views:

220

answers:

3

I'm trying to create a memoization interface for functions with arbitrary number of arguments, but I'm failing miserably I feel like my solution is not very flexible. I tried to define an interface for a function which gets memoized automatically upon execution and each function will have to implement this interface. Here is an example with a two parameter Exponential Moving Average function:

class EMAFunction:IFunction
{
    Dictionary<List<object>, List<object>> map;

    class EMAComparer : IEqualityComparer<List<object>>
    {
        private int _multiplier = 97;

        public bool Equals(List<object> a, List<object> b)
        {
            List<object> aVals = (List<object>)a[0];
            int aPeriod = (int)a[1];

            List<object> bVals = (List<object>)b[0];
            int bPeriod = (int)b[1];

            return (aVals.Count == bVals.Count) && (aPeriod == bPeriod);
        }

        public int GetHashCode(List<object> obj)
        {
            // Don't compute hash code on null object.
            if (obj == null)
            {
                return 0;
            }

            List<object> vals = (List<object>) obj[0];
            int period = (int) obj[1];

            return (_multiplier * period.GetHashCode()) + vals.Count;

        }
    }

    public EMAFunction()
    {
        NumParams = 2;
        Name = "EMA";
        map = new Dictionary<List<object>, List<object>>(new EMAComparer());
    }
    #region IFunction Members

    public int NumParams
    {
        get;
        set;
    }

    public string Name
    {
        get;
        set;
    }

    public object Execute(List<object> parameters)
    {
        if (parameters.Count != NumParams)
            throw new ArgumentException("The num params doesn't match!");

        if (!map.ContainsKey(parameters))
        {
            //map.Add(parameters,
            List<double> values = new List<double>();
            List<object> asObj = (List<object>)parameters[0];
            foreach (object val in asObj)
            {
                values.Add((double)val);
            }
            int period = (int)parameters[1];

            asObj.Clear();
            List<double> ema = TechFunctions.ExponentialMovingAverage(values, period);
            foreach (double val in ema)
            {
                asObj.Add(val);
            }
            map.Add(parameters, asObj);
        }
        return map[parameters];
    }

    public void ClearMap()
    {
        map.Clear();
    }

    #endregion
}

Here are my tests of the function:

private void MemoizeTest()
{
    DataSet dataSet = DataLoader.LoadData(DataLoader.DataSource.FROM_WEB, 1024);
    List<String> labels = dataSet.DataLabels;

    Stopwatch sw = new Stopwatch();
    IFunction emaFunc = new EMAFunction();
    List<object> parameters = new List<object>();
    int numRuns = 1000;
    long sumTicks = 0;
    parameters.Add(dataSet.GetValues("open"));
    parameters.Add(12);

    // First call

    for(int i = 0; i < numRuns; ++i)
    {
        emaFunc.ClearMap();// remove any memoization mappings
        sw.Start();
        emaFunc.Execute(parameters);
        sw.Stop();
        sumTicks += sw.ElapsedTicks;
        sw.Reset();
    }
    Console.WriteLine("Average ticks not-memoized " + (sumTicks/numRuns));


    sumTicks = 0;
    // Repeat call
    for (int i = 0; i < numRuns; ++i)
    {
        sw.Start();
        emaFunc.Execute(parameters);
        sw.Stop();
        sumTicks += sw.ElapsedTicks;
        sw.Reset();
    }
    Console.WriteLine("Average ticks memoized " + (sumTicks/numRuns));
}

Update:
Thanks for pointing out my n00bish error... I always forget to call Reset on the stopwatch!

I've seen another approach to memoization as well... it doesn't offer n-argument memoization, but my approach with the Interface is not much more advantageous since I have to write a class for each function. Is there a reasonable way that I can merge these ideas into something more robust? I want to make it easier to memoize a function without making the user write a class for each function that they intend to use.

+1  A: 

StopWatch.Stop does not reset the stopwatch so you are accumulating time on each start/stop.

For example

  Stopwatch sw = new Stopwatch();

  sw.Start();
  System.Threading.Thread.Sleep(100);
  sw.Stop();
  Debug.WriteLine(sw.ElapsedTicks);

  sw.Start();
  System.Threading.Thread.Sleep(100);
  sw.Stop();
  Debug.WriteLine(sw.ElapsedTicks);

Gives the following results

228221
454626

You can use StopWatch.Restart (Framework 4.0) to restart the stopwatch each time, or if not Framework 4.0, you can use StopWatch.Reset to reset the stopwatch.

