2 unstable releases

0.2.0 Nov 17, 2020
0.1.0 May 27, 2019

#29 in Machine learning

MIT license

37KB
784 lines



Orkhon: ML Inference Framework and Server Runtime

Latest Release Crates.io
License Crates.io
Build Status Build Status
Downloads Crates.io
Gitter

What is it?

Orkhon is Rust framework for Machine Learning to run/use inference/prediction code written in Python, frozen models and process unseen data. It is mainly focused on serving models and processing unseen data in a performant manner. Instead of using Python directly and having scalability problems for servers this framework tries to solve them with built-in async API.

Main features

  • Sync & Async API for models.
  • Easily embeddable engine for well-known Rust web frameworks.
  • API contract for interacting with Python code.
  • High processing throughput
    • ~4.8361 GiB/s prediction throughput
    • 3_000 concurrent requests takes ~4ms on average
  • Python Module caching

Installation

You can include Orkhon into your project with;

[dependencies]
orkhon = "0.2"

Dependencies

You will need:

  • Rust Nightly needed (for now. until async support fully lands in)
  • Python dev dependencies installed and have proper python runtime to use Orkhon with your project.
  • Point out your PYTHONHOME environment variable to your Python installation.

Python API contract

For Python API contract you can take a look at the Project Documentation.

Examples

Request a Tensorflow prediction asynchronously

 use orkhon::prelude::*;
 use orkhon::tcore::prelude::*;
 use orkhon::ttensor::prelude::*;
 use rand::*;
 use std::path::PathBuf;

let o = Orkhon::new()
    .config(
        OrkhonConfig::new()
            .with_input_fact_shape(InferenceFact::dt_shape(f32::datum_type(), tvec![10, 100])),
    )
    .tensorflow(
        "model_which_will_be_tested",
        PathBuf::from("tests/protobuf/manual_input_infer/my_model.pb"),
    )
    .shareable();

let mut rng = thread_rng();
let vals: Vec<_> = (0..1000).map(|_| rng.gen::<f32>()).collect();
let input = tract_ndarray::arr1(&vals).into_shape((10, 100)).unwrap();

let o = o.get();
let handle = async move {
    let processor = o.tensorflow_request_async(
       "model_which_will_be_tested",
       ORequest::with_body(TFRequest::new().body(input.into())),
    );
    processor.await
};
let resp = block_on(handle).unwrap();

Request an ONNX prediction synchronously

This example needs onnxmodel feature enabled.

use orkhon::prelude::*;
use orkhon::tcore::prelude::*;
use orkhon::ttensor::prelude::*;
use rand::*;
use std::path::PathBuf;

 let o = Orkhon::new()
     .config(
         OrkhonConfig::new()
             .with_input_fact_shape(InferenceFact::dt_shape(f32::datum_type(), tvec![10, 100])),
     )
     .onnx(
         "model_which_will_be_tested",
         PathBuf::from("tests/protobuf/onnx_model/example.onnx"),
     )
     .build();

 let mut rng = thread_rng();
 let vals: Vec<_> = (0..1000).map(|_| rng.gen::<f32>()).collect();
 let input = tract_ndarray::arr1(&vals).into_shape((10, 100)).unwrap();

 let resp = o
     .onnx_request(
         "model_which_will_be_tested",
         ORequest::with_body(ONNXRequest::new().body(input.into())),
     )
     .unwrap();
 assert_eq!(resp.body.output.len(), 1);

License

License is MIT

Documentation

Official documentation is hosted on docs.rs.

Getting Help

Please head to our Gitter or use StackOverflow

Discussion and Development

We use Gitter for development discussions. Also please don't hesitate to open issues on GitHub ask for features, report bugs, comment on design and more! More interaction and more ideas are better!

Contributing to Orkhon Open Source Helpers

All contributions, bug reports, bug fixes, documentation improvements, enhancements and ideas are welcome.

A detailed overview on how to contribute can be found in the CONTRIBUTING guide on GitHub.

Dependencies

~19MB
~398K SLoC