Enhance class SoftmaxRegression

[stereomatchingkiss]

1 : add function Serialize to class SoftmaxRegression
2 : add function Train to class SoftmaxRegression
3 : fix bug--did not initialize fitIntercept

The test case "SoftmaxRegressionFunctionEvaluate" always fail on my machine

[rcurtin]

For consistency, do you mind if I change this to () initialization vs. {} initialization?

[stereomatchingkiss]

No problem(I use {} to avoid narrow conversion)

[rcurtin]

Looks good to me; I made a few comments. Let me know what you think. I'll add a test for SoftmaxRegression::Serialize() once I merge. Do you mind if I add your name/email to the list of authors? If that's okay, which name/email should I use?

[stereomatchingkiss]

Do you mind if I add your name/email to the list of authors?
My pleasure

name : Tham Ngap Wei
email : thamngapwei@gmail.com

Thanks for adding the unit test.

[rcurtin]

Okay, I made some simple changes in cbeb3ea..f85a9b2; let me know if you see anything wrong. If you are interested, here are some future possibilities for refactoring/improving the SoftmaxRegression code further:

• Storing inputSize and numClasses is, I think, redundant, and I think you can just get these from the parameters matrix. So the constructors can be cleaned up a little (although numClasses will still be necessary in the constructor).
• The type of label would be better as arma::Row<size_t> not arma::vec.
• The InitializeWeights() function could be made static or moved elsewhere so that creating a SoftmaxRegressionFunction to initialize the weights in the non-training constructor isn't necessary.
• A softmax_regression_main.cpp program could be written that could allow users to use softmax regression from the command line.

Don't feel obligated to do any of these -- I'm just tossing them out there as ideas. :)

Thanks for your contribution! :)

[stereomatchingkiss]

Your suggestions are good, about

Storing inputSize and numClasses is, I think, redundant

Some(or many) classes of mlpack always ask for the inputSize(feature size) from the users, I think they could be removed too, because many algorithms should be able to find out the inputSize from the training data

Thanks for your contribution! :)
You are welcome:), thanks for the helps from community too