Differentiable Programming Manifesto

• Authors: Richard Wei, Dan Zheng, Marc Rasi, Bart Chrzaszcz
• Status: Partially implemented on master, feature gated under import _Differentiation

Table of contents

• Introduction
• Motivation
  • Background
  • Intelligent applications
  • Type-safe machine learning
  • Calculus is fun
    • Animations
    • Games
    • Simulations
    • Robotics
    • Rendering and ray tracing
• History of differentiation algorithms
  • Numerical differentiation
  • Symbolic differentiation
  • Automatic differentiation
• Approaches to automatic differentiation
  • Embedded domain-specific languages
  • Source code transformation tools
  • First-class language support
  • Why bake differentiation into Swift?
    • Maximal coverage of Swift language features
    • Extensibility
    • Static warnings and errors
    • The pursuit for user-defined code transformations
• Math introduction
  • What is a derivative?
  • Iterative optimization
  • Derivatives of functions with arbitrary inputs
• Proposed solution
  • The Differentiable protocol
  • The @differentiable declaration attribute
  • @differentiable function types
  • @derivative and @transpose attributes
  • Differential operators
• Detailed design
  • Differentiable data structures
    • The Differentiable protocol
    • Differentiable conformances
    • Compiler-synthesized conformances
      • Synthesis conditions
      • Default synthesis
      • Shortcut synthesis
  • Differentiable function declarations
    • The @differentiable declaration attribute
    • Conformance and subclassing
      • Protocol dispatch
      • Class dispatch
  • Make a function differentiable using @derivative or @transpose
    • Linear maps
      • Typing rules
        • Linearity parameters
        • Top-level functions
        • Static methods
        • Instance methods
        • Linearity generic requirements
      • Examples
    • Derivative functions
      • Typing rules
        • Differentiability parameters
        • Differentiability generic requirements
      • Examples
  • Differentiable function types
    • Function subtyping and runtime representation
    • The @differentiable function type attribute
    • Type conversion
      • Coercing function declarations into @differentiable function values
      • Upcasting to non-@differentiable functions
    • Implied generic constraints
    • Non-differentiable parameters
    • Higher-order functions and currying
  • Differential operators
    • Differential-producing differential operators
    • Pullback-producing differential operators
    • Example usage
    • List of differential operators
    • Static analysis
      • Cross-module opacity
      • Non-differentiable type conversions
      • Accidental data flow mistakes
• Examples of differentiable programming
  • Linear regression
  • Deep learning
    • Feed-forward neural networks (FFNN)
    • Convolutional neural networks (CNN)
    • Recurrent neural networks (RNN)
• Future directions
  • Higher-order differentiation
  • Naming conventions for numerical computing
• Source compatibility
  • Differentiable protocol
  • Differential operators
• Alternatives considered
  • Not support differentiable programming
  • Use another language or framework for differentiable programming
  • Other approaches to differentiable programming
• Acknowledgements

Introduction

This proposal introduces first-class differentiable programming to Swift. First-class differentiable programming includes five core additions:

• The Differentiable protocol.
• @differentiable function types.
• The @differentiable declaration attribute for defining differentiable functions.
• The @derivative and @transpose attributes for defining custom derivatives.
• Differential operators (e.g. derivative(of:)) in the standard library.

Differentiable programming is a new paradigm for programming in which programs can be differentiated throughout. At a glance, differentiable programming lets you take the derivative of functions whose parameters and results conform to the Differentiable protocol.

@differentiable
func f(_ x: Float) -> Float {
    x * x
}
let dfdx = derivative(of: f)
dfdx(3) // 6

The ability to get derivatives of programs enables a new world of numerical computing applications, notably machine learning. With first-class support, gradient-based learning algorithms can even be built using standard library types such as Float and SIMD64<Float> and be differentiated using protocol-oriented APIs such as valueWithGradient(at:in:).

struct Perceptron: @memberwise Differentiable {
    var weight: SIMD2<Float> = .random(in: -1..<1)
    var bias: Float = 0

    @differentiable
    func callAsFunction(_ input: SIMD2<Float>) -> Float {
        (weight * input).sum() + bias
    }
}

var model = Perceptron()
let andGateData: [(x: SIMD2<Float>, y: Float)] = [
    (x: [0, 0], y: 0),
    (x: [0, 1], y: 0),
    (x: [1, 0], y: 0),
    (x: [1, 1], y: 1),
]
for _ in 0..<100 {
    let (loss, 𝛁loss) = valueWithGradient(at: model) { model -> Float in
        var loss: Float = 0
        for (x, y) in andGateData {
            let ŷ = model(x)
            let error = y - ŷ
            loss = loss + error * error / 2
        }
        return loss
    }
    print(loss)
    model.weight -= 𝛁loss.weight * 0.02
    model.bias -= 𝛁loss.bias * 0.02
}

Differentiable programming scales up to full machine learning models, built with third-party libraries like TensorFlow.

import TensorFlow

let model = Sequential {
    Dense<Float>(inputSize: 784, outputSize: 100, activation: relu)
    Dense<Float>(inputSize: 100, outputSize: 30, activation: relu)
    Dense<Float>(inputSize: 30, outputSize: 3, activation: identity)
}

var classifier = Model()
let optimizer = SGD(for: classifier, learningRate: 0.02)
Context.local.learningPhase = .training
let x: Tensor<Float> = ...
let y: Tensor<Int32> = ...

for _ in 0..<1000 {
    let 𝛁model = gradient(at: classifier) { classifier -> Tensor<Float> in
        let ŷ = classifier(x)
        let loss = softmaxCrossEntropy(logits: ŷ, labels: y)
        print("Loss: \(loss)")
        return loss
    }
    optimizer.update(&classifier, along: 𝛁model)
}

While the differentiation APIs are flexible and fully dynamic, differentiation is based on a program transformation that happens at compile-time. This enables many static analyses that not only help produce more efficient programs, but also detect common numerical programming mistakes such as non-differentiable functions and zero derivatives.

let grad = gradient(at: 1.0) { x in
    3.squareRoot()
}

test.swift:2:18: warning: result does not depend on differentiation arguments and will always have a zero derivative; do you want to add 'withoutDerivative(at:)' to make it explicit?
    3.squareRoot()
    ^
     withoutDerivative(at:)

With a first-class differentiable programming language, some of the most common runtime errors in machine learning become directly debuggable without library boundaries. Simply step through backpropagation using LLDB to debug derivatives.

Motivation

Background

In mathematics, a derivative of a function of a real variable is another function that computes the sensitivity to changes in the output of the original function with respect to changes in the original function's arguments. Differentiation is the process of computing derivatives. See the “Math Introduction” section below for more details.

Derivatives are a fundamental tool in calculus and have applications in many domains, notably deep learning. As an expressive, high-performance language, Swift is a great fit for numerical applications. The Swift Numerics library and recent Swift Evolution proposals have paved the way for low-level numerical computing in Swift: AdditiveArithmetic, SIMD [1] [2], generic math functions. However, high-level numerical computing applications, including machine learning and artificial intelligence, require more work.

We believe that first-class differentiable programming is a big step towards high-level numerical computing support and will make Swift a real contender in the numerical computing and machine learning landscape. Differentiable programming will enable intelligent applications, machine learning models, scientific experiments, physical simulations, and more.

Intelligent applications

Intelligent applications are smart: they use machine learning techniques to enhance user experiences. Intelligent applications can make predictions, provide suggestions, and learn user preferences: all of these can be powered by differentiable programming.

The core of an intelligent application is a function with real-valued parameters. Differentiation can be used to systematically optimize (i.e. find “good” values for) these parameters via gradient descent. (Optimizing these parameters via conventional algorithms is typically difficult or intractable.)

For example, consider a podcast player that tries to automatically adjust the playback speed based on the podcast type and the podcast section.

enum PodcastCategory {
    case comedy
    case news
    ...
}

enum PodcastSection {
    case advertisement
    case introduction
    case body
    case conclusion
}

struct PodcastState {
    let category: PodcastCategory
    let section: PodcastSection
}

struct PodcastSpeedModel {
    var minSpeed, maxSpeed: Float
    var categoryMultipliers: [PodcastCategory: Float]
    var sectionMultipliers: [PodcastSection: Float]

    /// Returns a podcast speed multiplier prediction for the given podcast category
    /// and section.
    func prediction(for state: PodcastState) -> Float {
        let speed = categoryMultipliers[state.category] * sectionMultipliers[state.section]
        if speed < minSpeed { return minSpeed }
        if speed > maxSpeed { return maxSpeed }
        return speed
    }
}

This podcast speed model parameters that determine how quickly the podcast should play under different circumstances: minSpeed, maxSpeed, categoryMultipliers, and sectionMultipliers. A priori, it is not clear what good parameter values are, and different users may prefer different parameter values.

An intelligent application could determine personalized parameter values as follows:

1. Let the user set the speed manually, and record observations whenever the user changes the speed.

2. After collecting enough observations, search for parameter values such that the model predicts speeds close to the user's preferred speed. If such values are found, offer to start automatically setting the speed.

“Gradient descent” is an algorithm that performs this search, and a language that supports differentiable programming makes it easy to implement gradient descent. Here is some pseudocode illustrating gradient descent.

First, we need an objective function for gradient descent to minimize. Mean absolute error is used here:

struct Observation {
    var podcastState: PodcastState
    var userSpeed: Float
}

func meanError(for model: PodcastSpeedModel, _ observations: [Observation]) -> Float {
    var error: Float = 0
    for observation in observations {
        error += abs(model.prediction(for: observation.podcastState) - observation.userSpeed)
    }
    return error / Float(observations.count)
}

Next, we implement the gradient descent algorithm.

var model = PodcastModel()
let observations = storage.observations()
for _ in 0..<1000 {
    // The language differentiates `meanError` to get a "gradient", which is a value indicating
    // how to change `model` in order to decrease the value of `meanError`.
    let gradient = gradient(at: model) { meanError(for: $0, observations) }

    // Change `model` in the direction that decreased the value of `meanError`.
    model -= 0.01 * gradient
}

Type-safe machine learning

Today, machine learning is predominantly done in dynamically-typed languages like Python: these languages are concise and easy to use. However, some people prefer safer programming: features like type checking and static diagnostics help catch errors early and improve productivity.

Differentiable programming in Swift enables safe, expressive machine learning. Custom differentiable data structures can be declared and checked at compile-time. Thanks to protocol-oriented programming, differentiable types are generalized by a protocol, enabling differential operators to be defined as higher-order functions constrained on such a protocol. Mathematical optimization algorithms such as neural network optimizers can also be defined generically over such a protocol and work with all differentiable types.

Calculus is fun

Calculus is fun, and differentiation in the Swift toolbox will let programmers explore that fun. Here are some interesting applications:

Animations

Easing functions specify the rate of change of parameters for animations. Differentiation enables easy manipulation of these functions.

Games

Physics equations can be modeled using differentiable functions in game engines. Intelligent agents in games can be trained using techniques like machine learning that are enabled by differentiation.

Simulations

Many simulation techniques for fluids and other physical processes are based on approximate solutions to equations defined in terms of derivatives, like the Euler equations and Navier-Stokes. Being able to differentiate functions is an important building block for implementing algorithms to solve these equations.

Robotics

Control algorithms used in robotics and mechanical engineering rely on (often higher-order) derivatives of functions that model the behavior of joints and other physical systems. A language like Swift that can efficiently compute these derivatives without incurring the unpredictable runtime overhead of garbage collection may be well-placed to run aboard robots.

Rendering and ray tracing

Traditional rendering systems are black boxes that consume data structures with scene geometry and produce images, but the physical processes they simulate are made up of differentiable functions. Building a ray tracer out of differentiable building blocks unlocks applications like inverse rendering (going from an image to scene geometry). [1] [2]

History of differentiation algorithms

There are three main algorithms for computing derivatives: numerical differentiation, symbolic differentiation, and automatic differentiation.

Numerical differentiation

Numerical differentiation is a technique for estimating derivatives of mathematical functions using values of the functions. The simplest method uses the difference quotient formula, introduced in elementary calculus courses:

Numerical differentiation is easy to implement and generalizes to higher-order derivatives. However, as an estimation approach, it is known to produce inaccurate results, so it is rarely used when more accurate methods are available.

