skimit / orbital

Distribute private resources, such as machine learning models, through AWS.

PyOrbital

Distribute private resources, such as machine learning models, through AWS.

Motivation

Sputnik is a great library that manages data packages for another library, e.g. trained models for a machine learning library. However, Sputnik assumes packages will be hosted behind a webserver, which creates a fair bit of scaffolding work. We would like data packages to live on Amazon S3 instead.

Installation

pip install sputnik-orbital

Usage

Please refer to Sputnik's README for full details on how to structure a package so it can be managed by Sputnik. Essentially, the process is:

1. Create a resource (on machine A)
2. Publish resource (on machine A)
3. Install resource (on machine B)

A full example can be found in orbital/test/test_orbital.py.

Creation

Write your data resource as follows:

.
└── sputnik_sample
    ├── data
    │   └── model.pkl
    └── package.json

Here, model.pkl is the model that we want to distribute, and package.json is a manifest containing metadata about the model, e.g.

{
    "name": "orbital_test_model",
    "description": "This is a demo model, but it is still awesome.",
    "include": [["data", "*"]],
    "version": "2.0.0",
    "license": "Proprietary",
    "compatibility": {
        "my_library": ">=1.1.1"
    }
}

Then build the package for distribution:

from orbital import sputnik

package = sputnik.build("sputnik_sample")

Note we do not import Sputnik directly, but through Orbital. This applies the patches needed to use S3 as the storage layer.

Publishing

from orbital import sputnik

sputnik.upload("myapp", "1.0.0", package.path)

This uploads the package to an S3 bucket. This can be public or private.

Installation

from orbital import sputnik

sputnik.install("my_library", "1.0.0", "orbital_test_model==2.0.0")

This downloads and unpacks the required model version into a local directory.

Use installed model

package = sputnik.package("my_library", "1.1.3.", "orbital_test_model==2.0.0")
path_to_load = package.file_path(model_file_name)

Then load the model as usual, e.g. pickle.

S3 setup

Orbital does not create the S3 bucket where resources will be stored. You have to do that manually. The name of the bucket has to be provided as an environment variable to the upload script, e.g.

BUCKET="my_private_s3_bucket" python upload_all_models.py

To upload to a private bucket, you also need to create an AWS IAM use and give them R/W access to the bucket. Provide the user's credentials to your script, as described in the boto tutorial. The easiest thing to do is to specify the credentials as environment variables, e.g.

AWS_ACCESS_KEY_ID=AAAA AWS_SECRET_ACCESS_KEY=BBB BUCKET="my_private_s3_bucket" python upload_all_models.py

Alternatively, put the credentials in ~/.aws/credentials or ~/.boto.

Running tests

PYTHONPATH=. py.test