Custom loss function: number of calls of the custom loss function

[lyhybrid]

Hi!
I am currently trying to use the integration value as my loss function. My ODE function is written as: dc_pred_dt, _ = eval_tree_array(tree, X, options) , and the loss function is written as: sum((C_pred .- dataset.y) .^ 2) / length(dataset.y).
However, when I set niteration = 1 and population = 1 to test the code, it seems to call the loss function too many times( around 800). It should be 1(niteration) * 1(population) = 1 (time) if I get this right? Wondering is there any potential problem here, or why is it evaluating the loss function so many times?

Thank you so much for your help!
Best regards

objective = """
using DifferentialEquations
let counter = 0

function custom_loss(tree, dataset::Dataset{T,L}, options)::L where {T,L}
    counter += 1
    println("counter:", counter)
    
    function ode_func(dc, c, t)
        X = Matrix{Float32}(undef, 1, 1)
        X[1,1] = Float32(c[1])
        
        dc_pred_dt, _ = eval_tree_array(tree, X, options)  
        dc[1]= dc_pred_dt[1]
    end

    c1_0 = [dataset.y[1]]
    tspan = (0.0, 300.0)
    prob = DifferentialEquations.ODEProblem(ode_func, c1_0, tspan)
    
    try
        sol = solve(prob,(Rosenbrock23())) 
        C_pred = [u[1] for u in sol.u]
        return sum((C_pred .- dataset.y) .^ 2) / length(dataset.y)
    catch e
        #@warn "ODE_fail: $e"
        return Inf
    end
end
end
"""
model = PySRRegressor(
    niterations=1,
    populations=1,
    binary_operators=["+", "*", "/", "^"],
    unary_operators=[],
    loss_function =objective,
    maxsize=20,
    progress=True,
    batching= False,
    constraints={'^': (-1, 1)}
)

[MilesCranmer]

The # of evaluations is much more complicated (e.g., a BFGS constant optimization step can do 100s of evaluations). There is also the ncycles_per_iteration and population_size. Also, often the evaluations are cached if the algorithm knows that the loss should not change (like a simplification step).

Easiest is to measure it with max_evals.