Symbolic differentiation

Symbolic differentiation is a technique for computing derivatives of math expressions via symbolic manipulation, like differentiating an expression using pen and paper in elementary calculus. This technique is used by computer algebra systems like Mathematica, but it produces inefficient code when applied to computer programs due to code bloat with common subexpressions.

Automatic differentiation

Automatic differentiation (AD) is a technique for computing derivatives of functions. Unlike symbolic differentiation, which operates on math expressions, automatic differentiation operates on code.

Automatic differentiation leverages the chain rule of differentiation and the ability to define temporary values in a program. There are two styles of automatic differentiation in the traditional sense: forward-mode AD starts with partial derivatives at inputs and ends by computing partial derivatives at outputs, while reverse-mode automatic differentiation starts with partial derivatives at outputs and ends by computing partial derivatives at inputs.

Mathematically, forward-mode AD corresponds to a fully-right association of the chain rule of differentiation, and reverse-mode AD corresponds to a fully-left association. Different associations of the chain rule produce the same result but may differ in computational complexity†.

Both forward-mode AD and reverse-mode AD are well-explored. Forward-mode AD can be implemented simply by overloading math operations to compute both original values and derivatives. Traditionally, reverse-mode AD has been perceived as being more complicated: implementations typically involve non-local program transformation and/or mutable tape data structures, though recent research aims to demystify the subject [1] [2].

†: Finding the optimal association of the chain rule of differentiation is analogous to the matrix chain multiplication problem and can be solved in O(n^3) time. More efficient algorithms also exist.

Approaches to automatic differentiation

In practice, automatic differentiation is the most common differentiation algorithm because it is precise and efficient. This section summarizes approaches to automatic differentiation.

Embedded domain-specific languages

A domain-specific language (DSL) is a language designed to solve problems for a specific domain. Some DSLs are external: these are standalone languages with their own syntax and semantics, like HTML (a markup language) and SQL (a database query language). Other DSLs are embedded within a more general “host” language: these DSLs leverage host language constructs and features to define interesting behavior. Advantages of embedded DSLs include flexibility and portability: embedded DSLs can be imported as a library. Examples of embedded DSLs include React (a UI language embedded in JavaScript) and LINQ (a query language embedded in C#).

One approach to differentiable programming is to define an embedded DSL for differentiation as a library. This can be done via operator overloading: the DSL can define a “dual number” type (representing a pair of a real number and its derivative) and overload differentiable math operations to compute both original values and derivative values.

struct RealWithDerivative<T: FloatingPoint> {
    var value: T
    var derivative: T = 0
}
extension RealWithDerivative {
    static func + (lhs: Self, rhs: Self) -> Self {
        RealWithDerivative(
            value: lhs.value + rhs.value,
            derivative: lhs.derivative + rhs.derivative)
    }
    static func * (lhs: Self, rhs: Self) -> Self {
        RealWithDerivative(
            value: lhs.value * rhs.value,
            derivative: lhs.derivative * rhs.value + lhs.value * rhs.derivative)
    }
}

var x = RealWithDerivative(value: 3, derivative: 1)
// Original:   x^2 + x^3 = 3^2 + 3^3 = 36.
// Derivative: 2x + 3x^2 = 2*3 + 3(3)^2 = 33.
var result = x*x + x*x*x
print(result)
// RealWithDerivative<Double>(value: 36.0, derivative: 33.0)

Such a DSL could be extended to be more useful. For example, the Real type could be generalized to multidimensional arrays and more differentiable operations could be added.

However, embedded DSLs have some limitations:

• DSL functionality is often restricted to specific types and APIs. DSLs often use specialized abstractions rather than general ones for simplicity and to enable optimizations. For example, many machine learning frameworks are DSLs that support differentiation only for a particular multidimensional array type and only using a particular algorithm (reverse-mode automatic differentiation). Extending a differentiation DSL beyond these limitations is difficult and may require extra boilerplate: see below.

• They typically involve some boilerplate. As a host language, Swift currently supports limited metaprogramming for reducing boilerplate code. For example, libraries cannot define automatic conformance derivation for library protocols (though Swift provides it for Equatable, Hashable, and Codable), so users must write boilerplate conformances for their custom types.

• They are limited by the metaprogramming capabilities of the host language. It is not currently possible to define non-trivial code transformations (e.g. reverse-mode automatic differentiation) in a Swift library on Swift code. (Note: SwiftSyntax enables Swift AST transformations but has the extra indirection of parsing Swift code from a file - it is not possible to evaluate transformed Swift code from the same file without a general “eval” mechanism.) To cope with this, some DSLs require explicit program “graph” building and/or global mutable data structures to mimic the effects of code transformation, which obfuscate the original transformation semantics.

• They may not work well with all host language constructs. Embedded DSLs only support a subset of the host language's features. In particular, some differentiation DSLs do not support native mutation (e.g. assigning to a var) or native control flow (e.g. if constructs) due to technical limitations, even though supporting them would be ideal. Restricting/diagnosing unsupported host language features (e.g. preventing DSL users from using var in Swift) is difficult or not possible.

• Producing good diagnostics may be difficult or impossible. DSLs have limited access to source location information. When indirections like code transformations are involved, showing the appropriate source locations in diagnostic messages may be difficult. Without the aid of compiler utilities, statically detecting and diagnosing dataflow-based errors is not possible.

Source code transformation tools

Source code transformation tools are another approach to differentiable programming. Tool users write code, select various differentiation configuration options (the name of the function-to-differentiate, the independent and dependent variable, etc), and provide them to the tool. The tool analyzes the input code and generates output code that computes derivatives according to the options.

Historically, this is one of the oldest approaches for automatic differentiation. Tools like Tapenade and ADIC/ADIFOR compute derivatives of Fortran and C code.

An advantage of source code transformation tools is that they are essentially static compilers: they can perform static analyses on input code to generate optimized derivative-computing output code. For example, Tapenade performs “activity analysis” to determine variables that do not need a derivative and “TBR (to-be-recorded) analysis” to remove unnecessary intermediate variables during differentiation.

However, these tools are not ideal for usability: users must interact with an external GUI to specify inputs and they receive a textual program as output. This external workflow is an extra indirection that takes users out of their natural programming environment. Exposing the tool-provided differentiation features within a language would be more ergonomic.

First-class language support

Another class of differentiable programming approaches is by integrating the differentiation semantics and code transformations into a programming language to some degree. While there are no mainstream programming languages that support differentiable programming, research systems like Stalin∇ add first-class differential operators (e.g. grad) into the language and the reverse-mode automatic differentiation transformation into the compiler.

First-class language support for differentiation can reap the benefits of source code transformation techniques (e.g. language coverage, performant derivative code) without requiring programmers to use an external tool. Well-designed, powerful differentiation primitives enable users to define their own custom differentiation APIs that would otherwise not be possible in differentiation libraries.

Why bake differentiation into Swift?

First-class language support for differentiation will enable convenient, extensible, and performant differentiable programming in Swift.

Maximal coverage of Swift language features

First-class support for differentiation in Swift enables differentiation to work nicely with a maximal number of Swift language features, including mutation and control flow. Users of differentiable programming do not need to write in a restricted subset of Swift: just write normal code and use differentiation.

Extensibility

First-class language support enables an extensible differentiable programming system.

Custom types can be extended to be differentiable with minimal boilerplate. Custom derivative functions can be retroactively registered for existing functions. Users can define custom differentiation APIs using the powerful primitive operators defined in the standard library and supported by the type system.

Static warnings and errors

Some functions perform non-differentiable operations (on the path from parameters to result) and thus cannot be differentiated. Functions that do not use their parameters to compute the result are technically differentiable, but the derivative is trivially always zero.

With language support for differentiation, the compiler can identify these cases statically via data flow analysis and produce a non-differentiability error or warning. These diagnostics improve productivity and help users catch errors ahead of time. Library-based differentiation approaches cannot generally provide these diagnostics.

For details on static warnings and errors, see the “Static analysis” section in the detailed design below.

The pursuit for user-defined code transformations

The key code transformation enabling differentiable programming is “derivative code generation”. Derivative code generation implements automatic differentiation: given an “original function” to differentiate, a derivative function is generated by replacing function applications in the original function with corresponding derivative function applications. The algorithm is described in detail in the Swift Differentiable Programming Implementation Overview document.

Some languages provide the ability to define custom code transformations:

• Macros enable syntax-based code transformations at compile-time. Hygienic macros (macro systems that avoid accidental variable capture) are available in a variety of languages, including Lisp, Julia, Rust, and Scala, to name a few. As an example: generated type-safe schema wrappers can implemented using hygienic macros in Scala.

• Compiler plugin systems enable programmers to write plugins that extend the behavior of a compiler. Compiler plugins are more popular in bootstrapped languages, like Haskell, Rust and Scala, where the plugin can be written in the language itself. As an example: a continuation-passing-style code transformation can be implemented as a compiler plugin in Scala.

One might make the case that derivative code generation for differentiation is better implemented as a custom code transformation. While that may be true in theory, Swift does not yet support custom code transformations in practice. This proposal presents differentiable programming as a system of high-level language features and semantics; derivative code generation is an implementation detail. If a system for custom code transformations is added to Swift one day, it may be possible to reimplement derivative code generation using that system without changing the high-level differentiable programming features proposed here.

Math introduction

What is a derivative?

The derivative of a function f measures how quickly the function‘s output changes when you make small changes to the function’s input. The value of this measurement depends on the input x that you start with, and we call the value of the measurement starting at that input "the derivative of f at x.

For a single variable real function (a function with a single real input and a single real output), the derivative of f at x can be summarized as a single real number f'(x) such that f(x + ε) ~= f(x) + f'(x) * ε. In other words, changing the input by a tiny amount epsilon changes the output by f'(x) * ε.

Iterative optimization

Iterative optimization algorithms use derivatives to optimize functions (i.e. find the inputs that minimize or maximize the output of the function). For example, the simple “gradient descent” algorithm starts with an arbitrary input x and uses the derivative of the function at x to determine whether it needs to increase or decrease x to decrease the output of the function. Then it mutates x slightly along the appropriate direction and repeats until the output stops decreasing.

Derivatives of functions with arbitrary inputs

Real world programs deal with data more complicated than single real variables. Fortunately, there are mathematical theories that extend derivatives to functions with nearly arbitrary inputs and outputs.

Recall our original description of derivative: “The derivative of a function f measures how quickly the function‘s output changes when you make small changes to the function’s input.” This makes sense for arbitrary input and output types, as long as we can describe small changes in them.

It is easy to describe small changes in nested structures of real numbers: they are just small changes in all the components' real numbers. For example, consider:

struct Point {
    var x, y: Float
}

struct PointPair {
    var p1, p2: Point
}

A small change in Point might be “add 0.01 to x and add 0.02 to y”. A small change in PointPair might be “add 0.01 to p1.x and add 0.01 to p2.x”.

We can define new types that capture the values of these small changes. We call these types “tangent vectors”, a term from math. For example:

extension Point {
    struct TangentVector {
        // `dx` and `dy` are small changes in `x` and `y`, respectively.
        var dx, dy: Float
    }
}

extension PointPair {
    struct TangentVector {
        // `dp1` and `dp2` are small changes in `p1` and `p2`, respectively.
        var dp1, dp2: Point.TangentVector
    }
}

In terms of these tangent vectors, the small changes that we described in words above would be:

Point.TangentVector(dx: 0.01, dy: 0.02)

PointPair.TangentVector(
    p1: Point.TangentVector(dx: 0.01, dy: 0),
    p2: Point.TangentVector(dx: 0.01, dy: 0))

In terms of tangent vectors, the derivative of a function f: (A) -> B is a function df: (A, A.TangentVector) -> B.TangentVector. In other words, df takes a starting value of type A and a small change A.TangentVector and tells you what the resulting small change in B is.

The gradient descent iterative optimization algorithm can run on any function f: (A) -> Float as long as A is a type for which we can define a tangent vector. It iteratively walks around different values of A, searching for a value that minimizes the output of f.

Proposed solution

To push Swift's capabilities to the next level in numerics and machine learning, we introduce differentiable programming as a new language feature, which includes standard library APIs and small additive changes to the type system.

