Microsoft ONNXRuntime AI and Machine Learning Library for Freepascal / Delphi

Introduction

This is an implementation of Microsoft's Open Neural Network Exchange (ONNXRuntime) for Freepascal 🐾 and Delphi ⚔️

ONNXRuntime libraries comes shipped with most of modern Windows releases after Windows 8, as of the time this is written, version 1.13.1 is the most recent release, it can be installed on MacOS and most of Linux releases, for development and updates please visit ONNXRuntime Github Page.

How to install libraries

Windows

onnxruntime.dll is already shipped with windows, you can find it in %WINDIR%\SysWOW64\onnxruntime.dll or%WINDIR%\System32\onnxruntime.dll

MacOS and linux

check https://github.com/microsoft/onnxruntime/releases

Usage

From your Lazarus or Delphi project at the header of the pascal unit include the files

  unit formUnit; 
  {$h+}

  interface
  uses onnxruntime_pas_api, onnxruntime, Classes etc... ;

Load a Model

  var 
    session : TORTSession;
  begin
    session := TORTSession.Create('./mymodel/filname.onnx'); 
  { 
  *****************************************************************
      Check your model requirements for input/output 
      names and value dimensions before preparing the inputs.
      to explore the model before preparing use :
        session.GetInputCount and session.GetOutputCount
        session.GetInputName and session.GetOutputName
        session.GetInputTypeInfo and session.GetOutputTypeInfo
   ****************************************************************
  }

Prepare an input tensor with the desired shape using TORTTensor<type> and your inputs using TORTNameValueList

var 
  x,y:integer;
  imageData : array of array of single;
  inTensor : TORTTensor<single> ; 
  inputs : TORTNameValueList  ;
begin
  // assuming the model input name is 'image' and the tensor shape is [width, height]
  inTensor := TORTTensor<single>.Create([width, height{, depth ,etc...}]);
  for y:=0 to inTensor.shape[1]-1 do
      for x:=0 to inTensor.shape[0]-1 do
          inTensor.index2[x, y]:= imageData[x, y];  // your float values
  inputs['image'] := inTensor;        

Inference

  var
    myDetection : array of single;
    i:integer;
    outputs : TORTNameValueList;
    outTensor : TORTTensor<single>
  begin 
      outputs   := session.run(inputs);
      outTensor := outputs['result']

      for i:=0 to outTensor.shape[0] do
        myDetection[i] := outTensor.index1[i]

Training

More examples Coming soon..

Examples

• CPU : Faster RCNN10 example folder

  Download FasterRCNN-10.onnx from here

• GPU : Yolo V7 (DirectML) folder

  Download and extract yolov7_640x640.onnx from here

More information about ONNXRuntime API

• Check ONNXRuntime API Documents

Contributions and suggestions are most welcome.