Onnx for FSharp

This is an experimental F# API for Onnx.

The problem this library aims to solve is to combine the ease of use of an Eager Execution API with the speed of a Graph API. The API is code generated from the Onnx operator definitions.

It's possible to write Onnx graphs directly from high level F# functions which mix in normal code (that is evaluated at transform time) and Onnx functions.

Example

type on = FSharp.ML.Onnx.API.Snake.Onnx

type MNIST() = 
    let p193 = getTensorF("Parameter193", [|16L; 4L; 4L; 10L|])
    let p87  = getTensorF("Parameter87",  [|16L; 8L; 5L; 5L|])
    let p5   = getTensorF("Parameter5",  [|8L; 1L; 5L; 5L|])
    let p6   = getTensorF("Parameter6", [|8L; 1L; 1L|])
    let p88  = getTensorF("Parameter88", [|16L; 1L; 1L|])
    let p194 = getTensorF("Parameter194", [|1L; 10L|]) 

    [<ReflectedDefinition>]
    member this.Layer(x:Tensor<float32>,p1,p2,k) = 
       on.max_pool(on.relu(on.add(on.conv(x,p1,auto_pad = "SAME_UPPER"),p2)),kernel_shape = [|k;k|], strides = [|k;k|]) |> fst

    [<ReflectedDefinition>]
    member this.Forward(x: Tensor<float32>) = 
        on.add(on.mat_mul(on.reshape((this.Layer(this.Layer(x,p5,p6,2L),p87,p88,3L)),[|1;256|]),on.reshape(p193,[|256;10|])),p194)

DONE

• Eager Execution API for snake and pascal
• F# Quotation -> Onnx transform
• at transform time - support things like static conditionals which are useful for graph construction
• Support symbolic ops as OnnxFunctions
  • this is different to normal functions which are expanded in the transform

TODO

• fix tests 'converted graph' and 'defining new functions'

Limitations

• Eager Execution needs to clone memory which penalizes performance
• Normal quotation limitations apply for the F# Quotation -> Onnx Graph transform
• The generic arguments in the expression only support a limited subset of structual typing
  • Tuples and Records of Tensors of types Int32, Int64, Float32, or Float64
  • Arrays are not supported
• Not all Onnx Ops are supported at this time

Out of scope

• Nuget package
• Partial graph evaluation - this requires external knowledge of the generated graph which breaks the abstraction
• C++ wrapper interface for Onnx
• Onnx training
• Code Analyzer to check code is transformable at design time
• Encapsulate Tensor type to support operator overloads
• Shape Analyzer to graph has valid shapes at design time