The Differentiable protocol

Differentiable is a standard library protocol that generalizes all data structures that can be a parameter or result of a differentiable function. The compiler derives protocol requirement implementations when a @memberwise conformance is declared.

extension Float: Differentiable {
    typealias TangentVector = Self
}
struct Perceptron: @memberwise Differentiable {
    var weight: SIMD64<Float>
    var bias: Float
}

The @differentiable declaration attribute

The @differentiable declaration attribute is an attribute that marks function-like declarations (function declarations, initializers, properties, and subscripts) as being differentiable.

@differentiable
func cubed(_ x: Float) -> Float {
    x * x * x
}
extension Perceptron {
    @differentiable
    func callAsFunction(_ input: SIMD64<Float>) -> Float {
        (weight * input).sum() + bias
    }
}

@differentiable function types

Differentiable functions are first-class values, identified by a @differentiable attribute in the function type. A @differentiable function type is a subtype of its corresponding normal function type (i.e. without a @differentiable attribute) with an extended ABI, which stores metadata that allows their values to be differentiated anywhere the function is passed. A @differentiable(linear) function type is a subtype of its corresponding @differentiable function type. A normal function can be implicitly converted to a @differentiable or @differentiable(linear) function with appropriate compile-time checks.

func addOne(_ x: Float) -> Float { x + 1 }
let _: @differentiable (Float) -> Float = addOne
let _: @differentiable(linear) (Float) -> Float = addOne

@derivative and @transpose attributes

@derivative and @transpose attributes are used for declaring custom derivative functions for some other function declaration.

import Glibc

@derivative(of: expf)
func _(_ x: Float) -> (value: Float,
                       differential: @differentiable(linear) (Float) -> Float) {
    let y = expf(x)
    return (value: y, differential: { v in v * y })
}

Differential operators

Differential operators are APIs defined in the standard library that take @differentiable functions and return derivative functions or compute derivative values.

// In the standard library:
//
//     func derivative<T: FloatingPoint, R>(
//         of body: @escaping @differentiable (T) -> R
//     ) -> (T) -> R where T.TangentVector: FloatingPoint

@differentiable
func f(_ x: Float) -> Float {
    x * x
}
let dfdx = derivative(of: f)
dfdx(3) // 6

Detailed design

Differentiable data structures

Speaking in terms of elementary calculus, only functions are “differentiable”: only functions have derivatives and can be differentiated. In programming languages, types are isomorphic to mathematical spaces, and functions are isomorphic to mathematical functions over those spaces. Differentiability depends heavily on the continuity and smoothness of points in a space (or values of a type). For example, the Int type represents the space of integers, which are discrete values, so functions over integers cannot be differentiated. In general, when a type is said to be differentiable, it means that one can do calculus with its values. As such, real numbers, real vector spaces, and complex vector spaces are differentiable, but characters, strings, and integers are not.

For full flexibility and extensibility, a protocol is introduced in the Swift standard library to generalize all data structures that can be a parameter or a result of a differentiable function.

The Differentiable protocol

The Differentiable protocol defines operations and structures required for a type to be differentiated.

public protocol Differentiable {
    /// A type that can be used to represent derivatives with respect to a
    /// value whose type is `Self`. Mathematically, this is equivalent to the
    /// tangent bundle of the differentiable manifold represented by the
    /// differentiable type.
    associatedtype TangentVector: Differentiable & AdditiveArithmetic
        where TangentVector == TangentVector.TangentVector

    /// Moves `self` along the given direction. In Riemannian geometry, this is
    /// equivalent to exponential map, which moves `self` on the geodesic
    /// surface along the given tangent vector.
    mutating func move(along direction: TangentVector)
    
    /// A closure that produces a zero tangent vector and does not capture `self`.
    ///
    /// In some cases, the zero tangent vector of `self` is equal to
    /// `TangentVector.zero`. In other cases, the zero tangent vector depends on
    /// information in `self`, such as shape for an n-dimensional array type.
    /// For differentiable programming, it is more memory-efficient to define a
    /// custom `zeroTangentVectorInitializer` property which returns a closure
    /// that captures and uses only the necessary information to create a zero
    /// tangent vector. For example:
    ///
    /// ```swift
    /// struct Vector {
    ///     var scalars: [Float]
    ///     var count: Int { scalars.count }
    ///     init(repeating repeatedElement: Float, count: Int) { ... }
    /// }
    /// 
    /// extension Vector: Differentiable {
    ///     typealias TangentVector = Vector
    ///
    ///     @noDerivative
    ///     var zeroTangentVectorInitializer: () -> TangentVector {
    ///         let count = self.count
    ///         return { TangentVector(repeating: 0, count: count) }
    ///     }
    /// }
    /// ```
    ///
    @noDerivative
    var zeroTangentVectorInitializer: () -> TangentVector { get }
}

extension Differentiable {
    /// A tangent vector such that `move(along: zeroTangentVector)` will not modify
    /// `self`.
    @noDerivative
    var zeroTangentVector: TangentVector { zeroTangentVectorInitializer() }
}

Specifically, Differentiable generalizes types to satisfy the following requirements from real-world use cases: Functions over these types can be differentiable. Besides types, a function‘s differentiability also depends on the function’s body. Values of these types can be updated based on derivative values. For full flexibility, differentiable types should not be required to be a vector space. For example, a differentiable neural network layer can store a Bool flag in addition to differentiable parameters.

Intuitively, a Differentiable-conforming type allows one to do calculus with its values. In elementary calculus, a derivative of a real-valued function at a point is the slope of the tangent line at this point. The tangent line is the best linear approximation of the differentiated function near that input value. The same definition applies to vector-valued functions when they are split into their coordinate functions. The derivative of a vector-valued function at a certain point is called a tangent vector. Beyond real numbers and vector spaces, there is a widely accepted mathematical framework, differential geometry, which generalizes calculus beyond Euclidean space. By bringing ideas from this mathematical framework into the Swift standard library and the Swift compiler, differentiable programming becomes more flexible and expressive than ever.

Mathematically speaking, types that conform to Differentiable are considered smooth Riemannian manifolds. When differentiating a function over these manifolds, a derivative value is a vector in the tangent bundle of this manifold and has type TangentVector. The associated type TangentVector is required to conform to AdditiveArithmetic because additive group properties zero and +(_:_:) are necessary for initializing and accumulating derivative values.

The move(along:) method is equivalent to the mathematical notion of exponential map, which takes a tangent vector (e.g. a derivative), and moves the value along the direction specified by the tangent vector on the geodesic surface of the manifold. In vector spaces where the tangent vector is of the same vector space as the original differentiable space, move(along:) is equivalent to vector addition. Mathematical optimization algorithms such as gradient descent will make use of this method.

public extension Differentiable where Self == TangentVector {
    mutating func move(along direction: TangentVector) {
        self += direction
    }
}

The zeroTangentVector property returns a tangent vector such that calling move(along:) on the vector will not modify self. A zero tangent vector is often used in the initialization of mathematical optimization, where tangent vectors are initially zero and modified iteratively. This property may be different from TangentVector.zero because some tangent vectors depend on instance properties of self, e.g. the count property in Array.

Differentiable conformances

Conforming a type to Differentiable tells Swift that changes in values of this type can be differentiated, and makes functions over this type be compatible with all differentiation APIs in the standard library. Floating point numeric types and vector types, including Float, Double, Float80, and SIMD vector types, are extended to conform to Differentiable, and their TangentVectors equal themselves.

Besides numeric types, collections of numeric types are also powerful data structures in differentiable programming. For example, the Array type in the standard library conforms to Differentiable conditionally when the Element type conforms to Differentiable. This makes it possible to differentiate functions over arrays, and makes it easy to express dynamic differentiable algorithms. Similarly, other common container types in the standard library such as Optional, Dictionary, and Result can also be made differentiable via a conditional protocol conformance.

// struct Array<Element>
extension Array: Differentiable where Element: Differentiable {
    // Note: `Array.TangentVector` cannot be `Array` because `Array.+` is used for
    // concatenation and therefore cannot satisfy the `AdditiveArithmetic`
    // conformance constraint.
    public struct TangentVector: Differentiable, AdditiveArithmetic {
        public typealias TangentVector = Self
        @differentiable
        public var elements: [Element.TangentVector]
        @differentiable
        public init(_ elements: [Element.TangentVector]) { self.elements = elements }
        ...
    }

    public mutating func move(along direction: TangentVector) {
        for i in indices {
            self[i].move(along: Element.TangentVector(direction.elements[i]))
        }
    }

    @noDerivative
    public var zeroTangentVectorInitializer: () -> TangentVector {
        { [zeroInits = map(\.zeroTangentVectorInitializer)] in
            TangentVector(zeroInits.map { $0() })
        }
    }
}

// struct Dictionary<Key: Hashable, Value>
extension Dictionary: Differentiable where Value: Differentiable {
    public struct TangentVector: Differentiable, AdditiveArithmetic {
        public typealias TangentVector = Self
        @differentiable
        public var elements: [Key: Value.TangentVector]
        @differentiable
        public init(_ elements: [Key: Value.TangentVector]) {
            self.elements = elements
        }
        ...
    }

    public mutating func move(along direction: TangentVector) {
        for i in indices {
            self[i].move(along: Value.TangentVector(direction.elements[i]))
        }
    }

    @noDerivative
    public var zeroTangentVectorInitializer: () -> TangentVector {
        { [keys = self.keys] in
            let pairs = zip(keys, sequence(first: .zero, next: {$0}))
            return TangentVector(Dictionary(uniqueKeysWithValues: pairs))
        }
    }
}

// enum Optional<Wrapped>
extension Optional: Differentiable where Wrapped: Differentiable {
    public struct TangentVector: Differentiable, AdditiveArithmetic {
        public typealias TangentVector = Self
        @differentiable
        public var value: Wrapped.TangentVector?
        @differentiable
        public init(_ value: Wrapped.TangentVector?) { self.value = value }
        ...
    }

    public mutating func move(along direction: TangentVector) {
        if let value = direction.value {
            self?.move(along: value)
        }
    }

    @noDerivative
    public var zeroTangentVectorInitializer: () -> TangentVector {
        switch self {
        case nil:
            return { TangentVector(nil) }
        case let x?:
            return { [zeroTanInit = x.zeroTangentVectorInitializer] in
                TangentVector(zeroTanInit())
            }
        }
    }
}

Compiler-synthesized conformances

In numerics and machine learning, high-level data structures such as neural network layers and models are formed from smaller components stored as properties in structure types and class types. In order to use these types for differentiation, one must extend these types to conform to the Differentiable protocol. Luckily, this need not be done manually in most cases—the compiler automatically synthesizes conformances when a memberwise Differentiable conformance is declared.

Synthesis conditions

The compiler automatically synthesizes implementations of Differentiable protocol requirements for struct and class types. Here are the conditions for synthesis: The type must declare a conformance to Differentiable with a @memberwise attribute before the protocol name, either on the type declaration or on an extension in the same file. All stored properties of the conforming type must either be a var that conforms to Differentiable or be marked with the @noDerivative attribute. If a non-Differentiable or a let stored property is not marked with @noDerivative, then it is treated as if it has @noDerivative and the compiler emits a warning (with a fix-it in IDEs) asking the user to make the attribute explicit.

Default synthesis

By default, the compiler synthesizes a nested TangentVector structure type that contains the TangentVectors of all stored properties that are not marked with @noDerivative. In other words, @noDerivative makes a stored property not be included in a type's tangent vectors.

The synthesized TangentVector has the same effective access level as the original type declaration. Properties in the synthesized TangentVector have the same effective access level as their corresponding original properties.

A move(along:) method is synthesized with a body that calls move(along:) for each pair of the original property and its corresponding property in TangentVector.

Similarly, when memberwise derivation is possible, zeroTangentVectorInitializer is synthesized to return a closure that captures and calls each stored property's zeroTangentVectorInitializer closure. When memberwise derivation is not possible (e.g. for custom user-defined TangentVector types), zeroTangentVectorInitializer is synthesized as a { TangentVector.zero } closure.

