Let sympy use log2(x) instead of log(x)/log(2)

[nerai]

Fixes #711

The case of log10 is also included.

I was not immediately able to run the unit tests, but I tested it with this code:

import numpy as np
from pysr import PySRRegressor
import sympy
from sympy.codegen.cfunctions import log2

X = np.abs(np.random.randn(500, 1)) + 2
y = np.log2(X[:, 0] ** 2) + 10 / np.log10(X[:, 0])
model = PySRRegressor(
	unary_operators=[
		"log2",
		"log10",
	],
	maxsize = 20,
)
model.fit(X, y)
print(model.equations_["sympy_format"])

Equation: y = (10.003 / log10(x₀)) + log2(x₀ * x₀)
Translated to sympy: log2(x0*x0) + 10.003168/log10(x0)

[MilesCranmer]

codegen looks like an internal utility? https://docs.sympy.org/latest/modules/utilities/codegen.html Is there anything simpler we can use?

[coveralls]

Pull Request Test Coverage Report for Build 10703019687

Details

• 0 of 0 changed or added relevant lines in 0 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+0.2%) to 93.982%

---

Totals
Change from base Build 10518940981:	0.2%
Covered Lines:	1140
Relevant Lines:	1213

---

💛 - Coveralls

[nerai]

codegen looks like an internal utility? https://docs.sympy.org/latest/modules/utilities/codegen.html Is there anything simpler we can use?

I'm not familiar with sympy, so I'm not sure. The docs indicate that the regular log function is translated into the paired representation, though it is not explicit:

To get a logarithm of a different base b, use log(x, b), which is essentially short-hand for log(x)/log(b).

Grepping the sympy source code for instances of log2 did not reveal anything of note, though I may have missed it.

[MilesCranmer]

I am a bit worried about using this version over just log(x, 2) because codegen looks to be pretty uncommon in user code: https://github.com/search?q=/sympy.codegen.cfunctions/+language:python+&type=code. Is there anything in the sympy docs that recommend this approach over just log(x, 2)?

[nerai]

Good point. I asked over at sympy, and it seems it's a rare but supported function. It is documented at https://docs.sympy.org/latest/modules/codegen.html#sympy.codegen.cfunctions.log2 .

[MilesCranmer]

Thanks! Okay I’m on board then. Now, one other thing is we need the codegen import:

The codegen callable is not in the sympy namespace automatically, to use it you must first import codegen from sympy.utilities.codegen

This seems to have been done by some other package, but ideally we should also run this within PySR.

[nerai]

Thanks! Okay I’m on board then. Now, one other thing is we need the codegen import:

The codegen callable is not in the sympy namespace automatically, to use it you must first import codegen from sympy.utilities.codegen

This seems to have been done by some other package, but ideally we should also run this within PySR.

I think this refers only to the callable, i.e. the function codegen, not the namespace? That's how I read the example at https://docs.sympy.org/latest/modules/codegen.html#codegen-sympy-utilities-codegen . It is a bit confusing for sure.

[MilesCranmer]

Just checking in on this – could you add the line

from sympy.utilities.codegen import codegen  # noqa: F401

to the imports? This is needed according to the docs page.

[nerai]

Can you check if that's correct? My understanding of the docs page is that it refers to the codegen callable, not the codegen namespace. The code only uses the namespace, so the import should not be not required.