Mathias Brandewinder on .NET, F#, VSTO and Excel development, and quantitative analysis / machine learning.
22. March 2015 17:27

I will admit it, I got a bit upset by James McCaffrey’s column in MSDN magazine this month, “Gradient Descent Training Using C#”. While the algorithm explanations are quite good, I was disappointed by the C# sample code, and kept thinking to myself “why oh why isn’t this written in F#”. This is by no means intended as a criticism of C#; it’s a great language, but some problems are just better suited for different languages, and in this case, I couldn’t fathom why F# wasn’t used.

Long story short, I just couldn’t let it go, and thought it would be interesting to take that C# code, and do a commented rewrite in F#. I won’t even go into why the code does what it does – the article explains it quite well – but will instead purely focus on the implementation, and will try to keep it reasonably close to the original, at the expense of some additional nifty things that could be done.

The general outline of the code follows two parts:

• Create a synthetic dataset, creating random input examples, and computing the expected result using a known function,
• Use gradient descent to learn the model parameters, and compare them to the true value to check whether the method is working.

You can download the original C# code here. Today we’ll focus only on the first part, which is mainly contained in two methods, MakeAllData and MakeTrainTest:

static double[][] MakeAllData(int numFeatures, int numRows, int seed)
{
Random rnd = new Random(seed);
double[] weights = new double[numFeatures + 1]; // inc. b0
for (int i = 0; i < weights.Length; ++i)
weights[i] = 20.0 * rnd.NextDouble() - 10.0; // [-10.0 to +10.0]

double[][] result = new double[numRows][]; // allocate matrix
for (int i = 0; i < numRows; ++i)
result[i] = new double[numFeatures + 1]; // Y in last column

for (int i = 0; i < numRows; ++i) // for each row
{
double z = weights[0]; // the b0
for (int j = 0; j < numFeatures; ++j) // each feature / column except last
{
double x = 20.0 * rnd.NextDouble() - 10.0; // random X in [10.0, +10.0]
result[i][j] = x; // store x
double wx = x * weights[j + 1]; // weight * x
z += wx; // accumulate to get Y
}
double y = 1.0 / (1.0 + Math.Exp(-z));
if (y > 0.55)  // slight bias towards 0
result[i][numFeatures] = 1.0; // store y in last column
else
result[i][numFeatures] = 0.0;
}
Console.WriteLine("Data generation weights:");
ShowVector(weights, 4, true);

return result;
}

MakeAllData takes a number of features and rows, and a seed for the random number generator so that we can replicate the same dataset repeatedly. The dataset is represented as an array of array of doubles. The first columns, from 0 to numFeatures – 1, contain random numbers between –10 and 10. The last column contains a 0 or a 1. What we are after here is a classification model: each row can take two states (1 or 0), and we are trying to predict them from observing the features. In our case, that value is computed using a logistic model: we have a set of weights (which we also generate randomly), corresponding to each feature, and the output is

logistic [ x1; x2; … xn ] = 1.0 / (1.0 + exp ( - (w0 * 1.0 + w1 * x1 + w2 * x2 + … + wn * xn))

Note that w0 plays the role of a constant term in the equation, and is multiplied by 1.0 all the time. This is adding some complications to a code where indices are already flying left and right, because now elements in the weights array are mis-aligned by one element with the elements in the features array. Personally, I also don’t like adding another column to contain the predicted value, because that’s another implicit piece of information we have to remember.

In that frame, I will make two minor changes here, just to keep my sanity. First, as is often done, we will insert a column containing just 1.0 in each observation, so that the weights and features are now aligned. Then, we will move the 0s and 1s outside of the features array, to avoid any ambiguity.

Good. Instead of creating a Console application, I’ll simply go for a script. That way, I can just edit my code and check live whether it does what I want, rather than recompile and run every time.

Let’s start with the weights. What we are doing here is simply creating an array of numFeatures + 1 elements, populated by random values between –10.0 and 10.0. We’ll go a bit fancy here: given that we are also generating random numbers the same way a bit further down, let’s extract a function that generates numbers uniformly between a low and high value:

let rnd = Random(seed)
let generate (low,high) = low + (high-low) * rnd.NextDouble()
let weights = Array.init (numFeatures + 1) (fun _ -> generate(-10.0,10.0))

The next section is where things get a bit thornier. The C# code creates an array, then populates it row by row, first filling in the columns with random numbers, and then applying the logistic function to compute the value that goes in the last column. We can make that much clearer, by extracting that function out. The logistic function is really doing 2 things:

• first, the sumproduct of 2 arrays,
• and then, 1.0/(1.0 + exp ( – z ).