Here's an example:

struct Foo<T: Differentiable, U: Differentiable>: @memberwise Differentiable {
    // `x` and `y` are the "differentiation properties".
    var x: T
    var y: U
    @noDerivative var customFlag: Bool
    @noDerivative let helperVariable: T

    // The compiler synthesizes:
    //
    //     struct TangentVector: Differentiable, AdditiveArithmetic {
    //         var x: T.TangentVector
    //         var y: U.TangentVector
    //     }
    //
    //     mutating func move(along direction: TangentVector) {
    //         x.move(along: direction.x)
    //         y.move(along: direction.y)
    //     }
    //
    //     @noDerivative
    //     var zeroTangentVectorInitializer: () -> TangentVector {
    //         { [xTanInit = x.zeroTangentVectorInitializer,
    //            yTanInit = y.zeroTangentVectorInitializer] in
    //             TangentVector(x: xTanInit(), y: yTanInit())
    //         }
    //     }
}

Shortcut synthesis

In certain cases, it is not ideal to keep Self and TangentVector as separate types. A most obvious example of this is when all of the following conditions are met: Self is declared to conform to AdditiveArithmetic. All stored properties are declared to conform to AdditiveArithmetic. There are no @noDerivative stored properties.

In these cases, the compiler will make TangentVector be a type alias for Self. Method move(along:) will not be synthesized because a default implementation already exists.

struct Point<T: Real>: @memberwise Differentiable, @memberwise AdditiveArithmetic {
    // `x` and `y` are the "differentiation properties".
    var x, y: T

    // The compiler synthesizes:
    //
    //     typealias TangentVector = Self
    //
    //     @noDerivative
    //     var zeroTangentVectorInitializer: () -> TangentVector {
    //         { [xTanInit = x.zeroTangentVectorInitializer,
    //            yTanInit = y.zeroTangentVectorInitializer] in
    //             TangentVector(x: xTanInit(), y: yTanInit())
    //         }
    //     }
}

Differentiable function declarations

At the heart of a differentiable programming language is the ability to express differentiable functions, from abstract manifold operations all the way down to floating point addition. Because differentiable programming is a flexible and extensible language feature in Swift, the compiler is agnostic of actual mathematical operations—it does not have special knowledge of standard library operators such as Float.+(::), nor does it distinguish between primitive operations and normal functions. A function can be differentiated with respect to certain Differentiable-conforming parameters if it satisfies one of the following requirements:

• Base case 1: It is linear with respect to those parameters.

• Base case 2: A derivative function for it with respect to those parameters exists in code.

• Recursive case: All function calls, initializer calls, subscript accesses, property accesses, variable assignments along the path from those parameters to the result can be differentiated.

The @differentiable declaration attribute

The @differentiable declaration attribute can be used to mark function declarations, initializers, properties, and subscripts as being differentiable. When one of these entities is marked with @differentiable, the compiler attempts to differentiate it with respect to all parameters (including any implicit self parameter) that conform to the Differentiable protocol. One can specify explicit parameters via a wrt: clause, e.g. @differentiable(wrt: x) and @differentiable(wrt: (self, x)). In generic algorithms, one can also provide a where-clause to specify generic constraints for parameters or the result to make the function differentiable only when the generic constraints are satisfied, e.g. @differentiable(wrt: x where Scalar: FloatingPoint).

@differentiable // differentiable with respect to 'x'
func silly(_ x: Float, _ n: Int) -> Float {
    print("Running 'silly' on \(x) and \(n)!")
    return sin(cos(x))
}

Computed property getters behave like methods in that they accept exactly one argument, self. If a computed property is marked with @differentiable, the compiler attempts to differentiate its getter with respect to self. @differentiable can also be applied to an explicit getter declaration.

extension Float {
    @differentiable
    var reciprocal: Float {
        1 / self
    }
}

Among these language constructs, stored properties are the least method-like in that they are stored values and cannot have a user-defined getter. However, access to stored properties can be considered as a projection of self. Therefore, stored properties can be marked @differentiable and be differentiated as a function as well. However, an explicit @differentiable is only necessary for public properties in public structs or classes to support library evolution, and are implicitly synthesized by the compiler when the parent type's Differentiable conformance is synthesized by the compiler (not user-defined).

public struct Vector: @memberwise Differentiable {
    @differentiable // Okay, though the compiler has synthesized it.
    public var x, y: Float
}

Conformance and subclassing

Protocol requirements and class members can be made differentiable with a @differentiable attribute. Semantically, this means that this member is guaranteed to be differentiable, and that any conformance implementation or inheritance must maintain the differentiability.

Protocol dispatch

The @differentiable attribute can be used on protocol requirements. A @differentiable protocol requirement requires that all conforming types implement this requirement with a differentiable body with respect to the specified parameters. Conforming implementations are not required to be marked with @differentiable attribute unless they are public.

public protocol Layer: Differentiable {
    associatedtype Input: Differentiable
    associatedtype Output: Differentiable
    @differentiable // w.r.t. `input` and `self`
    func callAsFunction(_: Input) -> Output
}
struct Perceptron: @memberwise Differentiable, Layer {
    var weight: SIMD4<Float>
    var bias: Float

    func callAsFunction(_ input: SIMD4<Float>) -> Float {
        (weight * input).sum() + b
    }
}

In a protocol hierarchy, one can override a differentiable protocol requirement with a @differentiable attribute that declares differentiability with respect to more parameters.

public protocol Module: Differentiable {
    associatedtype Input
    associatedtype Output: Differentiable
    @differentiable(wrt: self)
    func callAsFunction(_: Input) -> Output
}

public protocol Layer: Module where Input: Differentiable {
    @differentiable(wrt: (self, input))
    func callAsFunction(_: Input) -> Output
}

In the example above, types that are declared to conform to Layer (the protocol with a refined callAsFunction(_:) method) can omit the @differentiable(wrt: self) attribute on the method implementation and use @differentiable(wrt: (self, input)) (or just @differentiable) only.

Differentiable protocol requirements are not allowed to use a where-clause in the @differentiable attribute. This is to simplify the programming model where protocol requirement overrides are more powerful.

Class dispatch

A differentiable non-final class method, property or subscript can be overridden by a subclass implementation. The overriding implementation must be @differentiable if the original overridden declaration is marked with @differentiable. When a method/subscript call or a property access that is dynamically dispatched is being differentiated, the derivative of the subclass implementation will be used.

class Superclass {
    @differentiable
    func foo(_ x: SIMD8<Float>) -> Float {
        x.sum()
    }
}

class Subclass: Superclass {
    @differentiable
    override func foo(_ x: SIMD8<Float>) -> Float {
        (x * x).sum()
    }
}

Make a function differentiable using @derivative or @transpose

Any function that has Differentiable-conforming parameters and result can be made differentiable by extending the function to have either an associated derivative function or a linear transpose. In other words, derivative functions and transpose functions provide differentiability for other functions.

The @derivative attribute is used for marking a function as producing a custom derivative for another function, hence making the other function differentiable. The @transpose attribute is used for marking a function as transposing another function, hence making the other function linear.

A protocol requirement or class method/property/subscript can be made differentiable via a derivative function or transpose function defined in an extension. When a protocol requirement is not marked with @differentiable but has been made differentiable by a @derivative or @transpose declaration in a protocol extension, a dispatched call to such a member can be differentiated, and the derivative or transpose is always the one provided in the protocol extension.

Linear maps

Linear maps are a fundamental concept in differentiation. Differentiating a function between two differentiable manifolds at a certain point produces a linear map between the tangent space at that point in the input manifold and the tangent space at the corresponding point at the output manifold. This linear map is called a differential (or pushforward), which applies the chain rule to compute directional derivatives. Gradients, on the other hand, are computed by a linear map called pullback, which is the transpose of a differential, where transposition can be thought of as transposing the matrix representing the linear map. It is important that functions that are used for chaining derivatives are implemented as linear maps provided with a transpose (e.g. scalar multiplication, matrix transposition, and matrix multiplication), because gradients can only be computed when the differential can be transposed.

To make an original function be linear, define a transpose function with a @transpose attribute that specifies the original function.

Typing rules

A function declaration does not have a fixed transpose type. This is because there can be multiple transpose functions that transpose the original function differently, e.g. with respect to different parameters, transposing under different generic constraints, etc.

Given an original function declaration, a transpose function's type is determined from the following configurations:

• Parameters to transpose with respect to.
• Additional generic constraints that make the original function linear.

The type of the transpose function under such configurations is a function that takes one argument whose type is the original function's result type and returns results that correspond to each original function parameter that is transposed with respect to. This definition, however, is a rough definition because there are differences among top-level functions, instance methods, and static methods.

Linearity parameters

Linearity parameters are parameters with respect to which a function is linear. The @transpose attribute accepts a wrt: argument which specifies a set of linearity parameters of the original function. If wrt: is not specified, linearity parameters default to all parameters. A wrt: argument in @derivative attributes can be a parameter index, a self, or a tuple of parameter indices and self. When there are more than one linearity parameters specified, parameter indices must be ascending, and self must be the first parameter when exists. All linearity parameters must have a type that conforms to both Differentiable and AdditiveArithmetic and satisfies Self == Self.TangentVector.

When linearity parameters do not include all of the original function's parameters, those parameters must be taken in the front of the parameter list of the transpose function.

The argument labels of original non-linearity parameters must be preserved in the transpose function. Other argument labels can be named freely. When there are multiple linearity parameters, it is useful to label the elements in the result tuple to distinguish between transposes with respect to different parameters.

Top-level functions

Note: Since both transpose functions and derivative functions are difficult to name and need not be referenced directly, we make these functions unnamed (with base name being an underscore). This is not yet valid in the official Swift language, but the developers of the differentiable programming feature will prototype and pitch this change through Swift Evolution.

func foo<T: Differentiable & AdditiveArithmetic>(_ x: T, _ y: T, _ z: T) -> T
    where T == T.TangentVector { ... }

// Transpose with respect to all parameters, making `foo(_:_:_:)` linear with
// with respect to all parameters.
@transpose(of: foo)
func _<T: Differentiable & AdditiveArithmetic>(_ v: T) -> (x: T, y: T, z: T)
    where T == T.TangentVector { ... }

// Transpose with respect to original parameter `x`, making `foo(_:_:_:)` 
// linear with respect to `x`.
@transpose(of: foo, wrt: 0)
func _<T: Differentiable & AdditiveArithmetic>(y: T, z: T, v: T) -> T
    where T == T.TangentVector { ... }

// Transpose with respect to original parameters `x` and `z`, making
// `foo(_:_:_:)` linear with respect to `x` and `z`.
@transpose(of: foo, wrt: (0, 2))
func _<T: Differentiable & AdditiveArithmetic>(y: T, v: T) -> (x: T, z: T)
    where T == T.TangentVector { ... }

Static methods

A transpose of a static method is exactly like top-level functions except that it must also be defined as a static method in the same type. The implicit self parameter cannot be a linearity parameter, because metatypes cannot conform to Differentiable & AdditiveArithmetic.

extension MyType {
    static func foo<T: Differentiable & AdditiveArithmetic>(_ x: T, _ y: T, _ z: T) -> T
        where T == T.TangentVector { ... }
}

extension MyType {
    // Transpose with respect to all parameters, making `foo(_:_:_:)` linear with
    // with respect to all parameters.
    @transpose(of: foo)
    static func _<T: Differentiable & AdditiveArithmetic>(_ v: T) -> (x: T, y: T, z: T)
        where T == T.TangentVector { ... }
    
    // Transpose with respect to original parameter `x`, making `foo(_:_:_:)` 
    // linear with respect to `x`.
    @transpose(of: foo, wrt: 0)
    static func _<T: Differentiable & AdditiveArithmetic>(y: T, z: T, v: T) -> T
        where T == T.TangentVector { ... }
    
    // Transpose with respect to original parameters `x` and `z`, making
    // `foo(_:_:_:)` linear with respect to `x` and `z`.
    @transpose(of: foo, wrt: (0, 2))
    static func _<T: Differentiable & AdditiveArithmetic>(y: T, v: T) -> (x: T, z: T)
        where T == T.TangentVector { ... }
}

The numeric addition operator AdditiveArithmetic.+(_:_:) is linear, and the multiplication operator Numeric.*(_:_:) is bilinear (i.e. linear with respect to each parameter). Here's how they are made differentiable in the standard library.

extension FloatingPoint 
    where Self: Differentiable & AdditiveArithmetic, Self == TangentVector
{
    @transpose(of: +)
    static func _(_ v: Self) -> (Self, Self) { (v, v) }

    @transpose(of: *, wrt: 0)
    @transpose(of: *, wrt: 1)
    static func _(lhs: Self, rhs: Self) -> Self { lhs * rhs }
}

As shown, transpose functions may be defined in a type extension or a protocol extension that has more generic constraints than the original +(_:_:) and *(_:_:) declarations. This makes the original functions linear only when these extra generic constraints are satisfied. Moreover, transpose functions for *(_:_:) are defined per-parameter due to the nature of bilinearity (x + y is a flat plane while x * y is not), but fortunately its transpose functions with respect to each parameter are just *(_:_:) itself.