Chris Taylor
OK, I'm ashamed now! I **always** forget to reset the stopwatch!
Lirik
+3  A: 

First, you need to call sw.Reset() between your tests. Otherwise your results for the second test will be in addition to the time from the first.

Second, you probably shouldn't use vals.GetHashCode() in your GetHashCode() override on the comparer, as this will lead you to getting different hash codes for objects that would evaluate to true for your Equals override. For now, I would worry about making sure that equivalent objects always get the same hash code rather than trying to get an even distribution of the codes. If the hash codes don't match, Equals will never be called, so you'll end up processing the same parameters multiple times.

Adam Robinson
OK, I fixed the GetHashCode function and I called Reset... things look much better now. I also saw another good way to get the memoization done, any suggestions on how it can be generalized: http://stackoverflow.com/questions/633508/two-argument-memoization
Lirik
+8  A: 

How about this? First write a one-argument memoizer:

static Func<A, R> Memoize<A, R>(this Func<A, R> f)
{
    var d = new Dictionary<A, R>();
    return a=> 
    {
        R r;
        if (!d.TryGetValue(a, out r))
        {
            r = f(a);
            d.Add(a, r);
        }
        return r;
    };
}  

Straightforward. Now write a function tuplifier:

static Func<Tuple<A, B>, R> Tuplify<A, B, R>(this Func<A, B, R> f)
{
    return t => f(t.Item1, t.Item2);
}

And a detuplifier:

static Func<A, B, R> Detuplify<A, B, R>(this Func<Tuple<A, B>, R> f)
{
    return (a, b) => f(Tuple.Create(a, b));
}

and now a two-argument memoizer is easy:

static Func<A, B, R> Memoize(this Func<A, B, R> f)
{
    return f.Tuplify().Memoize().Untuplify();
}

To write a three-argument memoizer just keep following this pattern: make a 3-tuplifier, a 3-untuplifier, and a 3-memoizer.

Of course, if you don't need them, there's no need to make the tuplifiers nominal methods:

static Func<A, B, R> Memoize(this Func<A, B, R> f)
{
    Func<Tuple<A, B>, R>> tuplified = t=>f(t.Item1, t.Item2);
    Func<Tuple<A, B>, R>> memoized = tuplified.Memoize();
    return (a, b)=>memoized(Tuple.Create(a, b));
}

UPDATE: You ask what to do if there is no tuple type. You could write your own; it's not hard. Or you could use anonymous types:

static Func<T, R> CastByExample<T, R>(Func<T, R> f, T t) { return f; }

static Func<A, B, R> Memoize<A, B, R>(this Func<A, B, R> f)
{
    var example = new { A=default(A), B=default(B) };
    var tuplified = CastByExample(t=>f(t.A, t.B), example);
    var memoized = tuplified.Memoize();
    return (a, b)=>memoized(new {A=a, B=b});
}

Slick, eh?

Eric Lippert
@Eric, that's a great idea... unfortunately I'm using .NET 3.5 and I won't be able to switch to 4.0 for some time. Is there a .NET 3.5 compatible version of the above?
Lirik
@Lirik: writing your own tuple type is not very hard. The trick is to make sure that you get the equality and hash code computation correct. Or, there are really tricky ways to make this work with anonymous types, which are effectively the same as tuples.
Eric Lippert
I love the anon type as an arbitrary tuple, you might want to point out that when you hit the wall in regards to system defined [Func<...> ](http://msdn.microsoft.com/en-us/library/system.aspx) you do need to start implementing your own. Gotta say if you're doing this it's perhaps time to consider f# as a possible candidate language...
ShuggyCoUk
oh and a link to [CastByExample](http://blogs.msdn.com/alexj/archive/2007/11/22/t-castbyexample-t-object-o-t-example.aspx)
ShuggyCoUk
@Eric and @ShuggyCoUK, that's amazing! I can't believe it works so well! If I could up-vote this answer more than once, **I would**! LOL
Lirik