That is easy enough to implement:

let sumprod (v1:float[]) (v2:float[]) =
Seq.zip v1 v2 |> Seq.sumBy (fun (x,y) -> x * y)

let sigmoid z = 1.0 / (1.0 + exp (- z))

let logistic (weights:float[]) (features:float[]) =
sumprod weights features |> sigmoid

We can now use all this, and generate a dataset by simply first creating rows of random values (with a 1.0 in the first column for the constant term), applying the logistic function to compute the value for that row, and return them as a tuple:

open System

let sumprod (v1:float[]) (v2:float[]) =
Seq.zip v1 v2 |> Seq.sumBy (fun (x,y) -> x * y)

let sigmoid z = 1.0 / (1.0 + exp (- z))

let logistic (weights:float[]) (features:float[]) =
sumprod weights features |> sigmoid

let makeAllData (numFeatures, numRows, seed) =

let rnd = Random(seed)
let generate (low,high) = low + (high-low) * rnd.NextDouble()
let weights = Array.init (numFeatures + 1) (fun _ -> generate(-10.0,10.0))

let dataset =
[| for row in 1 .. numRows ->
let features =
[|
yield 1.0
for feat in 1 .. numFeatures -> generate(-10.0,10.0)
|]
let value =
if logistic weights features > 0.55
then 1.0
else 0.0
(features, value)
|]

weights, dataset

Done. Let’s move to the second part of the data generation, with the MakeTrainTest method. Basically, what this does is take a dataset, shuffle it, and split it in two parts, 80% which we will use for training, and 20% we leave out for validation.

static void MakeTrainTest(double[][] allData, int seed,
out double[][] trainData, out double[][] testData)
{
Random rnd = new Random(seed);
int totRows = allData.Length;
int numTrainRows = (int)(totRows * 0.80); // 80% hard-coded
int numTestRows = totRows - numTrainRows;
trainData = new double[numTrainRows][];
testData = new double[numTestRows][];

double[][] copy = new double[allData.Length][]; // ref copy of all data
for (int i = 0; i < copy.Length; ++i)
copy[i] = allData[i];

for (int i = 0; i < copy.Length; ++i) // scramble order
{
int r = rnd.Next(i, copy.Length); // use Fisher-Yates
double[] tmp = copy[r];
copy[r] = copy[i];
copy[i] = tmp;
}
for (int i = 0; i < numTrainRows; ++i)
trainData[i] = copy[i];

for (int i = 0; i < numTestRows; ++i)
testData[i] = copy[i + numTrainRows];
}

Again, there is a ton of indexing going on, which in my old age I find very hard to follow. Upon closer inspection, really, the only thing complicated here is the Fischer-Yates shuffle, which takes an array and randomly shuffles the order. The rest is pretty simply – we just want to shuffle, and then split into two arrays. Let’s extract the shuffle code (which happens to also be used and re-implemented later on):

let shuffle (rng:Random) (data:_[]) =
let copy = Array.copy data
for i in 0 .. (copy.Length - 1) do
let r = rng.Next(i, copy.Length)
let tmp = copy.[r]
copy.[r] <- copy.[i]
copy.[i] <- tmp
copy

We went a tiny bit fancy again here, and made the shuffle work on generic arrays; we also pass in the Random instance we want to use, so that we can control / repeat shuffles if we want, by passing a seeded Random. Does this work? Let’s check in FSI:

> [| 1 .. 10 |] |> shuffle (Random ());;
val it : int [] = [|6; 7; 2; 10; 8; 5; 4; 9; 3; 1|]

Looks reasonable. Let’s move on – we can now implement the makeTrainTest function.

let makeTrainTest (allData:_[], seed) =

let rnd = Random(seed)
let totRows = allData.Length
let numTrainRows = int (float totRows * 0.80) // 80% hard-coded

let copy = shuffle rnd allData
copy.[.. numTrainRows-1], copy.[numTrainRows ..]

Done. A couple of remarks here. First, F# is a bit less lenient than C# around types, so we have to be explicit when converting the number of rows to 80%, first to float, then back to int. As an aside, this used to annoy me a bit in the beginning, but I have come to really like having F# as this slightly psycho-rigid friend who nags me when I am taking a dangerous path (for instance, dividing two integers and hoping for a percentage).

