Performance 求解F#中的背包问题：性能

我发现一篇文章：

关于背包问题在F#中的实现。在我学习这门语言的过程中，我发现这真的很有趣，并尝试对此进行一些研究。以下是我编写的代码：

open System
open System.IO 
open System.Collections.Generic

let parseToTuple (line : string) =
    let parsedLine = line.Split(' ') |> Array.filter(not << String.IsNullOrWhiteSpace)         |> Array.map Int32.Parse
    (parsedLine.[0], parsedLine.[1])

let memoize f =
    let cache = Dictionary<_, _>()
    fun x ->
        if cache.ContainsKey(x)
            then cache.[x]
        else
            let res = f x
            cache.[x] <- res
            res

type Item =
    {
        Value : int
        Size  : int
    }  

type ContinuationBuilder() = 
    member b.Bind(x, f) = fun k -> x (fun x -> f x k)
    member b.Return x = fun k ->  k x
    member b.ReturnFrom x = x

let cont = ContinuationBuilder()

let set1 =
    [
        (4, 11)
        (8, 4)
        (10, 5)
        (15, 8)
        (4, 3)
    ]

let set2 =
    [
        (50, 341045); (1906, 4912); (41516, 99732); (23527, 56554); (559, 1818); (45136, 108372); (2625, 6750); (492, 1484)
        (1086, 3072); (5516, 13532); (4875, 12050); (7570, 18440); (4436, 10972); (620, 1940); (50897, 122094); (2129, 5558)
        (4265, 10630); (706, 2112); (2721, 6942); (16494, 39888); (29688, 71276); (3383, 8466); (2181, 5662); (96601, 231302)
        (1795, 4690); (7512, 18324); (1242, 3384); (2889, 7278); (2133, 5566); (103, 706); (4446, 10992); (11326, 27552)
        (3024, 7548); (217, 934); (13269, 32038); (281, 1062); (77174, 184848); (952, 2604); (15572, 37644); (566, 1832)
        (4103, 10306); (313, 1126); (14393, 34886); (1313, 3526); (348, 1196); (419, 1338); (246, 992); (445, 1390)
        (23552, 56804); (23552, 56804); (67, 634)
    ]

[<EntryPoint>] 
let main args =
    // prepare list of items from a file args.[0]
    let header, items = set1
                        |> function
                           | h::t -> h, t
                           | _    -> raise (Exception("Wrong data format"))

    let N, K = header
    printfn "N = %d, K = %d" N K
    let items = List.map (fun x -> {Value = fst x ; Size = snd x}) items |> Array.ofList

    let rec combinations =
        let innerSolver key =
            cont
                {
                    match key with
                    | (i, k) when i = 0 || k = 0        -> return 0
                    | (i, k) when items.[i-1].Size > k  -> return! combinations (i-1, k)
                    | (i, k)                            -> let item = items.[i-1]
                                                           let! v1 = combinations (i-1, k)
                                                           let! beforeItem = combinations (i-1, k-item.Size)
                                                           let v2 = beforeItem + item.Value
                                                           return max v1 v2
                }
        memoize innerSolver