In vector calculus, transpose functions become less trivial. For example, here is a hypothetical Tensor type, which has two transpose functions defined for Tensor.transposed(), the tensor transposition method, and matmul(_:_:), the matrix multiplication function.

extension Tensor where Scalar: FloatingPoint & Differentiable {
    @transpose(of: transposed, wrt: self)
    func _() -> Tensor {
        self.transposed()
    }
}

@transpose(of: matmul(_:_:), wrt: 0)
func _<T: FloatingPoint & Differentiable>(y: Tensor<T>, v: Tensor<T>) -> Tensor<T> {
    matmul(v, y.transposed())
}

@transpose(of: matmul(_:_:), wrt: 1)
func _<T: FloatingPoint & Differentiable>(x: Tensor<T>, v: Tensor<T>) -> Tensor<T> {
    matmul(x.transposed(), v)
}

Instance methods

A transpose of a static method is exactly like top-level functions except:

• When linearity parameters does not include self, it must be defined as an instance method in the same type.
• When linearity parameters include self, it must be defined as a static method in the same type.

extension MyType {
    func foo<T: Differentiable & AdditiveArithmetic>(_ x: T, _ y: T, _ z: T) -> T
        where T == T.TangentVector { ... }
}

extension MyType {
    // Transpose with respect to all parameters, making `foo(_:_:_:)` linear with
    // with respect to all parameters.
    @transpose(of: foo)
    func _<T: Differentiable & AdditiveArithmetic>(_ v: T) -> (x: T, y: T, z: T)
        where T == T.TangentVector { ... }
    
    // Transpose with respect to original parameter `x`, making `foo(_:_:_:)` 
    // linear with respect to `x`.
    @transpose(of: foo, wrt: 0)
    func _<T: Differentiable & AdditiveArithmetic>(y: T, z: T, v: T) -> T
        where T == T.TangentVector { ... }
    
    // Transpose with respect to original parameters `x` and `z`, making
    // `foo(_:_:_:)` linear with respect to `x` and `z`.
    @transpose(of: foo, wrt: (0, 2))
    func _<T: Differentiable & AdditiveArithmetic>(y: T, v: T) -> (x: T, z: T)
        where T == T.TangentVector { ... }
    
    // Transpose with respect to original parameters `self`, making `foo(_:_:_:)`
    // linear with respect to `self`.
    @transpose(of: foo, wrt: self)
    static func _<T: Differentiable & AdditiveArithmetic>(x: T, y: T, z: T, v: T) -> MyType
        where T == T.TangentVector { ... }
    
    // Transpose with respect to original parameters `self`, `x` and `z`, making
    // `foo(_:_:_:)` linear with respect to `self`, `x` and `z`.
    @transpose(of: foo, wrt: (self, 0, 2))
    static func _<T: Differentiable & AdditiveArithmetic>(y: T, v: T) -> (self: MyType, x: T, z: T)
        where T == T.TangentVector { ... }
}

Linearity generic requirements

A transpose function can have additional generic constraints, called linearity generic requirements. Linearity generic requirements usually serve the purpose of making generic parameter types conform to Differentiable & AdditiveArithmetic.

Linearity generic requirements are functionally equivalent to the where clause in @differentiable attributes.

func foo<T, U, V>(_ x: T, _ y: U, _ z: V) -> W { ... }

// Transpose with respect to `x` and `z`, requiring that `T` and `V` to conform
// to `Differentiable & AdditiveArithmetic` and equal their corresponding
`TangentVector` types.
@transpose(of: foo, wrt: (x, z))
func _<
    T: Differentiable & AdditiveArithmetic,
    U,
    V: Differentiable & AdditiveArithmetic
>(_ y: U, _ v: W) -> (x: T, z: V)
    where T.TangentVector == T, V.TangentVector == V { ... }

Examples

Many floating-point operations are linear. Addition and subtraction are linear. Multiplication is bilinear (linear with respect to each argument).

extension FloatingPoint where Self: Differentiable, Self == TangentVector {
    @inlinable
    @transpose(of: +)
    func _(_ v: Self) -> (Self, Self) {
        (v, v)
    }

    @inlinable
    @transpose(of: -)
    func _(_ v: Self) -> (Self, Self) {
        (v, -v)
    }

    @inlinable
    @transpose(of: *, wrt: 0)
    @transpose(of: *, wrt: 1)
    func _(_ x: Self, _ v: Self) -> Self {
        return x * v
    }
}

Complex differentiation is representable in our system. Complex numbers behave differently from real numbers and vectors (forum discussion: they have an additional conjugate operation which flips the sign of the imaginary component.

Since complex numbers are not yet defined in the standard library, we extended the complex number type defined in the NumericAnnex library to be differentiable. The full implementation is here. The implementation adopts the Autograd convention for derivatives of functions with complex arguments or results, so that we can define derivatives for non-holomorphic primitives.

struct Complex<Base: FloatingPoint>: Numeric {
    var real: Base
    var imaginary: Base

    @differentiable(linear where Base: Differentiable, Base == Base.TangentVector)
    init(real: Base = 0, imaginary: Base = 0) {
        self.real = real
        self.imaginary = imaginary
    }

    ...
}

extension Complex: @memberwise Differentiable where Base: Differentiable, Base == Base.TangentVector {}

extension Complex {
    @differentiable(where Base: Differentiable, Base == Base.TangentVector)
    func complexConjugate() -> Complex {
        Complex(real: real, imaginary: -imaginary)
    }
}

SIMD vectors are also differentiable: mathematically, they represent a vector space. Most SIMD operations are defined as SIMD protocol requirements, so derivatives of these operations can be defined generally in a protocol extension on SIMD.

extension SIMD where Self: Differentiable, TangentVector: SIMD, Scalar: BinaryFloatingPoint, Self == Self.TangentVector {
    @transpose(of: *, wrt: 0)
    @transpose(of: *, wrt: 1)
    static func _(v: Self, x: Self) -> Self {
        v * x
    }
}

Additionally, concrete types conforming to SIMD are extended to conditionally conform to Differentiable and AdditiveArithmetic. For SIMD conforming types, the TangentVector associated type is equal to Self.

extension SIMD${n}: AdditiveArithmetic where Scalar: BinaryFloatingPoint {}

extension SIMD${n}: Differentiable
where Scalar: Differentiable & BinaryFloatingPoint,
      Scalar.TangentVector : BinaryFloatingPoint {
    public typealias TangentVector = SIMD${n}
}

// `subscript` is defined on `SIMD`-conforming types, so the transpose is as well.
extension SIMDScalar where Self: Differentiable & BinaryFloatingPoint {
    @transpose(of: subscript)
    func _(index: Int) -> SIMD${n}<Self> {
        var result = SIMD${n}<Self>.zero
        result[index] = self
        return result
    }
}

The full implementation is in SIMDVector.swift and SIMDVectorTypes.swift.gyb on the tensorflow branch.

Derivative functions

A derivative function has the same parameters as the original function, but returns a linear differential function in addition to the original value. Computing both the original value and the differential is the most efficient way for the differential closure to capture anything it needs from the original computation, and is important for flexibility and performance.

In the following example, the 32-bit floating point exponential function expf(_:) is imported from the C standard library. The derivative function marked with @derivative makes expf(_:) a differentiable function.

import Glibc

@derivative(of: expf)
func _(_ x: Float) -> (value: Float,
                       differential: @differentiable(linear) (Float) -> Float) {
    let y = expf(x)
    return (value: y, differential: { v in v * y })
}

Typing rules

A function declaration does not have a fixed derivative type. This is because there can be multiple derivative functions that differentiate the original function differently, e.g. differentiating with respect to different parameters, differentiating with different generic constraints, etc.

Given an original function declaration, a derivative function's type is determined from the following configurations:

• Parameters to differentiate with respect to, aka. differentiability parameters.
• Additional generic constraints that make the original function differentiable.

The type of the derivative function under such configurations is a function that takes the original function‘s parameters and returns a tuple of an original result (labeled value) and a differential (labeled differential). The differential is a linear map (@differentiable(linear)) function that takes the TangentVector nested types of all of the types of the original function’s parameters to differentiate with respect to, and returns the TangentVector nested type of the orgiinal function's result type.

Differentiability parameters

The @derivative attribute accepts a wrt: argument which specifies the differentiability parameters. If wrt: is not specified, the derivative function should be differentiating the original function with respect to all of its parameters, hence producing a differential that takes all of the original function‘s parameter types’ TangentVector types. A wrt: argument in @derivative attributes can be a parameter name, a parameter index, or a tuple of multiple parameter names or indices. All differentiability parameters must have a type that conforms to Differentiable.

A derivative function's argument labels must match those of the original function. Its parameter names do not have to match those of the original function. However, a wrt: argument in a @derivative attribute, when referring to parameters by names, must use parameter names in the derivative function.

func foo<T: Differentiable>(_ x: T, _ y: T, _ z: T) -> T { ... }

// Derivative with respect to all parameters.
@derivative(of: foo)
func _<T: Differentiable>(_ x: T, _ y: T, _ z: T) -> (
    value: T, 
    differential: @differentiable(linear) (T.TangentVector, T.TangentVector, T.TangentVector) -> T.TangentVector
) {
    ...
}

// Derivative with respect to `x`.
@derivative(of: foo, wrt: x)
func _<T: Differentiable>(_ x: T, _ y: T, _ z: T) -> (
    value: T, 
    differential: @differentiable(linear) (T.TangentVector) -> T.TangentVector
) {
    ...
}

// Derivative with respect to `x` and `z`.
@derivative(of: foo, wrt: (x, z))
func _<T: Differentiable>(_ x: T, _ y: T, _ z: T) -> (
    value: T, 
    differential: @differentiable(linear) (T.TangentVector, T.TangentVector) -> T.TangentVector
) {
    ...
}

One concrete example is sinf(_:) from the C standard library. It can be made differentiable by defining a derivative retroactively.

#if canImport(Darwin)
import func Darwin.sinf
#else
import func Glibc.sinf
#endif

// Imported:
//     public func sinf(Float) -> Float

@derivative(of: sinf)
public func _(_ x: Float) -> (
    value: Float, 
    differential: @differentiable(linear) (Float) -> Float
) {
    (value: sinf(x), differential: { v in cosf(x) * v })
}

Differentiability generic requirements

A derivative function can have additional generic constraints, called differentiability generic requirements. Differentiability generic requirements usually serve the purpose of making generic parameter types conform to Differentiable.