Besides that, I think the code is markedly clearer. The complexity of the shuffle has been nicely contained, and we just have to slice the array to get a training and test sets. As an added bonus, we got rid of the out parameters, and that always feels nice and fuzzy.

I’ll leave it at for today; next time we’ll look at the second part, the learning algorithm itself. Before closing shop, let me make a couple of comments. First, the code is a tad shorter, but not by much. I haven’t really tried, and deliberately made only the changes I thought were needed. What I like about it, though, is that all the indexes are gone, except for the shuffle. In my opinion, this is a good thing. I find it difficult to keep it all in my head when more than one index is involved; when I need to also remember what columns contain special values, I get worried – and just find it hard to figure out what is going on. By contrast, I think makeTrainTest, for instance, conveys pretty directly what it does. makeAllData, in spite of some complexity, also maps closely the way I think about my goal: “I want to generate rows of inputs” – this is precisely what the code does. There is probably an element of culture to it, though; looping over arrays has a long history, and is familiar to every developer, and what looks readable to me might look entirely weird to some.

Easier, or more complicated than before? Anything you like or don’t like – or find unclear? Always interested to hear your opinion! Ping me on Twitter if you have comments.

15. February 2014 12:51

My favorite column in MSDN Magazine is Test Run; it was originally focused on testing, but the author, James McCaffrey, has been focusing lately on topics revolving around numeric optimization and machine learning, presenting a variety of methods and approaches. I quite enjoy his work, with one minor gripe –his examples are all coded in C#, which in my opinion is really too bad, because the algorithms would gain much clarity if written in F# instead.

Back in June 2013, he published a piece on Amoeba Method Optimization using C#. I hadn’t seen that approach before, and found it intriguing. I also found the C# code a bit too hairy for my feeble brain to follow, so I decided to rewrite it in F#.

In a nutshell, the Amoeba approach is a heuristic to find the minimum of a function. Its proper respectable name is the Nelder-Nead method. The reason it is also called the Amoeba method is because of the way the algorithm works: in its simple form, it starts from a triangle, the “Amoeba”; at each step, the Amoeba “probes” the value of 3 points in its neighborhood, and moves based on how much better the new points are. As a result, the triangle is iteratively updated, and behaves a bit like an Amoeba moving on a surface.

Before going into the actual details of the algorithm, here is how my final result looks like. You can find the entire code here on GitHub, with some usage examples in the Sample.fsx script file. Let’s demo the code in action: in a script file, we load the Amoeba code, and use the same function the article does, the Rosenbrock function. We transform the function a bit, so that it takes a Point (an alias for an Array of floats, essentially a vector) as an input, and pass it to the solve function, with the domain where we want to search, in that case, [ –10.0; 10.0 ] for both x and y:

#load "Amoeba.fs"

open Amoeba
open Amoeba.Solver

let g (x:float) y =
100. * pown (y - x * x) 2 + pown (1. - x) 2

let testFunction (x:Point) =
g x.[0] x.[1]

solve Default [| (-10.,10.); (-10.,10.) |] testFunction 1000

Running this in the F# interactive window should produce the following:

val it : Solution = (0.0, [|1.0; 1.0|])
>

The algorithm properly identified that the minimum is 0, for a value of x = 1.0 and y = 1.0. Note that results may vary: this is a heuristic, which starts with a random initial amoeba, so each run could produce slightly different results, and might at times epically fail.

More...

6. May 2012 12:48

I am digging back into the Bumblebee code base, to clean it up before talking at the New England F# user group in Boston in June. As usual for me, it’s a humbling experience to face my own code, 6 months later, or, if you are an incorrigible optimist, it’s great to see that I am so much smarter today than a few months ago…

In any case, while toying with one of the samples, I noted that performance was degrading pretty steeply as the size of the problem was increasing. Most of the action revolved around producing random shuffles of a list, so I figured it would be interesting to look into it and see where I messed up how this could be improved upon.

