Compute derivative from fitted model

[llamm-de]

Dear @MilesCranmer & community,
first of all, thanks for developing the wonderful SR package. It is a pleasure working with it.

Once I have trained my model using the MLJ interface, I would not only like to use predict(mach, X) but also compute the derivative of my model expression with respect to X. I know that there exists the eval_diff_tree_array() function, but I am not sure on how to call this function outside e.g. a custom loss function. I would somehow need to get access to the tree of the best fit from my model to pass it to eval_diff_tree_array(). Is this possible? Or does anyone have a better idea on how to tackle the issue?

Thanks a lot in advance.

Bests
Lukas

[MilesCranmer]

Hi @llamm-de,

Thanks for the message. This is a bit tricky as MLJ does not have an interface for this, so you will have to use the low-level SymbolicRegression.jl calls directly. The easiest way to do this right now would be as follows:

using MLJ
using SymbolicRegression as SR

...
model = ...
mach = ...
...

# Get the tree:
r = report(mach)
tree = r.equations[r.best_idx]

# Get the transposed X:
X_t, _, _ = SR.MLJInterfaceModule.get_matrix_and_info(X, model.dimensions_type)

# Get options from mach:
options = mach.fitresult.options

# Use:
eval_grad_tree_array(tree, X_t, options; variable=true/false)  # or eval_diff_tree_array

But it would definitely nice if there was a function to pass this through automatically.

Cheers,
Miles

[llamm-de]

Hey @MilesCranmer,

thanks for the quick answer. It works like a charm. 😊 👍

Bests Lukas