Differentiability generic requirements are functionally equivalent to the where clause in @differentiable attributes.

func foo<T, U, V>(_ x: T, _ y: U, _ z: V) -> W { ... }

// Derivative with respect to `x` and `z`, requiring that `T` and `V` to conform
// to `Differentiable`.
@derivative(of: foo, wrt: (x, z))
func foo<T: Differentiable, U, V: Differentiable>(
    _ x: T, _ y: U, _ z: V
) -> (
    value: W, 
    differential: (T.TangentVector, V.TangentVector) -> W.TangentVector
) {
    ...
}

Examples

The ElementaryFunctions protocol introduced in SE-0246 defines generic elementary functions, which are non-linear. By defining derivatives using the @derivative attribute for these protocol requirements in an extension, all conforming types now have differentiable elementary functions.

public protocol ElementaryFunctions {
    static func sqrt(_ x: Self) -> Self
    static func cos(_ x: Self) -> Self
    static func asinh(_ x: Self) -> Self
    static func exp(_ x: Self) -> Self
    static func exp10(_ x: Self) -> Self
    static func log(_ x: Self) -> Self
    static func pow(_ x: Self, _ y: Self) -> Self
    ...
}

public extension ElementaryFunctions
where Self: Differentiable & FloatingPoint, Self == Self.TangentVector {
    @inlinable
    @derivative(of: sqrt)
    static func _(_ x: Self) -> (value: Self, differential: @differentiable(linear) (Self) -> Self) {
        (sqrt(x), { dx in (1 / 2) * (1 / sqrt(x)) * dx })
    }

    @inlinable
    @derivative(of: cos)
    static func _(_ x: Self) -> (value: Self, differential: @differentiable(linear) (Self) -> Self) {
        (cos(x), { dx in -sin(x) * dx })
    }

    @inlinable
    @derivative(of: asinh)
    static func _(_ x: Self) -> (value: Self, differential: @differentiable(linear) (Self) -> Self) {
        (asinh(x), { dx in 1 / (1 + x * x) * dx })
    }

    @inlinable
    @derivative(of: exp)
    static func _(_ x: Self) -> (value: Self, differential: @differentiable(linear) (Self) -> Self) {
        let ret = exp(x)
        return (ret, { dx in ret * dx })
    }

    @inlinable
    @derivative(of: exp10)
    static func _(_ x: Self) -> (value: Self, differential: @differentiable(linear) (Self) -> Self) {
        let ret = exp10(x)
        return (ret, { dx in exp(10) * ret * dx })
    }

    @inlinable
    @derivative(of: log)
    static func _(_ x: Self) -> (value: Self, differential: @differentiable(linear) (Self) -> Self) {
        (log(x), { dx in 1 / x * dx })
    }

    @inlinable
    @derivative(of: pow)
    static func _(_ x: Self, _ y: Self) -> (value: Self, differential: @differentiable(linear) (Self, Self) -> Self) {
        (pow(x, y), { (dx, dy) in
            let l = y * pow(x, y-1) * dx
            let r = pow(x, y) * log(x) * dy
            return l + r
        })
    }

    ...
}

Default derivatives and transposes

In a protocol extension, class definition, or class extension, providing a derivative or transpose for a protocol extension or a non-final class member is considered as providing a default derivative/transpose for that member. Types that conform to the protocol or inherit from the class can inherit the default derivative/transpose.

If the original member does not have a @differentiable attribute, a default derivative/transpose is implicitly added to all conforming/overriding implementations.

protocol P {
    func foo(_ x: Float) -> Float
}

extension P {
    @derivative(of: foo(x:))
    func _(_ x: Float) -> (value: Float, differential: (Float) -> Float) {
        (value: foo(x), differential: { _ in 42 })
    }
}

struct S: P {
    func foo(_ x: Float) -> Float {
        33
    }
}

let s = S()
let d = derivative(at: 0) { x in
   s.foo(x)
} // ==> 42

When a protocol requirement or class member is marked with @differentiable, it is considered as a differentiability customization point. This means that all conforming/overriding implementation must provide a corresponding @differentiable attribute, which causes the implementation to be differentiated. To inherit the default derivative/transpose without differentiating the implementation, add default to the @differentiable attribute.

protocol P {
    @differentiable
    func foo(_ x: Float) -> Float
}

extension P {
    @derivative(of: foo(x:))
    func _(_ x: Float) -> (value: Float, differential: (Float) -> Float) {
        (value: foo(x), differential: { _ in 42 })
    }
}

struct S: P {
    @differentiable(default) // Inherits default derivative for `P.foo(_:)`.
    func foo(_ x: Float) -> Float {
        33
    }
}

let s = S()
let d = derivative(at: 0) { x in
   s.foo(x)
} // ==> 42

Access control

Derivative and transpose functions provide differentiability for other functions, and the access level of the differentiability can be controlled precisely with access modifiers on derivative/transpose functions.

When a function‘s differentiability is provided by a derivative/transpose function, the access scope of differentiability is identical to the derivative/transpose function’s access scope. For example, a fileprivate derivative function in B.swift only overrides the original function's derivative in B.swift.

// File A.swift:
internal func foo(_ x: Float) -> Float {
    x * x
}
let dfdx_A = derivative(at: 3, in: foo)
// dfdx_A ==> 6

// File B.swift:
@derivative(of: foo)
fileprivate func _(_ x: Float) -> (
    value: Float, 
    differential: @differentiable(linear) (Float) -> Float
) {
    (value: foo(x), differential: { _ in 42 })
}
let dfdx_B = derivative(at: 3, in: foo)
// dfdx_B ==> 42

// File C.swift:
let dfdx_C = derivative(at: 3, in: foo)
// dfdx_C ==> 6

Differentiable function types

Differentiability is a fundamental mathematical concept that applies not only to declarations of functions, initializers, subscripts, and properties, but also to function types. In Swift, functions are first-class values of function types that can be passed around, applied, or converted. Because an important part of differentiable programming is to be able to define differential operators and custom algorithms on differentiable functions, Swift's type system has been extended to be able to express differentiable functions as first-class values.

A differentiable function type is a special function type that has a different runtime representation than a normal function type, and is a subtype of a non-differentiable function type with the same parameter types and result type.

Function subtyping and runtime representation

Subtyping of function types already exists in Swift and is primarily used for representing different foreign calling conventions for language interoperability. Function types and function pointer types in C, e.g. int(*)(int), are imported to Swift as function types with a @convention(c) attribute, e.g. @convention(c) (Int) -> Int, with all parameter types and return types converted to the corresponding Swift ones.

These function types are also subtypes of a function type with the same parameter types and result types but without the @convention(c) attribute. For example, you can implicitly convert a @convention(c) function value to a Swift function value and use it directly as an argument to higher-order functions such as map(_:).

// In a C file:
int addOne(int x) { return x + 1; }
int (*addOneFunctionPointer)(int) = addOne;
// Swift equivalent:
//   let addOneFunctionPointer: (Int) -> Int = addOne

// In a Swift file that imports the C file:
// Global variable `addOneFunctionPointer` imported as `@convention(c) (Int) -> Int`.
[1, 2, 3].map(addOneFunctionPointer) // [2, 3, 4]

One of the main differences between a Swift function value and a C function value is their runtime representation. A C function cannot capture values from the context where it is defined, so the runtime representation of a C function value is just a pointer to the function in memory. A Swift function, however, can capture values from the context, and thus contains a pointer to the heap-allocated (or sometimes stack-allocated) context storing captured values.

In differentiable programming, differentiable function types contain more information than its non-differentiable counterparts. A differentiable function contains the original function pointer so that it can be efficiently converted to or called like the original function type. It also contains a derivative function that will be called when this function is being differentiated. All of these functions share the same context. A linear map, which is differentiable by definition and whose differential at any point is itself, does not need to store derivative functions but just a linear transpose function instead.

The @differentiable function type attribute

A @differentiable attribute on a function type specifies the function‘s differentiability, just like @differentiable on function declarations. A @differentiable(linear) attribute specifies the function’s linearity with respect to differentiation. All linear maps are infinitely differentiable, therefore @differentiable(linear) is a subtype of @differentiable.

@differentiable requires the enclosing function type to have differentiable parameters and results. Each parameter and result must conform to the Differentiable protocol unless marked @noDerivative. @differentiable(linear) requires the closing function to have “differentiable vector space” parameters and results, that is, each parameter and result, unless marked @noDerivative, must conform to Differentiable & AdditiveArithmetic and satisfy Self == Self.TangentVector.

Type conversion

The subtyping relation among @differentiable(linear), @differentiable, and non-@differentiable function types allow functions of different types to be conditionally convertible to each other. Such conversions do not always succeed: Conversion from a function declaration (func) to a @differentiable function value succeeds if and only if the function can be differentiated. Conversion from a @differentiable function value to a non-@differentiable function value always succeeds. Conversion from a non-@differentiable function value to a @differentiable function value always fails, because the function's body is opaque to the compiler.

Coercing function declarations into @differentiable function values

A function declaration can be implicitly coerced into a @differentiable function value, when there is a contextual @differentiable function type. Such conversions succeed either if the function declaration has been marked with a @differentiable declaration attribute, or if the function declaration is defined in the same module and the function can be differentiated as if it were marked with @differentiable. When neither of these conditions are met, the function cannot be differentiated, and thus cannot be converted to a @differentiable function value, in which case the compiler will produce an error.

func addOne(_ x: Float) -> Float { x + 1 }
let _: @differentiable (Float) -> Float = addOne // Okay!
let _: @differentiable(linear) (Float) -> Float = addOne // Okay!

let _: @differentiable(linear) (Float) -> Float = coshf(_:)
// Error: `coshf(_:)` is from a different module and has not been marked with
// `@differentiable`.

func mySin(_ x: Float) -> Float { sin(x) * 2 }
let _: @differentiable (Float) -> Float = mySin // Okay!
let _: @differentiable(linear) (Float) -> Float = mySin
// Error: When differentiating `mySin(_:)` as a linear map, `sin` is not linear.

func addOneViaInt(_ x: Float) -> Float { Float(Int(x) + 1) }
let _: @differentiable (Float) -> Float = addOneViaInt
// Error: When differentiating `addOneViaInt(_:)`, `Int(x)` is not differentiable.

Upcasting to non-@differentiable functions

As shown in the function subtyping and runtime representation subsection, a @differentiable function value's runtime representation contains the original function along with extra information that allows the function to be differentiated (or transposed, if it is @differentiable(linear)). A @differentiable or @differentiable(linear) function value can be called like a non-@differentiable function. A @differentiable(linear) function value can be implicitly converted to a @differentiable one, which can be implicitly converted to a non-@differentiable one.

func addOne(_ x: Float) -> Float { x + 1 }
let f0: @differentiable(linear) (Float) -> Float = addOne
let f1: @differentiable (Float) -> Float = f0
let f2: (Float) -> Float = f1

A @differentiable function can also be converted to a function which is identical except that more of its parameters are marked with @noDerivative.

func addOne(_ x: Float) -> Float { x + 1 }
let f0: @differentiable (Float, Float, Float) -> Float = addOne
let f1: @differentiable (@noDerivative Float, Float, Float) -> Float = f0
let f2: @differentiable (@noDerivative Float, Float, @noDerivative Float) -> Float = f1

Implied generic constraints

In the declaration of a generic higher-order function, when a function type is marked with @differentiable as a parameter or a result and uses generic parameters from the parent function declaration, type inference will add implicit generic constraints that make the @differentiable function type's parameter types and result type conform to Differentiable.

// With all explicit generic constraints:
func foo<T: Differentiable, U: Differentiable, V: Differentiable>(
    _ f: @differentiable (T, U) -> V
) {
    ...
}

// With implied constraints:
//     where T: Differentiable, U: Differentiable, V: Differentiable
func foo<T, U, V>(_ f: @differentiable (T, U) -> V) {
    ...
}

Similarly, when such parameters or results are marked with @differentiable(linear), implicit generic constraints will add additional constraints that make the @differentiable(linear) function type's parameter types and result type conform to Differentiable & AdditiveArithmetic and satisfy Self == Self.TangentVector.

// With all explicit generic constraints:
func foo<T: Differentiable & AdditiveArithmetic,
         U: Differentiable & AdditiveArithmetic,
         V: Differentiable & AdditiveArithmetic>(
    _ f: @differentiable(linear) (T, U) -> V
) where T.TangentVector == T, U.TangentVector == U, V.TangentVector == V
{
    ...
}

// With implied constraints:
//     where T: Differentiable & AdditiveArithmetic,
//           U: Differentiable & AdditiveArithmetic,
//           V: Differentiable & AdditiveArithmetic,
//           T.TangentVector == T,
//           U.TangentVector == U,
//           V.TangentVector == V
func foo<T, U, V>(_ f: @differentiable(linear) (T, U) -> V) {
    ...
}

By extending the type system with the ability to represent differentiable functions as first-class values, users are able to define arbitrary algorithms and data structures that deal with differentiable function values, including:

Arbitrary higher-order functions that require arguments to be differentiable functions. Differential operators, e.g. derivative(of:), described in the differential operators and differentiation APIs section. Differentiable higher-order functions for collections, e.g. Array.differentiableReduce(_:_:). Data structures that store @differentiable functions as a property. Neural network layers that store activation functions, e.g. Dense. Neural network trainer objects that store loss functions, e.g. Learner in the fast.ai Swift notebooks.

