cookbook - CARTree - classification tree

[OXPHOS]

• CART takes a SGVector in parameter list. What I had in the code is the most reasonable way I found to initialize a SGVector in meta - but it didn't work. I got error:

/Users/zora/Github/shogun/build/examples/meta/cpp/classifier/randomforest.cpp:32:16: error: 
      no template named 'CSGVector'; did you mean 'SGVector'?
auto ft = some<CSGVector<bool>>(2);
               ^~~~~~~~~
               SGVector
/Users/zora/Github/shogun/src/shogun/lib/SGString.h:22:25: note: 'SGVector'
      declared here
template<class T> class SGVector;

• Need more work on the introduction of cart and I didn't check the style yet as make cookbook fails because of the changes I made for the first point.

[karlnapf]

It should be easier to use a bool vector...

[karlnapf]

@OXPHOS currently the meta language does not support

• instantiating SG* types
• accessing elements within

I think we should add both. @sorig do you have an idea how much work that is? We would need to adjust the include path stuff (similar to what @OXPHOS did in this patch), and then wrapping the [] and () operators might be nice as well. What do you think?

@OXPHOS Let's see what @sorig says before we continue here

[OXPHOS]

I rebased based on #3285 but got the same error here and at local.

I removed parser.out, parsetab.py, parsetab.pyc, lextab.py and lextab.pyc in examples/meta/generator.

I was using:

#![set_attribute_types]
BoolVector ft(2)
ft[0] = False
ft[1] = False
#![set_attribute_types]

[karlnapf]

1. Remove the build dir
2. Clean up the examples/meta/generator/ parser output

It works on my machine. Ask travis, it might confirm

[sorig]

Let me know if you need help with this.

Java error is caused by java.json#L14 not using the correct type. You need to figure out how BoolVectors are mapped to Java and use the appropriate type here.

C# is also an error with mapped types. Take the generated C# file, figure out how to make it compile, and update the csharp.json file.

cartree.cs(26,22): error CS1502: The best overloaded method match for `CARTree.CARTree(BoolVector, EProblemType, int, bool)' has some invalid arguments
/opt/shogun/build/src/interfaces/csharp_modular/modshogun.dll (Location of the symbol related to previous error)
cartree.cs(26,34): error CS1503: Argument `#1' cannot convert `bool[]' expression to type `BoolVector'

The errors in the other languages are probably similar, but need investigation.

[OXPHOS]

@sorig Hey I have got some results, but not solutions...
So I need to use a new BoolVector for cartree, but I looked into the errors in Java and found that there's no BoolVector type in jblas. I think this is why Heiko commented out this.
If we just use 0 and 1 for false and true, apparently csharp doesn't like it. And use IntVector instead of BoolVector won't work either. So I am not sure how to deal with the problem.

There's another problem in Java, even if the above one is solved, also appeared in #3303:

//![set_attribute_types]
DoubleMatrix ft = new DoubleMatrix(2);
ft.put(0, 0);
ft.put(1, 0);
//![set_attribute_types]

If I create a vector of length 2 in meta language (and translated in cartree.java as shown in the above code), and pass the vector to the constructor of CCarTree:

//![create_instance]
CARTree classifier = new CARTree(ft, PT_MULTICLASS, 5, true);
//![create_instance]

and output the length of the vector in CCarTree constructor:

CCARTree::CCARTree(SGVector<bool> attribute_types, EProblemType prob_type, int32_t num_folds, bool cv_prune): CTreeMachine<CARTreeNodeData>()
{
    std::cerr << "vector length: " << attribute_types.vlen << std::endl;
}

I will get: vector length: 1, when ft has size of 2.
I have tried initializing a matrix of size(1, 2) or size(2, 1), and I got the same error.
So in summary whatever size vector/matrix I pass to the constructor of CCarTree as param, it will end up with size of 1. I have no idea why this is happening. I am not sure whether you could help look into this? Thanks!