    let res = combinations (N, K) id
    printfn "%d" res
    0

开放系统
开放系统
open System.Collections.Generic
让parseToTuple（行：string）=
让parsedLine=line.Split（“”）|>Array.filter（而不是Array.map Int32.Parse
（解析线[0]，解析线[1]）
让我们回忆一下f=
让缓存=字典（）
乐趣x->
if cache.ContainsKey（x）
然后缓存
其他的
设res=fx
缓存.x]x（乐趣x->f x k）
成员b.返回x=fun k->k x
成员b.ReturnFrom x=x
设cont=ContinuationBuilder（）
让set1=
[
(4, 11)
(8, 4)
(10, 5)
(15, 8)
(4, 3)
]
让set2=
[
(50, 341045); (1906, 4912); (41516, 99732); (23527, 56554); (559, 1818); (45136, 108372); (2625, 6750); (492, 1484)
(1086, 3072); (5516, 13532); (4875, 12050); (7570, 18440); (4436, 10972); (620, 1940); (50897, 122094); (2129, 5558)
(4265, 10630); (706, 2112); (2721, 6942); (16494, 39888); (29688, 71276); (3383, 8466); (2181, 5662); (96601, 231302)
(1795, 4690); (7512, 18324); (1242, 3384); (2889, 7278); (2133, 5566); (103, 706); (4446, 10992); (11326, 27552)
(3024, 7548); (217, 934); (13269, 32038); (281, 1062); (77174, 184848); (952, 2604); (15572, 37644); (566, 1832)
(4103, 10306); (313, 1126); (14393, 34886); (1313, 3526); (348, 1196); (419, 1338); (246, 992); (445, 1390)
(23552, 56804); (23552, 56804); (67, 634)
]
[] 
让主参数=
//准备文件args中的项目列表。[0]
let header，items=set1
|>作用
|h:：t->h，t
|->raise（异常（“错误数据格式”））
设N，K=header
printfn“N=%d，K=%d”N K
让items=List.map（funx->{Value=fstx；Size=sndx}）items}>Array.ofList
let rec组合=
让innerSolver键=
续
{
匹配键
|（i，k）当i=0 | | k=0时->返回0
|（i，k）当项目[i-1]。大小>k->返回！组合（i-1，k）
|（i，k）->设项=项。[i-1]
设！v1=组合（i-1，k）
let！beforeItem=组合（i-1，k-item.Size）
设v2=beforeItem+item.Value
返回最大v1 v2
}
记忆内部解算器
设res=组合（N，K）id
printfn“%d”res
0

然而，这个实现的问题是速度非常慢（实际上，我无法解决50个项目和大约300000个容量的问题，这是我用C#在不到1s的时间内简单实现的）

---

你能告诉我我是否在某个地方犯了错误吗？或者可能实现是正确的，这只是解决此问题的低效方法。

当你天真地应用这样一个通用的回忆体，并使用延续传递时，回忆体缓存中的值是延续，而不是常规的“最终值”结果。因此，当你得到一个缓存命中时，你没有得到一个最终的结果，你得到的是一个函数，该函数承诺在你调用它时计算结果。这个调用可能会很昂贵，可能会调用其他各种延续，可能最终会再次命中记忆缓存本身，等等

有效地记忆连续传递函数，这样a）缓存工作完全有效，b）函数保持尾部递归是相当困难的。阅读讨论，当你完全理解它时再回来

你链接的博客文章的作者正在使用一种更复杂、更不通用的回忆录，它特别适合这个问题。诚然，我还没有完全研究它（博客上的代码不完整/不完整，很难尝试），但我认为它的要点是，它在缓存最终整数结果之前“强制”连续链

为了说明这一点，这里有一个代码的快速重构，它是完全自包含的，并跟踪相关信息：

open System
open System.Collections.Generic

let mutable cacheHits = 0
let mutable cacheMisses = 0

let memoize f =
    let cache = Dictionary<_, _>()
    fun x ->
        match cache.TryGetValue(x) with
        | (true, v) -> 
            cacheHits <- cacheHits + 1
            printfn "Hit for %A - Result is %A" x v
            v
        | _ ->
            cacheMisses <- cacheMisses + 1
            printfn "Miss for %A" x
            let res = f x
            cache.[x] <- res
            res

type Item = { Value : int; Size  : int }  

type ContinuationBuilder() = 
    member b.Bind(x, f) = fun k -> x (fun x -> f x k)
    member b.Return x = fun k ->  k x
    member b.ReturnFrom x = x

let cont = ContinuationBuilder()

let genItems n = 
   [| for i = 1 to n do
         let size = i % 5
         let value = (size * i)
         yield { Value = value; Size = size }
   |]

let N, K = (5, 100)
printfn "N = %d, K = %d" N K

let items = genItems N

let rec combinations_cont =
    memoize (
     fun key ->
       cont {
                match key with
                | (0, _) | (_, 0)                   -> return 0
                | (i, k) when items.[i-1].Size > k  -> return! combinations_cont (i - 1, k) 
                | (i, k)                            -> let item = items.[i-1]
                                                       let! v1 = combinations_cont (i-1, k)
                                                       let! beforeItem = combinations_cont (i-1, k - item.Size)
                                                       let v2 = beforeItem + item.Value
                                                       return max v1 v2
        }
    )