Non-differentiable parameters

Like function declarations with a @differentiable attribute, differentiable function values can also be differentiable with respect to a subset of parameters. This is expressed as part of type information, in @differentiable and @differentiable(linear) function types, using a @noDerivative attribute at each parameter that is not being differentiated with respect to.

By default, all parameters are being differentiated with respect to. When a @noDerivative attribute is specified for a parameter in a @differentiable function type, values of this function type are not differentiable (or linear) with respect to the parameter.

let f0: @differentiable (Float, Float) -> Float = { $0 * $1 }
let f1: @differentiable(linear) (Float, Float) -> Float = { $0 + $1 }
let f2: @differentiable(linear) (Float, @noDerivative Float) -> Float = { $0 * $1 }
let f3: @differentiable (@noDerivative Int, Float, @noDerivative Int) -> Float = {
  $0 ? Float($1) + $2 : 0
}

Differentiability of parameters in a function type is important for type conversions and is part of the subtyping rule: Any @differentiable or @differentiable(linear) function type is a subtype of the same function type with more @noDerivative parameters than there originally are.

let f0: @differentiable (Float, Float) -> Float = { $0 * $1 }
_ = f0 as @differentiable (Float, @noDerivative Float) -> Float
_ = f0 as @differentiable (@noDerivative Float, Float) -> Float
_ = f0 as @differentiable (@noDerivative Float, @noDerivative Float) -> Float

Higher-order functions and currying

As defined above, the @differentiable function type attributes requires all non-@noDerivative arguments and results to conform to the @differentiable attribute. However, there is one exception: when the type of an argument or result is a function type, e.g. @differentiable (T) -> @differentiable (U) -> V. This is because we need to differentiate higher-order funtions.

Mathematically, the differentiability of @differentiable (T, U) -> V is similar to that of @differentiable (T) -> @differentiable (U) -> V in that differentiating either one will provide derivatives with respect to parameters T and U. Here are some examples of first-order function types and their corresponding curried function types:

First-order function type	Curried function type
@differentiable (T, U) -> V	@differentiable (T) -> @differentiable (U) -> V
@differentiable (T, @noDerivative U) -> V	@differentiable (T) -> (U) -> V
@differentiable (@noDerivative T, U) -> V	(T) -> @differentiable (U) -> V

A curried differentiable function can be formed like any curried non-differentiable function in Swift.

func curry<T, U, V>(
    _ f: @differentiable (T, U) -> V
) -> @differentiable (T) -> @differentiable (U) -> V {
    { x in { y in f(x, y) } }
}

The way this works is that the compiler internally assigns a tangent bundle to a closure that captures variables. This tangent bundle is existentially typed, because closure contexts are type-erased in Swift. The theory behind the typing rules has been published as The Differentiable Curry.

Differential operators

The core differentiation APIs are the differential operators. Differential operators are higher-order functions that take @differentiable functions as inputs and return derivative functions or evaluate derivative values.

Differential-producing differential operators

Among these differential operators, two base APIs, valueWithDifferential(at:in:) and transpose(of:), are used for implementing all other differential operators and differentiation APIs.

/// Returns `body(x)` and the differential of `body` at `x`.
func valueWithDifferential<T, R>(
    at x: T, in body: @differentiable (T) -> R
) -> (value: R,
      differential: @differentiable(linear) (T.TangentVector) -> R.TangentVector) {
    // Compiler built-in.
    Builtin.applyDerivative_arity1(body, x)
}


/// Returns the transpose of the linear map `body`.
func transpose<T, R>(
    of body: @escaping @differentiable(linear) (T) -> R
) -> @differentiable(linear) (R) -> T {
    // Compiler built-in.
    { x in Builtin.applyTranspose_arity1(body, x) }
}

The most common differential operators are the ones that compute directional derivatives. These differential operators are defined to take a differentiable function whose parameter is a real number.

func valueWithDerivative<T: FloatingPoint, R>(
    at x: T, in body: @differentiable (T) -> R
) -> (value: R, derivative: R.TangentVector) where T.TangentVector: FloatingPoint {
    let (value, df) = valueWithDifferential(at: x, in: body)
    return (value, df(T.TangentVector(1)))
}

func derivative<T: FloatingPoint, R>(
    at x: T, in body: @differentiable (T) -> R
) -> R.TangentVector where T.TangentVector: FloatingPoint {
    valueWithDerivative(at: x, in: body).derivative
}

func derivative<T: FloatingPoint, R>(
    of body: @escaping @differentiable (T) -> R
) -> (T) -> R.TangentVector where T.TangentVector: FloatingPoint {
    return { x in derivative(at: x, in: body) }
}

Pullback-producing differential operators

Unlike directional derivatives, gradients are computed by pullbacks. Based on the differential-producing differential operator valueWithDifferential(at:in:), valueWithPullback(at:in:) is defined as returning the original value and the transpose of the differential, and valueWithGradient(at:in:) is defined as evaluating the pullback at 1 when the function being differentiated returns a real number.

func valueWithPullback<T, R>(
    at x: T, in body: @differentiable (T) -> R
) -> (value: R,
      pullback: @differentiable(linear) (R.TangentVector) -> T.TangentVector) {
    let (value, df) = valueWithDifferential(at: x, in: body)
    return (value, transpose(of: df))
}

func valueWithGradient<T, R: FloatingPoint>(
    at x: T, in body: @differentiable (T) -> R
) -> (value: R, gradient: T.TangentVector) where R.TangentVector: FloatingPoint {
    let (value, pullback) = valueWithPullback(at: x, in: body)
    return (value, pullback(R.TangentVector(1)))
}

func gradient<T, R: FloatingPoint>(
    at x: T, in body: @differentiable (T) -> R
) -> T.TangentVector where R.TangentVector: FloatingPoint {
    return valueWithGradient(at: x, in: body).gradient
}

func gradient<T, R: FloatingPoint>(
    of body: @escaping @differentiable (T) -> R
) -> (T) -> T.TangentVector where R.TangentVector: FloatingPoint {
    return { x in gradient(at: x, in: body) }
}

Example usage

All of these APIs are designed to work nicely with Swift's trailing closure syntax. Here is an example of training a simple deep learning model:

for _ in 0..<1000 {
    // Differentiate the loss with respect to the model `classifier` itself,
    // producing a tangent vector `𝛁model` that represents partial derivatives
    // with respect to all differentiable properties (trainable model parameters)
    // in the model
    let 𝛁model = gradient(at: classifier) { classifier -> Tensor<Float> in
        let ŷ = classifier(x)
        let loss = softmaxCrossEntropy(logits: ŷ, labels: y)
        print("Loss: \(loss)")
        return loss
    }
    optimizer.update(&classifier, along: 𝛁model)
}

List of differential operators

Differential operators	Description
transpose(of:)	Returns transpose of linear map.
valueWithDifferential(at:in:) valueWithDifferential(at:_:in:) (arity 2)	Returns original result and differential function.
valueWithPullback(at:in:) valueWithPullback(at:_:in:)	Returns original result and pullback function.
differential(at:in:) differential(at:_:in:) (arity 2)	Returns differential function.
pullback(at:in:) pullback(at:_:in:)	Returns pullback function.
derivative(at:in:) derivative(at:_:in:) (arity 2)	Returns partial derivatives with respect to arguments (“forward-mode”).
gradient(at:in:) gradient(at:_:in:)	Returns partial derivatives with respect to arguments (“reverse-mode”).
valueWithDerivative(at:in:) valueWithDerivative(at:_:in:) (arity 2)	Returns original result and partial derivatives with respect to arguments (“forward-mode”).
valueWithGradient(at:in:) valueWithGradient(at:_:in:)	Returns original result and partial derivatives with respect to arguments (“reverse-mode”).
derivative(of:) derivative(of:) (arity 2)	Returns derivative function, taking original arguments and returning and partial derivatives with respect to arguments (“forward-mode”).
gradient(of:) gradient(of:)	Returns gradient function, taking original arguments and returning and partial derivatives with respect to arguments (“reverse-mode”).
valueWithDerivative(of:) valueWithDerivative(of:) (arity 2)	Returns function taking original arguments and returning original result and partial derivatives with respect to arguments (“forward-mode”).
valueWithGradient(of:) valueWithGradient(of:)	Returns function taking original arguments and returning original result and partial derivatives with respect to arguments (“reverse-mode”).

Static analysis

Differentiable programming in Swift aims to provide the best static compiler diagnostics to help users catch mistakes. Beyond error diagnostics, the compiler and the standard library are equipped with static analyses and marker APIs that help the user write differentiable code with explicit annotations about non-obvious non-differentiable cases.

Cross-module opacity

Swift libraries are distributed as modules, which provide an API interface and an opaque binary format for client code to use. By importing a library, we can compute derivatives of functions that have been marked with @differentiable or that have been provided with a linear transpose function or a derivative function, but not of functions that have not been marked this way without defining a custom derivative for it. For example, if we try to differentiate sinf(_:) with the derivative(at:in:) API, the compiler will produce error messages at compile-time instead of producing zero derivatives.

let y = derivative(at: 1.0) { x in
    sinf(x)
}

test.swift:4:5: error: expression is not differentiable
    sinf(x)
    ^
test.swift:4:5: note: cannot differentiate functions that have not been marked '@differentiable' and that are defined in other modules
    sinf(x)
    ^

Non-differentiable type conversions

Calling functions that convert values to non-differentiable types and convert them back makes the function no longer differentiable. The compiler is able to detect these cases and provide error messages.

let d = derivative(at: 1.0) { x in
    Double(Int(x)) + 2
}

test.swift:1:27: error: function is not differentiable
let y = derivative(at: 1.0) { x in
                            ^~~~~~
test.swift:2:12: note: cannot differentiate through a non-differentiable result; do you want to add 'withoutDerivative(at:)'?
    Double(Int(x)) + 2
           ^

Accidental data flow mistakes

Even when there are no obvious non-differentiable operations on the path from parameters to the result (like non-differentiable type conversions), it is still possible to mistype a variable and cause numerical computation to be incorrect. As such, the compiler is able to leverage dependency analysis to determine whether the derivative is always zero and warns the user.

let grad = gradient(at: 1.0) { x in
    Double(3).squareRoot()
}

test.swift:4:18: warning: result does not depend on differentiation arguments and will always have a zero derivative; do you want to use 'withoutDerivative(at:)' to make it explicit?
    Double(3).squareRoot()
    ^
    withoutDerivative(at:)

Examples of differentiable programming

Linear regression

Linear Regression attempts to fit a line that best fits a set of data points. There are two different ways of finding a solution: the iterative and closed form methods. In the iterative method, we use gradient descent to slowly find better and better values for the slope and y-intercept. For a basic set of data points consisting of (x, y) value pairs, the model would look like the following:

struct Perceptron: @memberwise Differentiable {
    var weights: SIMD64<Float>
    var bias: Float

    @differentiable
    func callAsFunction(_ input: SIMD64<Float>) -> Float {
        weights.dot(input) + bias
    }
}

To train the model on a data set, it would look like the following:

let iterationCount = 160
let learningRate: Float = 0.00003

var model = Perceptron(weights: .zero, bias: 0)

for i in 0..<iterationCount {
    var (loss, 𝛁loss) = valueWithGradient(at: model) { model -> Float in
        var totalLoss: Float = 0
        for (x, y) in data {
            let pred = model(x)
            let diff = y - pred
            totalLoss = totalLoss + diff * diff / Float(data.count)
        }
        return totalLoss
    }
    𝛁loss.weight *= -learningRate
    𝛁loss.bias *= -learningRate
    model.move(along: 𝛁loss)
    if i.isMultiple(of: 10) {
        print("Iteration: \(iteration) Avg Loss: \(loss / Float(data.count))")
    }
}

Deep learning

Swift for TensorFlow is a numerics and machine learning library that uses the proposed differentiable programming feature. Swift for TensorFlow has been used to implement many machine learning models, from simple image classification models like ResNet to advanced models using Monte Carlo tree search to power a Go game engine.

