Export weights of a Random Forest

konstantint / SKompiler

A tool for compiling trained SKLearn models into other representations (such as SQL, Sympy or Excel formulas)

MIT License

171 stars 10 forks source link

Export weights of a Random Forest #10

Closed PARODBE closed 2 years ago

PARODBE commented 2 years ago

With this library, would be possible extract the weights of a Random Fores as for example in suppor vector machine or PLS or Neural Network? In this case, could you provide me some code, for after load the weights and training your model.

Thanks! Pablo

konstantint commented 2 years ago

I am not sure I completely understand the question. A couple of possibly related remarks:

A random forest is a forest of trees, and it does not have a notion of "weights" equivalent to purely algebraic models.
You can extract the internals of a random forest (e.g. the per-tree weight or the individual tree structures), however I don't think SKompiler is your best tool here. It might be easier and more natural to "manually" examine the internals of the respective data structure. You might find examining the source code of SKompiler useful here (as it also based on the analysis of the respective internal structures). Start here.
Using SKompiler to convert the random forest expression to raw Python code (skompile(m.predict).to('python/code')) can perhaps be a useful teaching aid for small forests of not-too-deep trees. Here is how this is used for decision trees, for example.

PARODBE commented 2 years ago

Thanks for so rapid answer!. When I do this:

Being my dataset, the iris_dataset. I obtain the next error:

However the shape of my inputs is:

Could you help me please?

Thanks!

konstantint commented 2 years ago

SKompiler expects the inputs argument to provide a list of symbolic arguments that will be referred to in the expression being generated. E.g.

from skompiler.dsl import ident
inputs = [ident('a'), ident('b'), ident('c'), ident('d')]

could probably work out in your case.

I don't see however how your current actions will help you in your quest. I suggested you to look at SKompiler code in the sense that it might help you better understand the internal data structures of SKLearn (e.g. you can see that the loop in the function goes over model.estimators_ which can be a useful hint here). Just running it is probably not going to help.