Strange error when I try to define my own operator

[xuyuon]

I was trying to add to the model an unary operator to return a value from a Gaussian distribution, so I coded something like this:

model=PySRRegressor(
...
unary_operators=[
        "exp",
        "gau(x) = exp(-0.5 * x^2.0)",
    ],
...

but this would raise a runtime error:
TypeError: <PyCall.jlwrap (in a Julia function called from Python) JULIA: TypeError: in typeassert, expected Float32, got a value of type Float64

Therefore, I changed my code to:

model=PySRRegressor(
...
unary_operators=[
        "exp",
        "gau(x::Float64) = exp(-0.5 * x^2.0)",
    ],
...

This produced another error:
RuntimeError: <PyCall.jlwrap (in a Julia function called from Python) JULIA: AssertionError: Your configuration is invalid - one of your operators (gau) is not well-defined over the real line. You can get around this by returning NaN for invalid inputs.

This error makes no sense because the function I wrote should be able to accept any real numbers. Anyway, I altered my code again:

model=PySRRegressor(
...
unary_operators=[
        "exp",
        "gau(x::Float64) = isfinite(x) ? exp(-0.5 * x^2.0) : NaN",
    ],
...

It was yet giving the same error:
RuntimeError: <PyCall.jlwrap (in a Julia function called from Python) JULIA: AssertionError: Your configuration is invalid - one of your operators (gau) is not well-defined over the real line. You can get around this by returning NaN for invalid inputs.

I am stuck at this point, does anyone have some advice on this? Thanks in advance!

[MilesCranmer]

If you are using precision=32 (default), then your operators have to use only Float32. So either use precision=64, or write your operator like this:

        "gau(x::T) where {T} = isfinite(x) ? exp(-T(0.5) * x^T(2)) : T(NaN)",

[MilesCranmer]

It might be nice if I convert the output automatically though... will see how hard this is to set up.

[xuyuon]

It works! Thank you very much!