[OXPHOS]

@karlnapf any thoughts? : )

[karlnapf]

@OXPHOS I can look at this, shall I just copy the meta example in this PR? Does it reproduce the errors.
With Java, there is a problem with strong types and implicit numerical downcasting ... Ill look intot his

[OXPHOS]

@karlnapf Maybe you could reproduce the error by using the code in #3303, where the error is simpler: When passing a vector

//![set_attribute_types]
DoubleMatrix ft = new DoubleMatrix(2);
ft.put(0, 0);
ft.put(1, 0);
//![set_attribute_types]

to the constructor:

CCHAIDTree(int32_t dependent_vartype, SGVector<int32_t> feature_types, int32_t num_breakpoints=0);

The vector will have size of 1.

[karlnapf]

The "other" problem should be solved in #3451
Looking into the other tomorrow or so.

[karlnapf]

There is another problem with csharp and boolean vectors --- we don't have a typemap, which means we cannot pass boolean vectors as input in the csharp interface, see #3452

This, sadly, means that for now, we have to change this example to not using boolean vectors that are hand-constructed as parameters. Instead, we can create an instance of the (mapped) BoolVector as an object and then use a method (say zero) to populate it (hacky)

[karlnapf]

Can you rebase? Should work now

[karlnapf]

I guess this needs an updated data version and a squash. It is all ready otherwise

[karlnapf]

Can you please run the tests locally (at least the cpp version) before putting things into the PR. This would have been detected locally.

Is done with
ctest -R integration_meta_cpp-multiclass_classifier-cartree -V

[karlnapf]

See the error is coming from a missing shogun-data update

The generated test passed:
https://travis-ci.org/shogun-toolbox/shogun/jobs/175608246#L3717

But checking against reference data failed:
https://travis-ci.org/shogun-toolbox/shogun/jobs/175608246#L4322

[karlnapf]

And it even tells you the line where it differs:
https://travis-ci.org/shogun-toolbox/shogun/jobs/175608246#L4427

[OXPHOS]

@karlnapf Thanks..! I saw only some of the tests failed so I thought it should be some specific errors without checking closely.. I updated the test dataset and it worked for me at local now.

[karlnapf]

So let's get it in then :)

[OXPHOS]

Ruby and CSharp don't like 0 as False value, and it might also be true for Octave. But it is weird that in R tests the integration_meta_cpp-multiclass_classifier-cartree failed as the generated output didn't match ref. I'll check whether there is chance to generate different results from the algorithm.

[karlnapf]

Ah sigh, I thought I had solved that. But we have a mechanic to fix this...checking ...

[karlnapf]

@OXPHOS rebase against develop, then it should work (I tried locally for all interfaces that failed in the last build here)

[karlnapf]

@OXPHOS You should always try these things locally before you push. At least in one modular language.

So here is the listing I used (no fixed random seed)
https://gist.github.com/karlnapf/6cf4186dc77861681ceba938f397f2c0

Note that meta language syntax is True False

Also, check out the docs for git commit --amend, which allows you to update your commit, so that the PR doesnt have like 5 commits, but just one

[karlnapf]

I dont understand the failure in octave. Checking again. The rest looks fine

[karlnapf]

Ok so works locally for oactave, I restarted the build
Just need to squash the commits, then we should be able to merge (fingers crossed travis is fine)

[karlnapf]

This seems to be specific to the travis setup, I don't have this problem locally ....
EDIT: Managed to reproduce now.

[karlnapf]

@OXPHOS I solved the problem with octave. Man, this is a beast to get merged.
But I think you can squash the next time before you push ( after rebasing against develop once #3559 is merged, EDIT: it is merged now), since all the rest works

[OXPHOS]

@karlnapf It finally works! Thanks for the advice and sorry that I skipped the local test because of laziness.

[karlnapf]

Nice one, this was a big effort to get these Boolean vectors working. But it really helped in figuring out corner cases of the whole system. Thanks for the patience :)

[karlnapf]

Some minor style updates would be good :)

[karlnapf]

Decition tree learning is based on trees as predictive models. (Remove the second sentence)

[karlnapf]

There are two types of decision trees:

[karlnapf]

Or even better: There are decision trees for both classification (integer-valued) and regression (real-valued).

[karlnapf]

The classification and regression tree (CART) algorithm is an umbrella method ....

[karlnapf]

In this example, we show how to apply CART to multi-class datasets.

Remove the rest and fix typos

[karlnapf]

These words you can also use above when you talk about regression/classification values

[karlnapf]

A reference to a wiki article would be good here