control PSRN's operators via SR.jl Options

[x66ccff]

No description provided.

[x66ccff]

Currently, the operators for PSRN are directly set in the source code. I hope to be able to set them directly through the options in SR.jl, but I haven't decided yet whether to adapt automatically or set them separately.

[x66ccff]

@MilesCranmer Any thoughts on this?

SymbolicRegressionGPU.jl/src/SymbolicRegressionGPU.jl

Lines 1147 to 1167 in 24211bd

	if options.populations > 0 # TODO I don' know how to add a option for control whether use PSRN or not, cause Option too complex for me ...
	println("Use PSRN")
	# N_PSRN_INPUT = 3
	N_PSRN_INPUT = 15 # TODO this can be tuned
	psrn_manager = PSRNManager(;
	N_PSRN_INPUT=N_PSRN_INPUT, # these operators must be the subset of options.operators
	operators=["Add", "Mul", "Sub", "Div", "Identity", "Cos", "Sin", "Exp", "Log"], # TODO maybe we can place this in options
	# operators=["Add", "Mul", "Sub", "Div", "Identity"], # TODO maybe we can place this in options
	# operators=["Add", "Mul", "Neg", "Inv", "Identity", "Cos", "Sin", "Exp", "Log"], # TODO maybe we can place this in options
	# operators = ["Sub", "Div", "Identity", "Cos", "Sin", "Exp", "Log"],
	# operators = ["Sub", "Div", "Identity"],
	# operators = ["Add", "Mul", "Neg", "Inv", "Identity"],
	n_symbol_layers=2, # TODO if use 3 layer, easily crash (segfault), don't know why
	options=options,
	max_samples=20,
	# max_samples = 10
	)
	else
	println("Not use PSRN")
	end

SymbolicRegressionGPU.jl/src/PSRNmodel.jl

Lines 49 to 64 in 24211bd

	# Operator dictionary
	const OPERATORS = Dict{String,Operator}(
	"Identity" => UnaryOperator("Identity", identity_kernel!, true, identity),
	"Sin" => UnaryOperator("Sin", sin_kernel!, true, sin),
	"Cos" => UnaryOperator("Cos", cos_kernel!, true, cos),
	"Exp" => UnaryOperator("Exp", exp_kernel!, true, exp),
	"Log" => UnaryOperator("Log", log_kernel!, true, safe_log),
	# "Add" => BinaryOperator("Add", add_kernel!, false, +),
	# "Mul" => BinaryOperator("Mul", mul_kernel!, false, *),
	"Add" => BinaryOperator("Add", add_kernel!, true, +),
	"Mul" => BinaryOperator("Mul", mul_kernel!, true, *),
	"Div" => BinaryOperator("Div", div_kernel!, true, /),
	"Sub" => BinaryOperator("Sub", sub_kernel!, true, -),
	"Inv" => UnaryOperator("Inv", inv_kernel!, true, x -> 1 / x),
	"Neg" => UnaryOperator("Neg", neg_kernel!, true, x -> 0 - x),
	)

[MilesCranmer]

Why are the strings needed? Is it to pass it back to Python?

If you were to do things from scratch I would probably just write a single kernel like

struct Kernel{F<:Function}
    f::F
end

(k::Kernel)(x, y) = k.f(x, y)
#= or extra functions with ::Kernel as an argument =#

and then define mappings like

to_string(::Base.Fix{typeof(+)}) = "Add"
to_string(::Base.Fix{typeof(*)}) = "Mul"
to_string(::Base.Fix{typeof(/)}) = "Div"

and so on

[MilesCranmer]

p.s., instead of 1 / x you should use Base.inv and instead of 0 - x you can just use - by itself