An operating system for machine learning

Installation

pip install modelos

Configuring

ModelOS requires and OCI compliant image registry. This can be set using any of:

• $MDL_IMAGE_REPO env var
• tool.modelos.image_repo in pyproject.toml
• image_repo in mdl.yaml file at project root
• image_repo function parameters

ModelOS requires a Kubernetes cluster >= 1.22.0 and will use your current kubeconfig context.

Usage

NOTE: ModelOS is pre-alpha and under heavy development, expect breakage

Objects

Objects are distributed persistent Python objects.

A sample object

frommodelosimportObjectclassHam(Object):
temp: intweight: intbrand: strdef__init__(self, weight: int, brand: str) ->None:
self.weight=weightself.brand=branddefbake(self, temp: int) ->None:
self.temp=tempdefcut(self, weight: int) ->int:
self.weight=self.weight-weightreturnself.weight

ModelOS objects work just like Python objects with some extra capabilities.

Developing Objects

Develop on the object remotely

# Create a client which can be used to generate remote instancesHamClient=Ham.client(hot=True)
# Creates a remote instance within the context then deletes itwithHamClient(weight=12, brand="black forest") asham:
ham.bake(temp=340)
# Store the objecturi=ham.store()

Example develop URI: acme.org/ml-project:obj.ham.f2oij-2oijr-f8g2n

Releasing Objects

Release the object to be used by others. This creates a semver for the object and client/server packages

HamClient=Ham.client()
# Creates a remote instance within the context then deletes itwithHamClient(weight=14, brand="honey baked") asham:
ham.bake(temp=340)
uri=ham.release()

Example release URI: acme.org/ml-project:obj.ham.v1.2.3

Versioning scheme

v1: interface version
v1.2: class version
v1.2.3: instance version

Using Objects

Install a client from a release and use it to generate a remote instance

frommodelosimportinstallinstall("acme.org/ml-project:obj.ham.v1")
fromml_project.ham.v1importHamClient# Use the latest releasewithHamClient(weight=10) asham:
ham.bake(temp=320)
# Specify a class release versionHamClient.version="v1.2"withHamClient(weight=10) asham:
ham.bake(temp=320)
# Specify an instance versionwithHamClient.instance("v1.2.3") asham:
ham.bake(temp=320)

Install a class and use locally or remotely

frommodelosimportinstallinstall("acme.org/ml-project:obj.ham.v1.2")
fromml_project.ham.v1_2importHam# locallyham=Ham(weight=12)
ham.bake(temp=300)
# remotelywithHam.client()(weight=10) asham:
ham.bake(temp=320)

Install an instance and use it locally

frommodelosimportinstallinstall("acme.org/ml-project:obj.ham.v1.2.3")
fromml_project.ham.v1_2_3importHam# load the object instance stateham=Ham.from_env()
ham.bake(temp=400)

An example working project can be found at https://github.com/pbarker/kvd

Packages

Packages are versioned filesystems

frommodelos.pkgimportPkg, clean# Create a new package from the ./data dirpkg=Pkg("foo", "./data", "A foo package", remote="acme.org/ml-project")
# See package contentspkg.show()
# Push the package to its remotepkg.push()
# List files in the packagefiles=pkg.ls()
# Open a file in the packagewithpkg.open("./foo.yaml") asf:
b=f.read()
# Release the packagepkg.release("v0.0.1", labels={"type": "foo"}, tags=["baz"])
# Check the latest package is our releaseassertpkg.latest() =="v0.0.1"# Delete the packagepkg.delete()
# Describe a remote packageinfo=Pkg.describe("acme.org/ml-project:pkg.fs.bar.v1.2.3")
# Use a remote packagebar_pkg=Pkg("bar", version="v1.2.3", remote="acme.org/ml-project")
# Clean packagesclean("acme.org/ml-project", "foo")

See the tests for more examples

Roadmap

• Releasing
• Properties
• Packages
• Environments
• Runtimes
• Extension objects
• Finding / indexing
• Docs / landing
• UI / CLI
• Schema
• Bi-directional streaming
• Smarter versioning