Skip to content

Latest commit

 

History

History
65 lines (49 loc) · 4.49 KB

README.md

File metadata and controls

65 lines (49 loc) · 4.49 KB
Raw

Swarm Learning Examples

Several examples of using Swarm Learning are provided. These examples use different models, datasets, ML platforms, and Swarm cluster configurations.

For details of running each example, see the below:

Prerequisite for all examples

  1. Start license server and install valid license before running any of the examples. Refer Installing licenses and starting license server.

  2. Install the Swarm Learning product using the Web UI installer. Refer Web UI installation

    For more information on starting license server and installing the Swarm Learning, see HPE Swarm Learning Installation and Configuration Guide.

  3. All examples require valid X.509 certificates to be used by different Swarm Learning components.

    A certificate generation utility (gen-cert) is provided with each example to enable users to run the examples quickly, using self-signed certificates.

    ./swarm-learning/examples/utils/gen-cert -h
Usage: gen-cert -e EXAMPLE-NAME -i HOST-INDEX
        -e Name of the example e.g. mnist, mnist-pyt, cifar10 etc.
        -i Host index like 1,2 etc.
        -h Show help.
    NOTE: Spire based example is automated and doesn't need any certificates to be generated. Spire agent and server containers will manage the certificates internally.
    NOTE: HPE recommends that users must use their own certificates in actual production environment.

How to resolve "ERROR: Invalid requirement: 'swarmlearning==client': Expected end or semicolon" error while building user container?

Latest version of pip has become stricter, especially in handling package names and version specifiers. You can resolve this error by downgrading pip to 24.0 in dockerfile or build task yaml file. You can check dockerfile of any of the 'breakhis', 'cancer-pred', 'cataract-detection' or 'nail-disease' examples for reference.

System setup for the examples

  1. The instructions in these examples assume that Swarm Learning runs on 1 to 2 host systems.

    • These systems have IP addresses 172.1.1.1 and 172.2.2.2.

    • License server is installed on 172.1.1.1 and running on default port 5814.

    • 172.1.1.1 runs the Sentinel SN node.

    • MNIST example has one more SN node. 172.2.2.2 runs the other SN node. A SL and ML node pair are spawned across these 2 host systems - 172.1.1.1 and 172.2.2.2.

    • Spawning of SL and ML nodes can be automatic or manual.

    • SWOP node, which runs on each host system, is automatically spawned using the task framework. MNIST example shows how Swarm training can be automatically launched using SWCI and SWOP nodes. SWCI runs on 172.1.1.1 host, while Swarm Operator runs on both 172.1.1.1 and 172.2.2.2 hosts.

    • CIFAR-10 example shows how Swarm training can be manually launched using run-sl script on each host.

    • Model training starts once minimum number (minPeers) of specified ML nodes are launched, either automatically or manually.

    • After training, the final Swarm model is saved by each ML node.

  2. The files created under the workspace directory that include certs, models, data, etc., are expected to have the minimum file permission as 664. Once the files are copied to the workspace directory, check the permissions of the files inside it. If desired permissions are not met, user might observe file permission denied in the respective swarm component's Docker logs. To overcome such cases, user can change the permissions of the files by using the chmod command.

  3. Finally, these instructions assume that swarm-learning is the current working directory on all the systems: cd swarm-learning.