Step operator

[AKHCE]

Hi,
I want to define customized operators which have a "if-then-else" structure. For example,
lambda x: 1 if x>0 else -1 if x<0 else 0
and
lambda x: 1 if x>0 else 0 if x<0 else 0.5

To do so in PySR, I wrote the following code:

model = PySRRegressor(
    model_selection="best", 
    binary_operators=["*", "+", "-", "/"],
    unary_operators=["Mystep", "Mysign"],
    extra_sympy_mappings={"Mystep": lambda x: 1 if x>0 else 0  if x<0 else 0.5, 
                                            "Mysign": lambda x: 1 if x>0 else -1 if x<0 else 0},
)

But, it doesn't work!
Any advice is much appreciated

[MilesCranmer]

Here's an implementation which might be what you're looking for: #125

Essentially you have to use sympy.Piecewise, rather than Python's if, in order to get a correct mapping to sympy

[AKHCE]

I do really appreciate your speedy reply. Yes, it is working, and I see heaviside(.) in correlation. Just wondering if you could please give a bit more elaboration on "T(x>=0)", what does this say? I mean what if x<0?
Many thanks

[MilesCranmer]

That is Julia syntax for a function. See https://docs.julialang.org/en/v1/manual/methods/#Parametric-Methods for more details.

x>=0 will turn into true if true, false if false.

Then, the T is whatever type the input is. WIth default settings in PySR, T would become a Float32.

Wrapping a value with a type, like Float32(x) would perform a conversion. Thus, Float32(true) would become 1.0f0, and Float32(false) would become 0.0f0.

[AKHCE]

I got it!
Thanks

[MilesCranmer]

Also, an alternate and maybe faster version would be:

heaviside(x::T) where {T} = x>=0 ? one(T) : zero(T)

By the way, if you are up for it, I'd welcome a PR adding this as a built-in operator! (so people don't have to define it themselves.)

[AKHCE]

Yes, please.
Your library is literally AWESOME ! and salient point is that I as a general user don't need to dig into detailed programming. So, then I recommend enriching operators a bit further e.g. heaviside.
Thank you so much