I know this sounds crazy for a 10-year-old, but because S4TF doesn't work for me, I'm building my own neural network library in Swift. (I haven't gotten that far.) I'm creating a structure that conforms to AdditiveArithmetic
. It also uses Philip Turner's Differentiable
, but that's unimportant.
Anyway, when defining the zero
variable, I call another variable dimen
, defined in the same structure. This raises an error: instance member 'dimen' cannot be used on type 'Electron<T>'
note: the structure I am creating is going to be used to create a multi-dimensional array for neural networks.
Total code (stripped down to remove unimportant bits):
public struct Electron<T> where T: ExpressibleByFloatLiteral, T: AdditiveArithmetic {
var energy: [[Int]: T] = [:]
var dimen: [Int]
public init(_ dim: [Int], with: ElectronInitializer) {
self.dimen = dim
self.energy = [:]
var curlay = [Int](repeating: 0, count: dimen.count)
curlay[curlay.count-1] = -1
while true {
var max: Int = -1
for x in 0..<curlay.count {
if curlay[curlay.count-1-x] == dimen[curlay.count-1-x]-1 {
max = curlay.count-1-x
}
else {break}
}
if max == 0 {break}
else if max != -1 {
for n in max..<curlay.count {
curlay[n] = -1
}
curlay[max-1] = 1
}
curlay[curlay.count-1] = 1
print(curlay)
energy[curlay] = { () -> T in
switch with {
case ElectronInitializer.repeating(let value):
return value as! T
case ElectronInitializer.random(let minimum, let maximum):
return Double.random(in: minimum..<maximum) as! T
}
}()
}
}
subscript(_ coordinate: Int...) -> T {
var convertList: [Int] = []
for conversion in coordinate {
convertList.append(conversion)
}
return self.energy[convertList]!
}
public mutating func setQuantum(_ replace: T, at: [Int]) {
self.energy[at]! = replace
}
}
extension Electron: AdditiveArithmetic {
public static func - (lhs: Electron<T>, rhs: Electron<T>) -> Electron<T> where T: AdditiveArithmetic, T: ExpressibleByFloatLiteral {
var output: Electron<T> = lhs
for value in lhs.energy {
output.energy[value.key] = output.energy[value.key]!-rhs.energy[value.key]!
}
return output
}
public static var zero: Electron<T> {
return Electron.init(dimen, with: ElectronInitializer.repeating(0.0))
}
static prefix func (x: Electron) -> Electron {
return x
}
public static func (lhs: Electron<T>, rhs: Electron<T>) -> Electron<T> where T: AdditiveArithmetic, T: ExpressibleByFloatLiteral {
var output: Electron<T> = lhs
for value in lhs.energy {
output.energy[value.key] = output.energy[value.key]! rhs.energy[value.key]!
}
return output
}
}
public enum ElectronInitializer {
case random(Double, Double)
case repeating(Double)
}
Error:
NeuralNetwork.xcodeproj:59:30: error: instance member 'dimen' cannot be used on type 'Electron' return Electron.init(dimen, with: ElectronInitializer.repeating(0.0))
I don't know what's happening, but thanks in advance. I'm new to Stack Overflow, so sorry if I did something wrong.
CodePudding user response:
The root of the problem is that dimen
is an instance property, while zero
is a static property. In a static
context, you don't have an instance from which to access dimen
, and so the compiler gives you the error. static
properties and methods are a lot like global variables and free-functions with respect to accessing instance properties and methods. You'd have to make an instance available somehow. For a static
function, you could pass it in, but for a static
computed property, you'd either have to store an instance in a stored static
property, which isn't allowed for generics, or you'd have to store it in a global variable, which isn't good either, and would be tricky to make work for all the possible T
.
There are ways to do what you need though. They all involve implementing some special behavior for a zero Electron
rather than relying on access to an instance property in your static
.zero
implementation. I made some off-the-cuff suggestions in comments, which would work; however, I think a more elegant solution is to solve the problem by creating a custom type for energy
, which would require very few changes to your existing code. Specifically you could make an Energy
type nested in your Electron
type:
internal struct Energy: Equatable, Sequence {
public typealias Value = T
public typealias Key = [Int]
public typealias Element = (key: Key, value: Value)
public typealias Storage = [Key: Value]
public typealias Iterator = Storage.Iterator
public typealias Keys = Storage.Keys
public typealias Values = Storage.Values
private var storage = Storage()
public var keys: Keys { storage.keys }
public var values: Values { storage.values }
public var count: Int { storage.count }
public init() { }
public subscript (key: Key) -> Value? {
get { storage.isEmpty ? .zero : storage[key] }
set { storage[key] = newValue }
}
public func makeIterator() -> Iterator {
storage.makeIterator()
}
}
The idea here is that when energy.storage
is empty, it returns 0 for any key, which allows you to use it as a .zero
value. I've made it internal
, because energy
defaults to internal
, and so I've done a minimalist job of wrapping a Dictionary
, mainly providing subscript
operator, and making it conform to Sequence
, which is all that is needed by code you provided.
The only changes needed in the rest of your code are to change the definition of energy
var energy: Energy
Then to set it in your initializer, by-passing the bulk of your init
when dim
is empty.
public init(_ dim: [Int], with: ElectronInitializer) {
self.dimen = dim
self.energy = Energy() // <- Initialize `energy`
// Empty dim indicates a zero electron which doesn't need the
// rest of the initialization
guard dim.count > 0 else { return }
var curlay = [Int](repeating: 0, count: dimen.count)
curlay[curlay.count-1] = -1
while true {
var max: Int = -1
for x in 0..<curlay.count {
if curlay[curlay.count-1-x] == dimen[curlay.count-1-x]-1 {
max = curlay.count-1-x
}
else {break}
}
if max == 0 {break}
else if max != -1 {
for n in max..<curlay.count {
curlay[n] = -1
}
curlay[max-1] = 1
}
curlay[curlay.count-1] = 1
print(curlay)
energy[curlay] = { () -> T in
switch with {
case ElectronInitializer.repeating(let value):
return value as! T
case ElectronInitializer.random(let minimum, let maximum):
return Double.random(in: minimum..<maximum) as! T
}
}()
}
}
And then of course, to change how you create it in your zero
property
public static var zero: Electron<T> {
return Electron.init([], with: ElectronInitializer.repeating(0.0))
}
ElectronInitializer
isn't actually used in this case. It's just a required parameter of your existing init
. This suggests an opportunity to refactor initialization, so you could have an init()
that creates a zero Electron
in addition to your existing init(dim:with:)