Feed-forward neural networks (FFNN)

A neural networks is a “parameterized function approximator”: it takes some input, produces some output, and is parameterized by weights. Neural networks are composed of layers, which are smaller “building block” parameterized functions. A loss function (or cost function) measures the difference between the output of a neural network versus the expected output. Neural networks can improve via training: networks are applied to “training data” (input/output pairs) and parameters are updated with their derivatives with respect to the loss function.

A feed-forward neural network is a simple neural network in which the output of each layer is fed as the input to the next layer. A multi-layer perceptron is an example of a feed-forward neural network: it is composed of multiple dense layers, each of which performs output = activation(matmul(weight, input) + bias).

import TensorFlow

struct MultiLayerPerception: Layer, @memberwise Differentiable {
    var dense1 = Dense<Float>(inputSize: 784, outputSize: 100, activation: relu)
    var dense2 = Dense<Float>(inputSize: 100, outputSize: 30, activation: relu)
    var dense3 = Dense<Float>(inputSize: 30, outputSize: 10, activation: softmax)

    @differentiable
    func callAsFunction(_ input: Tensor<Float>) -> Tensor<Float> {
        dense3(dense2(dense1(input)))
    }
}

Convolutional neural networks (CNN)

A convolution neural network is a feed-forward neural network that performs a cross-correlation operation, which is a “sliding dot product” over the input. The cross-correlation operation encodes spatial locality and translation invariance, making CNNs suited for applications like image recognition.

Here is a simple script that implements LeNet-5, a convolutional neural network for classifying handwritten digits.

import TensorFlow

// Original Paper:
// "Gradient-Based Learning Applied to Document Recognition"
// Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner
// http://yann.lecun.com/exdb/publis/pdf/lecun-01a.pdf
//
// Note: this implementation connects all the feature maps in the second convolutional layer.
// Additionally, ReLU is used instead of sigmoid activations.
struct LeNet: Layer, @memberwise Differentiable {
    var conv1 = Conv2D<Float>(filterShape: (5, 5, 1, 6), padding: .same, activation: relu)
    var pool1 = AvgPool2D<Float>(poolSize: (2, 2), strides: (2, 2))
    var conv2 = Conv2D<Float>(filterShape: (5, 5, 6, 16), activation: relu)
    var pool2 = AvgPool2D<Float>(poolSize: (2, 2), strides: (2, 2))
    var flatten = Flatten<Float>()
    var fc1 = Dense<Float>(inputSize: 400, outputSize: 120, activation: relu)
    var fc2 = Dense<Float>(inputSize: 120, outputSize: 84, activation: relu)
    var fc3 = Dense<Float>(inputSize: 84, outputSize: 10, activation: softmax)

    @differentiable
    func callAsFunction(_ input: Tensor<Float>) -> Tensor<Float> {
        let convolved = pool2(conv2(pool1(conv1(input))))
        return fc3(fc2(fc1(flatten(convolved))))
    }
}

Recurrent neural networks (RNN)

A recurrent neural network is a feed-forward neural network wrapped in a loop over a sequence of inputs. The feed-forward neural network within the loop is usually referred to as the “cell” of the RNN. An RNN cell, like other neural network layers, has a callAsFunction(_:) method that is differentiable with respect to self and input, where input is an element of the sequence that is the input to the RNN.

/// A recurrent neural network cell.
protocol RNNCell: Layer
where Input == RNNCellInput<TimeStepInput, State>,
      Output == RNNCellOutput<TimeStepOutput, State> {
    /// The input at a time step.
    associatedtype TimeStepInput: Differentiable
    /// The output at a time step.
    associatedtype TimeStepOutput: Differentiable
    /// The state that may be preserved across time steps.
    associatedtype State: Differentiable
    /// The zero state.
    var zeroState: State { get }
}

Below is the cell of a long short-term memory (LSTM) network, which is used widely in natural language processing and speech processing.

/// An LSTM cell.
struct LSTMCell<Scalar: TensorFlowFloatingPoint>: RNNCell, @memberwise Differentiable {
    var fusedWeight: Tensor<Scalar>
    var fusedBias: Tensor<Scalar>

    @noDerivative var stateShape: TensorShape { [1, fusedWeight.shape[1] / 4] }

    var zeroState: State {
        State(cell: Tensor(zeros: stateShape), hidden: Tensor(zeros: stateShape))
    }

    typealias TimeStepInput = Tensor<Scalar>
    typealias TimeStepOutput = State
    typealias Input = RNNCellInput<TimeStepInput, State>
    typealias Output = RNNCellOutput<TimeStepOutput, State>

    struct State: @memberwise Differentiable {
        var cell: Tensor<Scalar>
        var hidden: Tensor<Scalar>
    }

    @differentiable
    func callAsFunction(_ input: Input) -> Output {
        let gateInput = input.input.concatenated(with: input.state.hidden, alongAxis: 1)
        let fused = matmul(gateInput, fusedWeight) + fusedBias
        let (batchSize, hiddenSize) = (fused.shape[0], fused.shape[1] / 4)
        let fusedParts = fused.split(count: 4, alongAxis: 1)
        let (inputGate, updateGate, forgetGate, outputGate) = (
            sigmoid(fusedParts[0]),
            tanh(fusedParts[1]),
            sigmoid(fusedParts[2]),
            sigmoid(fusedParts[3])
        )
        let newCellState = input.state.cell * forgetGate + inputGate * updateGate
        let newHiddenState = tanh(newCellState) * outputGate
        let newState = State(cell: newCellState, hidden: newHiddenState)
        return Output(output: newState, state: newState)
    }
}

Since an RNN is a loop wrapped around the cell, it can be implemented as a generic struct with a Cell generic parameter that conforms to RNNCell.

struct RNN<Cell: RNNCell>: Layer {
    typealias Input = [Cell.TimeStepInput]
    typealias Output = [Cell.TimeStepOutput]

    var cell: Cell

    init(_ cell: @autoclosure () -> Cell) {
        self.cell = cell()
    }

    @differentiable(wrt: (self, input))
    func callAsFunction(_ input: [Cell.TimeStepInput]) -> [Cell.TimeStepOutput] {
        var currentHiddenState = zeroState
        var timeStepOutputs: [Cell.TimeStepOutput] = []
        for timeStep in input {
            let output = cell(input: timeStep, state: currentHiddenState)
            currentHiddenState = output.state
            timeStepOutputs.append(output.output)
        }
        return timeStepOutputs
    }
}

Using generics, one can compose RNN with different RNN cell types. Different RNN types can be defined in a library simply by creating a type alias.

typealias SimpleRNN<Scalar: TensorFlowFloatingPoint> = RNN<SimpleRNNCell<Scalar>>
typealias LSTM<Scalar: TensorFlowFloatingPoint> = RNN<LSTMCell<Scalar>>

Future directions

Higher-order differentiation

Distinct from differentiation of higher-order functions, higher-order differentiation refers to taking the derivative of a derivative of a function. As a natural next step after the first-order differentiation capability proposed here, higher-order differentiation can be designed and implemented in various different ways with trade-offs in performance, usability, and complexity.

Intuitively, higher-order differentiation will enable calling a differential operator on the result of a differential operator, e.g.

let f = derivative(of: derivative(of: derivative(of: { x in pow(x, 3.0) })))

This will require the differential operator derivative(of:) to return a @differentiable function, hence semantically changing @differentiable to mean infinite differentiability.

func derivative<T: FloatingPoint, U: Differentiable>(
    _ f: @differentiable (T) -> U
) -> @differentiable (T) -> U where T: FloatingPoint, T == T.TangentVector {
    { x in differential(at: x, in: f) }
}

Since derivative(of:) is implemented in term of derivative(at:in:), which is implemented in terms of valueWithDifferential(at:in:), both derivative(at:in:) and valueWithDifferential(at:in:) would need to be marked with @differentiatiable with respect to its x argument.

@differentiable(wrt: x)
func derivative<T: FloatingPoint, U: Differentiable>(
    at x: T, in body: @differentiable (T) -> U) -> U
) -> U.TangentVector where T: FloatingPoint, T == T.TangentVector {
    valueWithDifferential(at: x, in: body).differential(T(1))
}

@differentiable(wrt: x)
func valueWithDifferential<T: FloatingPoint, U: Differentiable>(
    at x: T, in body: @differentiable (T) -> U) -> U
) -> (value: U, differential: @differentiable(linear) (T.TangentVector) -> U.TangentVector)

To differentiate valueWithDifferential, we need to be able to differentiate its return value, a tuple of the original value and the differential, with respect to its x argument.

A kneejerk solution is to differentiate derivative functions generated by the differentiation transform at compile-time, but this leads to problems. For example, how do we repeatedly differentiate a function whose body is unavailable? Should a @differentiable function contain derivative functions for dynamically many orders? Would it require serializing SIL code as part of a @differentiable function and running the differentiation transform at runtime? Alternatively, is there a single closed-form formula that the compiler can generate once in the differentiation transform, without performing any runtime compilation or using large function representations? These questions are difficult to answer, due to the complexity in both mathematical formulae (e.g. Faà di Bruno's formula) and static compilation. Currently, we are exploring different theoretical and practical approaches to find a beautiful design that would help us deliver the best differentiable programming language.

Naming conventions for numerical computing

The API Design Guidelines encourages naming that is both easy-to-learn for beginners and unsurprising for experts.

Numerical computing is full of math terminology and notation; finding good names for math concepts is not always easy. Consider the formulas for gated recurrent neural networks:

Each of these mathematical variables needs a name in code. Consider the following names for the W_ih variable:

• var W_ih: the abbreviated name. May be difficult to learn for beginners.
• var inputHiddenWeight: the descriptive name. May be unfamiliar for experts, who are accustomed to the math notation.

Which name is the best? It is hard to say, as no naming precedent exists. Standardizing naming conventions for math terminology will be important as numerical computing becomes more prominent in Swift.

Source compatibility

This feature does not change any existing APIs. New implicit function conversions are added to the type system, which slightly increases type checking complexity. We have not observed source compatibility breakages so far.

Effect on ABI stability

This feature has additions to the ABI. Specifically, the @differentiable function representation will be added and must be kept stable.

Effect on API resilience

This feature adds the Differentiable protocol and differential operators to the standard library as public APIs. They introduce additions to the standard library.

Differentiable protocol

The Differentiable protocol contains all necessary requirements for a type to be differentiated. Without breaking API, it will be possible to add extensions to the Differentiable protocol and add new requirements with default implementations.

Differential operators

Differential operators (e.g. derivative(of:) and gradient(of:)) are added to the standard library as lightweight top-level higher-order functions. These APIs can be renamed or moved under some namespace without breaking ABI.

Alternatives considered

Not support differentiable programming

We believe first-class differentiable programming is a big step towards making Swift a real contender in the numerical computing and machine learning landscape. Differentiable programming will enable intelligent applications, machine learning models, scientific experiments, physical simulations, and more.

Use another language or framework for differentiable programming

Dynamic languages, like Python and Julia, have established library support for differentiable programming. While it is possible to interoperate with these libraries via Swift, we feel that first-class differentiable programming in Swift is leaps ahead in expressivity, usability, and safety.

Other approaches to differentiable programming

See “Approaches to automatic differentiation” above for an overview and comparison of automatic differentiation approaches. First-class language support for differentiation will enable convenient, extensible, and performant differentiable programming in Swift - more so than library-based approaches.

Acknowledgements

Many people have influenced the design and the implementation of the differentiable programming feature. The authors would like to thank these people (sorted alphabetically by last name) for their contributions in any form (inspirations, ideas, discussions, code, or bikeshedding): Gogul Balakrishnan, James Bradbury, Steve Canon, Casey Chu, Conal Elliott, Roy Frostig, Doug Gregor, Dominik Grewe, Dmitri Gribenko, Joe Groff, Sylvain Gugger, Tim Harley, Matthew Johnson, Chris Lattner, Dougal Maclaurin, John McCall, Bart van Merriënboer, Slava Pestov, Anthony Platanios, Gordon Plotkin, Alexey Radul, Brennan Saeta, Parker Schuh, and Dimitrios Vytiniotis.