[FEATURE] Multilayer Perceptron

[imdeep2905]

I have seen that you guys have MLP in TODO here.

I already have implemented NN in C++ (repo link). I am planning to make a pull request to add this. But as you might have seen I have created three header files. I am also going to modify code to suit the pull request. Can you please discuss your ideas about what should I do with header files and also some ideas about refactoring my code.

Thanks.

[kvedala]

Thats wonderful, thank you 👍

1. Ideally, the repo had been using at most of two files - one CPP and one H, if needed.
2. The matrix.hpp sounds quite generic and useful. Please review QR decomposition and LU decomposition programs in the repo that use similar tools but only that are needed. There is redundancy but it keeps the programs whole and easy to read and study with the documentation added and organized using namespaces.

[imdeep2905]

I will add proper documentation and namespaces. One more thing should I keep header files in the same folder (i.e. machine_learning/).

[kvedala]

I will add proper documentation and namespaces. One more thing should I keep header files in the same folder (i.e. machine_learning/).

yes, in the same folder please and with the same names, except for the type suffix - .h and .cpp.
Also, you can reuse the header files and the compiler will link them correctly. Thus, you can focus on the algorithm at hand - the ML while not reinventing the wheel :)
Note that std::valarray provides much of the vector operations that the QR and LU decomposition algorithms utilize.

[imdeep2905]

Is it okay if I provide only MSE as loss function and SGD as optimizer (along with various activation functions) ?

[kvedala]

Is it okay if I provide only MSE as loss function and SGD as optimizer (along with various activation functions) ?

Absolutely!

[imdeep2905]

Should I provide facility to save and load model ?
I mean it is already there in my code, but I was confused about that feature because it includes file handling (i.e. fstream).

[kvedala]

Should I provide facility to save and load model ?
I mean it is already there in my code, but I was confused about that feature because it includes file handling (i.e. fstream).

You can use any of the standard C++11 features from the STL library.

[imdeep2905]

For testing Can I upload dataset (for e.x. iris.csv) in the same folder ?

[kvedala]

I think we can have a small enough CSV file for testing in ML.
Is it https://gist.github.com/curran/a08a1080b88344b0c8a7#file-iris-csv?
Or you can generate simple test data as in other ML programs in the repo and use that for testing.

[imdeep2905]

Is it https://gist.github.com/curran/a08a1080b88344b0c8a7#file-iris-csv?

Yes.
It is only around 4KB in size.

[kvedala]

Is it https://gist.github.com/curran/a08a1080b88344b0c8a7#file-iris-csv?

Yes.
It is only around 4KB in size.

Ok, sounds fine.