wandb persist user account data?

When a Kubernetes instance is stopped, the W&B application bundles all the user account data into a tarball and uploads it to the Amazon S3 object store. W&B pulls previously uploaded tarball files when you restart an instance and provide the BUCKET environment variable. W&B will also load your user account information into the newly started Kubernetes deployment.

When an external object store is enabled, strong access controls should be enforced as it will contain all users data. W&B persists instance settings in the external bucket when it is configured. W&B also persist certificates, and secrets in the bucket.

Create and scale a shared instance

To enjoy the powerful collaborative features of W&B, you will need a shared instance on a central server, which you can set up on AWS, GCP, Azure, Kubernetes, or Docker.

caution

Trial Mode vs. Production Setup

In Trial Mode of W&B Local, you run the Docker container on a single machine. This setup is ideal for testing the product, but it is not scalable.

For production work, set up a scalable file system to avoid data loss. We suggest you:

• allocate extra space in advance,
• resize the file system proactively as you log more data
• configure external metadata and object stores for backup.

The instance will stop working if you run out of space. In this case, any additional data will be lost.

Contact sales to learn more about Enterprise options for W&B server.