let res = combinations_cont (N, K) id
printfn "Answer: %d" res
printfn "Memo hits: %d" cacheHits
printfn "Memo misses: %d" cacheMisses
printfn ""

let rec combinations_plain =
    memoize (
     fun key ->
                match key with
                | (i, k) when i = 0 || k = 0        -> 0
                | (i, k) when items.[i-1].Size > k  -> combinations_plain (i-1, k) 
                | (i, k)                            -> let item = items.[i-1]
                                                       let v1 = combinations_plain (i-1, k)
                                                       let beforeItem = combinations_plain (i-1, k-item.Size)
                                                       let v2 = beforeItem + item.Value
                                                       max v1 v2
    )

cacheHits <- 0
cacheMisses <- 0

let res2 = combinations_plain (N, K)
printfn "Answer: %d" res2
printfn "Memo hits: %d" cacheHits
printfn "Memo misses: %d" cacheMisses

开放系统
open System.Collections.Generic
设可变缓存命中数=0
设可变缓存未命中=0
让我们回忆一下f=
让缓存=字典（）
乐趣x->
将cache.TryGetValue（x）与
|（对，v）->
缓存命中
缓存未命中（f x k）
成员b.返回x=fun k->k x
成员b.ReturnFrom x=x
设cont=ContinuationBuilder（）
设genitem n=
[|对于i=1到n do
让大小=i%5
让值=（大小*i）
产量{Value=Value；Size=Size}
|]
设N，K=（5100）
printfn“N=%d，K=%d”N K
设items=genItems N
让记录组合\u cont=
回忆(
趣味键->
续{
匹配键
|（0，124;（124;，0）->返回0
|（i，k）当项目[i-1].大小>k->返回！组合\u cont（i-1，k）
|（i，k）->设项=项。[i-1]
设！v1=组合（i-1，k）
let！beforeItem=组合（i-1，k-项目大小）
设v2=beforeItem+item.Value
返回最大v1 v2
}
)
设res=cont（N，K）id的组合
printfn“答案：%d”res
printfn“备忘录命中率：%d”缓存命中率
printfn“备忘录未命中：%d”缓存未命中
printfn
open System
open System.Diagnostics
open System.Collections.Generic

let time f =
    System.GC.Collect()
    let sw = Stopwatch.StartNew()
    let r = f()
    sw.Stop()
    printfn "Took: %f" sw.Elapsed.TotalMilliseconds
    r

let mutable cacheHits = 0
let mutable cacheMisses = 0

let memoize f =
    let cache = Dictionary<_, _>()
    fun x ->
        match cache.TryGetValue(x) with
        | (true, v) -> 
            cacheHits <- cacheHits + 1
            //printfn "Hit for %A - Result is %A" x v
            v
        | _ ->
            cacheMisses <- cacheMisses + 1
            //printfn "Miss for %A" x
            let res = f x
            cache.[x] <- res
            res

type Item = { Value : int; Size  : int }  

type ContinuationBuilder() = 
    member b.Bind(x, f) = fun k -> x (fun x -> f x k)
    member b.Return x = fun k ->  k x
    member b.ReturnFrom x = x

let cont = ContinuationBuilder()

let genItems n = 
    [| for i = 1 to n do
            let size = i % 5
            let value = (size * i)
            yield { Value = value; Size = size }
    |]

let N, K = (80, 400)
printfn "N = %d, K = %d" N K

let items = genItems N

//let rec combinations_cont =
//    memoize (
//     fun key ->
//       cont {
//                match key with
//                | (0, _) | (_, 0)                   -> return 0
//                | (i, k) when items.[i-1].Size > k  -> return! combinations_cont (i - 1, k) 
//                | (i, k)                            -> let item = items.[i-1]
//                                                       let! v1 = combinations_cont (i-1, k)
//                                                       let! beforeItem = combinations_cont (i-1, k - item.Size)
//                                                       let v2 = beforeItem + item.Value
//                                                       return max v1 v2
//        }
//    )
//
//
//cacheHits <- 0
//cacheMisses <- 0