Here is the original code, a quick-and-dirty implementation of the Fisher-Yates shuffle:

open System

let swap fst snd i =
if i = fst then snd else
if i = snd then fst else
i

let shuffle items (rng: Random) =
let rec shuffleTo items upTo =
match upTo with
| 0 -> items
| _ ->
let fst = rng.Next(upTo)
let shuffled = List.permute (swap fst (upTo - 1)) items
shuffleTo shuffled (upTo - 1)
let length = List.length items
shuffleTo items length

[<EntryPoint>]
let main argv =
let test = [1..10000]
let random = new Random()
let shuffled = shuffle test random
System.Console.WriteLine("done...")
0

Running the test case in fsi, using #time, produces the following:

Real: 00:00:42.735, CPU: 00:00:42.734, GC gen0: 2307, gen1: 6, gen2: 0

(Digression: #time is absolutely awesome – just typing #time;; in a fsi session will automatically display performance information, allowing to quickly tweak a function and fine-tune it “on the fly”. I wish I had known about it earlier.)

My initial assumption was that the problem revolved around performing multiple permutations of a List. However, I figured it would be interesting to take the opportunity and use the Performance Analysis tools provided in VS11 – and here is what I got:

Uh-oh. Looks like the shuffle is spending most of its time doing comparisons in the swap function – and what the hell is HashCompare.GenericEqualityIntrinsic doing in here? Something is off.

Looking into the swap function provides a hint:

F# has identified that the function could be made generic. It’s great, but in our case it comes with overhead, because we simply want to compare integers. Let’s mark the function as inline, to avoid that problem (we could also make the function non-generic, by marking one of the inputs as integer):

More...

7. April 2012 10:50

For no clear reason, I got interested in Convex Hull algorithms, and decided to see how it would look in F#. First, if you wonder what a Convex Hull is, imagine that you have a set of points in a plane – say, a board – and that you planted a thumbtack on each point. Now take an elastic band, stretch it, and wrap it around the thumbtacks. The elastic band will cling to the outermost tacks, leaving some tacks untouched. The convex hull is the set of tacks that are in contact with the elastic band; it is convex, because if you take any pair of points from the original set, the segment connecting them remains inside the hull.

The picture below illustrates the idea - the blue thumbtacks define the Convex Hull; all the yellow tacks are included within the elastic band, without touching it.

There are a few algorithms around to identify the Convex Hull of a set of points in 2 dimensions; I decided to go with Andrew’s monotone chain, because of its simplicity.

The insight of the algorithm is to observe that if you start from the leftmost tack, and follow the elastic downwards, the elastic turns only clockwise, until it reaches the rightmost tack. Similarly, starting from the right upwards, only clockwise turns happen, until the rightmost tack is reached. Given that the left- and right-most tacks belong to the convex hull, the algorithm constructs the upper and lower part of the hull by progressively constructing sequences that contain only clockwise turns.

More...

1. April 2012 09:47

Last week’s StackOverflow newsletter contained a fun problem I had never seen before: Bipartite Matching. Here is the problem:

There are N starting points (purple) and N target points (green) in 2D. I want an algorithm that connects starting points to target points by a line segment (brown) without any of these segments intersecting (red) and while minimizing the cumulative length of all segments.

Image from the original post on StackOverflow

I figured it would be fun to try out Bumblebee, my artificial bee colony library, on the problem. As the accepted answer points out, the constraint that no segment should intersect is redundant, and we only need to worry about minimizing the cumulative length, because reducing the length implies removing intersections.

As usual with Bumblebee, I’ll go first with the dumbest thing that could work. The solution involves matching points from two lists, so we’ll define a record type for Point and represent a Solution as two (ordered) lists of points, packed in a Tuple:

type Point = { X: float; Y: float }
let points = 100
let firstList = [ for i in 0 .. points -> { X = (float)i ; Y = float(i) } ]
let secondList =  [ for i in 0 .. points -> { X = (float)i ; Y = float(i) } ]

let root = firstList, secondList
We’ll start with a silly problem, where the 2 lists are identical: the trivial solution here is to match each point with itself, resulting in a zero-length, which will be convenient to see how well the algorithm is doing and how far it is from the optimum.

How can we Evaluate the quality of a solution? We need to pair up the points of each of the lists, compute the distance of each pair, and sum them up – fairly straightforward:

let distance pair =
((fst pair).X - (snd pair).X) ** 2.0 + ((fst pair).Y - (snd pair).Y) ** 2.0

let evaluate = fun (solution: Point list * Point list) ->
List.zip (fst solution) (snd solution)
|> List.sumBy (fun p -> – distance p)

More...