Question about input X's dimension

[qizzhao]

I have seen many posts that tutorial about training on 2dimensional's X or 1dimensional's X. And I am going to make a symbolic regression model, which input X is 3-dimenstion and input y is 2-dimension. Could SymbolicRegression.jl handle this? If yes, how to do it. (Please give me a short example please). Thank you.

[MilesCranmer]

You can do it with a custom loss function in the loss_function parameter: https://astroautomata.com/SymbolicRegression.jl/dev/api/ or https://astroautomata.com/PySR/examples/#9-custom-objectives. The idea is to flatten X into a 2D array, and then reshape it to 3D within the loss function.

To use eval_tree_array you have to pass a 2D array. Imagine batching a computation over all pixels.

const nx = 32
const ny = 20  # For example
function default_objective(tree, dataset::Dataset{T,L}, options)::L where {T,L}
    X = dataset.X
    (prediction, completion) = eval_tree_array(tree, X, options)
    if !completion
        return L(Inf)
    end
    # Here, X is a 2D array of shape [nx * ny, num_features]
    # y is a 1D array of shape [nx * ny]
    
    prediction_reshaped = reshape(prediction, (nx, ny))
    y_reshaped = reshape(dataset.y, (nx, ny))

    # Some computation on the 2D predictions
    
    return L(loss)
end

[MilesCranmer]

I’m not sure how they would get passed to it though. But you could just write nx=32 within the function definition if that’s what you mean?

[MilesCranmer]

Alternatively you could store this info in the extra field of the Dataset class, which permits arbitrary names tuples.

[qizzhao]

For my case, nx and ny are not constant number, but changing variable number. Each time training would use different value of nx and ny. Ideally, function is like function default_objective(tree, dataset::Dataset{T,L}, options, nx, ny) , but it seems not possible.

How could I save in the extra field of the Dataset before start of training? ( Do you mean set in Options or EquationSearch ?)

Thank you.

[MilesCranmer]

Why not just do a function closure?

function build_loss(nx, ny)
    return function default_objective(tree, dataset::Dataset{T,L}, options)::L where {T,L}
        X = dataset.X
        (prediction, completion) = eval_tree_array(tree, X, options)
        if !completion
            return L(Inf)
        end
        # Here, X is a 2D array of shape [nx * ny, num_features]
        # y is a 1D array of shape [nx * ny]
    
        prediction_reshaped = reshape(prediction, (nx, ny))
        y_reshaped = reshape(dataset.y, (nx, ny))

        # Some computation on the 2D predictions
    
        return L(loss)
    end
end

then you can pass it to the options like

Options(loss_function=build_loss(32, 20))

so whenever you need to change those numbers, you just change it in that call.

[qizzhao]

Thank you.