machine Issue Tracker Rss Feed

Commented Issue: Implement SEARN functionality [9629]

SethJuarez — Wed, 10 Nov 2010 22:31:48 GMT

You said: "Although, my adviser did write a paper describing how to learn structure from a set of independent classifiers with a system he developed called SEARN. Maybe I should implement it". After reading the whole paper, I agree with that, and think it would be very benificial to many, please implement SEARN!
Comments: Cool. I will add it to my list.

Created Issue: Implement SEARN functionality [9629]

rickrat — Wed, 13 Oct 2010 21:17:08 GMT

Commented Issue: PerceptronModel.Generate fails in an unexpected way when given an empty collection. [9235]

spivey — Fri, 20 Aug 2010 19:49:29 GMT

I expected the following instructions to either fail silently, or throw an argument exception.
var model = new PerceptronModel<Student>();
var predictor = model.Generate(new List<Student>()); // notice empty set

Instead, an index out of range exception is thrown from inside the Matrix class. Consider adding guard clause or Code Contracts. For example:

using System.Diagnostics.Contracts;
using System.Linq;
...
public IPredict<T> Generate(IEnumerable<T> examples)
{
Contract.Requires(examples != null);
Contract.Requires(examples.Count() > 0);
Comments: ** Comment from web user: spivey **

Contracts have a number of advantages that make them worth considering. The most notable advantage is that they enable the use of Microsoft's static checker, which identifies any potential violations at compile time. Using the static checker, callers of the Generate function would receive a warning if it was possible that they could pass a null or empty set.

Advantages
* Contracts can be applied to the IModel interface, so exceptional behavior would be standard across all implementations of the interface.
* Contracts provide explicit documentation for users of the library.
* Contracts can be stated once, and then verified at runtime, verified by the static checker, read by humans, and parsed by documentation generators like SandCastle.

Contracts are cumbersome to add, so the benefits come at a cost. The QuickGraph (http://quickgraph.codeplex.com) library has been using them for some time. I wonder if they've benefited from them.

Commented Issue: Perceptron and LinearRegression should directly usable [9236]

SethJuarez — Fri, 20 Aug 2010 18:50:07 GMT

The IModel and IPredict interfaces provide useful abstractions, however they hide interesting information on the supervised classes. For example, the PerceptronPredictor.W and PerceptronPredictor.Features properties help explain the predictions.

Recommendations

1. Expose this information by making the properties more visibile

from
public class PerceptronPredictor<T> : IPredict<T>, IXmlSerializable
{
internal Vector W { get; set; }
internal Property[] Features { get; set; }
internal Property Target { get; set; }
internal double B { get; set; }

to
public class PerceptronPredictor<T> : IPredict<T>, IXmlSerializable
{
public Vector W { get; internal set; }
public Property[] Features { get; internal set; }
public Property Target { get; internal set; }
public double B { get; internal set; }

2. Making the PerceptronModel.Generate method static to eliminate the need to instantiate the PerceptronModel.

The current usage is
var model = new PerceptronModel<Student>();
var predictor = model.Generate(students);

If made static, this would be reduced to
var predictor = PerceptronModel<Student>.Generate(students);

3. The PerceptronModel.Generate method should return a PerceptronPredictor, in addition to the IPredict. This will allow it to be used directly, without casting.

Existing usage
PerceptronPredictor<Student> predictor = (PerceptronPredictor<Student>)model.Generate(data);

Proposed usage
PerceptronPredictor<Student> predictor = PerceptronModel<Student>.Generate(data);

The code to make this work is:
IPredict<T> IModel<T>.Generate(IEnumerable<T> examples)
{
return Generate(examples);
}

public PerceptronPredictor<T> Generate(IEnumerable<T> examples)
Comments: ** Comment from web user: SethJuarez **

Sounds good. I will add them sometime this week.

Commented Issue: PerceptronModel.Generate fails in an unexpected way when given an empty collection. [9235]

SethJuarez — Fri, 20 Aug 2010 18:47:31 GMT

Oops! Rookie mistake. I will fix that. It also looks like you are using code contracts. Would it be worth the time to look into those?

Created Issue: Perceptron and LinearRegression should directly usable [9236]

spivey — Fri, 20 Aug 2010 17:10:09 GMT

The IModel and IPredict interfaces provide useful abstractions, however they hide interesting information on the supervised classes. For example, the PerceptronPredictor.W and PerceptronPredictor.Features properties help explain the predictions.

Recommendations

1. Expose this information by making the properties more visibile

from
<code>
public class PerceptronPredictor<T> : IPredict<T>, IXmlSerializable
{
internal Vector W { get; set; }
internal Property[] Features { get; set; }
internal Property Target { get; set; }
internal double B { get; set; }
</code>
to
<pre>
public class PerceptronPredictor<T> : IPredict<T>, IXmlSerializable
{
public Vector W { get; internal set; }
public Property[] Features { get; internal set; }
public Property Target { get; internal set; }
public double B { get; internal set; }
</pre>

2. Making the PerceptronModel.Generate method static to eliminate the need to instantiate the PerceptronModel.

The current usage is
var model = new PerceptronModel<Student>();
var predictor = model.Generate(students);

If made static, this would be reduced to
var predictor = PerceptronModel<Student>.Generate(students);

3. The PerceptronModel.Generate method should return a PerceptronPredictor, in addition to the IPredict. This will allow it to be used directly, without casting.

Existing usage
PerceptronPredictor<Student> predictor = (PerceptronPredictor<Student>)model.Generate(data);

Proposed usage
PerceptronPredictor<Student> predictor = PerceptronModel<Student>.Generate(data);

The code to make this work is:
IPredict<T> IModel<T>.Generate(IEnumerable<T> examples)
{
return Generate(examples);
}

public PerceptronPredictor<T> Generate(IEnumerable<T> examples)

Created Issue: PerceptronModel.Generate fails in an unexpected way when given an empty collection. [9235]

spivey — Fri, 20 Aug 2010 17:06:48 GMT