//let res = time(fun () -> combinations_cont (N, K) id)
//printfn "Answer: %d" res
//printfn "Memo hits: %d" cacheHits
//printfn "Memo misses: %d" cacheMisses
//printfn ""

let rec combinations_plain =
    memoize (
        fun key ->
                match key with
                | (i, k) when i = 0 || k = 0        -> 0
                | (i, k) when items.[i-1].Size > k  -> combinations_plain (i-1, k) 
                | (i, k)                            -> let item = items.[i-1]
                                                       let v1 = combinations_plain (i-1, k)
                                                       let beforeItem = combinations_plain (i-1, k-item.Size)
                                                       let v2 = beforeItem + item.Value
                                                       max v1 v2
    )

cacheHits <- 0
cacheMisses <- 0

printfn "combinations_plain"
let res2 = time (fun () -> combinations_plain (N, K))
printfn "Answer: %d" res2
printfn "Memo hits: %d" cacheHits
printfn "Memo misses: %d" cacheMisses
printfn ""

let recursivelyMemoize f =
    let cache = Dictionary<_, _>()
    let rec memoizeAux x =
        match cache.TryGetValue(x) with
        | (true, v) -> 
            cacheHits <- cacheHits + 1
            //printfn "Hit for %A - Result is %A" x v
            v
        | _ ->
            cacheMisses <- cacheMisses + 1
            //printfn "Miss for %A" x
            let res = f memoizeAux x
            cache.[x] <- res
            res
    memoizeAux

let combinations_plain2 =
    let combinations_plain2Aux combinations_plain2Aux key =
                match key with
                | (i, k) when i = 0 || k = 0        -> 0
                | (i, k) when items.[i-1].Size > k  -> combinations_plain2Aux (i-1, k) 
                | (i, k)                            -> let item = items.[i-1]
                                                       let v1 = combinations_plain2Aux (i-1, k)
                                                       let beforeItem = combinations_plain2Aux (i-1, k-item.Size)
                                                       let v2 = beforeItem + item.Value
                                                       max v1 v2
    let memoized = recursivelyMemoize combinations_plain2Aux
    fun x -> memoized x

cacheHits <- 0
cacheMisses <- 0

printfn "combinations_plain2"
let res3 = time (fun () -> combinations_plain2 (N, K))
printfn "Answer: %d" res3
printfn "Memo hits: %d" cacheHits
printfn "Memo misses: %d" cacheMisses
printfn ""

let recursivelyMemoizeCont f =
    let cache = Dictionary HashIdentity.Structural
    let rec memoizeAux x k =
        match cache.TryGetValue(x) with
        | (true, v) -> 
            cacheHits <- cacheHits + 1
            //printfn "Hit for %A - Result is %A" x v
            k v
        | _ ->
            cacheMisses <- cacheMisses + 1
            //printfn "Miss for %A" x
            f memoizeAux x (fun y ->
                cache.[x] <- y
                k y)
    memoizeAux

let combinations_cont2 =
    let combinations_cont2Aux combinations_cont2Aux key =
        cont {
                match key with
                | (0, _) | (_, 0)                   -> return 0
                | (i, k) when items.[i-1].Size > k  -> return! combinations_cont2Aux (i - 1, k) 
                | (i, k)                            -> let item = items.[i-1]
                                                       let! v1 = combinations_cont2Aux (i-1, k)
                                                       let! beforeItem = combinations_cont2Aux (i-1, k - item.Size)
                                                       let v2 = beforeItem + item.Value
                                                       return max v1 v2
        }
    let memoized = recursivelyMemoizeCont combinations_cont2Aux
    fun x -> memoized x id

cacheHits <- 0
cacheMisses <- 0

printfn "combinations_cont2"
let res4 = time (fun () -> combinations_cont2 (N, K))
printfn "Answer: %d" res4
printfn "Memo hits: %d" cacheHits
printfn "Memo misses: %d" cacheMisses
printfn ""

N = 80, K = 400
combinations_plain
Took: 7.191000
Answer: 6480
Memo hits: 6231
Memo misses: 6552

combinations_plain2
Took: 6.310800
Answer: 6480
Memo hits: 6231
Memo misses: 6552

combinations_cont2
Took: 17.021200
Answer: 6480
Memo hits: 6231
Memo misses: 6552