BoundsError in custom eval_tree_array

[DenisSvirin]

What happened?

I tried to run '_eval_tree_array' with array that has different dimensions then cX in 'eval_tree_array'

And if i understand correctly after test runs l've got an error ''

Version

0.18.0

Operating System

macOS

Interface

Jupyter Notebook

Relevant log output

BoundsError: attempt to access 1×11 Matrix{Float32} at index [4, 1:11]

Extra Info

my version of 'eval_tree_array' :

function eval_tree_array(
    tree::Node{T}, cX::AbstractMatrix{T}, operators::OperatorEnum; turbo::Bool=false
)::Tuple{AbstractVector{T},Bool} where {T<:Number}
    if turbo
        @assert T in (Float32, Float64)
    end
    r = T.( [6.332061275761631 5.051237472144821 2.7 4.2690748412273125 3.3068111527572905 5.4 4.676537180435969 6.037383539249433 5.727564927611035 3.8183$
    result, finished = _eval_tree_array(
        tree, r, operators, (turbo ? Val(true) : Val(false))
    )

    @return_on_false finished result
    @return_on_nonfinite_array result  
    return result, finished
end

[MilesCranmer]

Could you share your full code, so I can reproduce it?

[DenisSvirin]

I've changed 'eval_tree_array' in the backend:

function eval_tree_array(
    tree::Node{T}, cX::AbstractMatrix{T}, operators::OperatorEnum; turbo::Bool=false
)::Tuple{AbstractVector{T},Bool} where {T<:Number}
    if turbo
        @assert T in (Float32, Float64)
    end
    r =T.( [6.332061275761631 5.051237472144821 2.7 4.2690748412273125 3.3068111527572905 5.4 4.676537180435969 6.037383539249433 5.727564927611035 3.818376618407357 1.9091883092036785])
    result, finished = _eval_tree_array(
        tree, r, operators, (turbo ? Val(true) : Val(false))
    )
@return_on_false finished result
    @return_on_nonfinite_array result  
    return result, finished
end

and to start i use python:

import numpy as np
from pysr import PySRRegressor
r = np.array([ 2.7       ,  3.        ,  3.2       ,  3.4       ,  3.48621994,
        3.52253473,  3.55884952,  3.59516431,  3.6314791 ,  3.66779389,
        3.70410868,  3.74042347,  3.77673827,  3.81305306,  4.        ,
        4.5       ,  5.        ,  5.5       ,  5.73684211,  5.97368421,
        6.        ,  6.21052632,  6.44736842,  6.68421053,  6.92105263,
        7.15789474,  7.39473684,  7.63157895,  7.86842105,  8.        ,
        8.10526316,  8.34210526,  8.57894737,  8.81578947,  9.05263158,
        9.28947368,  9.52631579,  9.76315789, 10.        , 35.        ]) 
e = np.array([ 20.        ,  -2.3341443 , -10.378765  , -13.874788  ,
       -14.534859  , -14.705036  , -14.821358  , -14.8896    ,
       -14.934813  , -14.90294   , -14.857485  , -14.782236  ,
       -14.681216  , -14.557151  , -13.653582  , -10.439137  ,
        -7.4652775 ,  -5.1739723 ,  -4.26172661,  -3.58361976,
        -3.5250227 ,  -3.01341022,  -2.53392988,  -2.1307423 ,
        -1.79170813,  -1.50661955,  -1.26689299,  -1.06531065,
        -0.89580319,  -0.78888653,  -0.753267  ,  -0.63341053,
        -0.53262508,  -0.44787616,  -0.37661211,  -0.31668728,
        -0.26629742,  -0.22392536,  -0.18829536,   0.        ])
model = PySRRegressor(
    model_selection="best",
    niterations=40,
    binary_operators=["+", "*", "-", "^"],
    loss="loss(x, y) = (x - y)^2")
model.fit(r.reshape(-1, 1), e)

[MilesCranmer]

Okay I think I might understand the issue. Does the error come during precompilation for you? Or during the search itself?

Can you try changing the loss function, rather than the evaluation code?

See https://astroautomata.com/PySR/examples/#9-custom-objectives ^ this is how you should do custom objectives, rather than changing the evaluation function itself.

[DenisSvirin]

precompilation works fine, it occurs during search.

Oh, then i'll try it this way.

[DenisSvirin]

It seems, that everything works fine this way

[MilesCranmer]

So the bug is fixed with this change?

[DenisSvirin]

yes, if i do it with full_objective

[MilesCranmer]

Great!

It's probably because there are other pieces of code that rely on eval_tree_array. It is a generic function for evaluating an expression and shouldn't be overloaded. eval_loss is for these purposes (which is what full_objective overrides).