Type conversion happens before matlab Set method is called in classes, problem with setting int64(Nan) = [] default is 0

I am trying to set the default behaviour of a class so that int64(NaN) is set to empty and Not to the default 0.

Here is a simple class example class example:

classdef TestClass < matlab.mixin.SetGet

% CLASSDEF TestClass class to test add NaN

%

properties

a int64 {mustBeNonNan} = []

end % methods

methods

function obj = TestClass()

end % TestClass

function set.a(obj, val)

disp(val)

class(val)

obj.a = val;

end

end % methods

end % class

If you try and set a = Nan here it gets set to 0, in the set method it is already a int64 and not a double NaN. Is there a way to change this? I don't really want to addListeners and handle preSet methods. I also want the class type to be an int64 rather than not give it a type and then set the type in the set method.

On another note why is int64(NaN) set to 0 anyway?

This would work, but then the class property would not be an int64.

classdef TestClass < matlab.mixin.SetGet

% CLASSDEF TestClass class to test add NaN

%

properties

a

end % methods

methods

function obj = TestClass()

obj.a = int64([]);

end % TestClass

function set.a(obj, val)

if isnan(val), val = []; end

obj.a = int64(val);

end

end % methods

end % class

[Matt J]

On another note why is int64(NaN) set to 0 anyway?

Integer types can't represent NaNs. What bit pattern would you choose to represent it? Whichever you choose, it would sacrifice one of the finite integers that int64 will normally hold for you.

[Steven Lord]

Having myInt64(NaN) (where myInt64 is your class name) return [] instead of 0 will cause problems when or if you start working with arrays. If you replaced int64 in the example below with the constructor of your int64-like class, do you want the variable a to be a "jagged" array? That doesn't open a can of worms, it opens a 55 gallon drum of worms.

a = int64([1 2; 3 NaN])

a = 2×2

1 2 3 0

[Matt J]

I also want the class type to be an int64 rather than not give it a type and then set the type in the set method.

I don't see why you would prefer that. One is just a shorthand for the other.

[Benjamin Klein]

Thank you for the answers, it's a good point about arrays, guess why int64(NaN) becomes a zero. In this case it's a singular value so in my class I define:

a(1,:) int64 = []

So that isn't a problem for me. It would be best to just leave it untyped and deal with that in the set method, the running code won't change, my problem with this is that only when users looks at the object they won't know what type of variable they have to set it to such as:

myclass with properties:

a: [1×0 int64]

b: [1×0 string]

c: [1×0 string]

I am really intested in the subsasgn answer (below) because Matlab treats that differently so that it doesn't evaluate val before it reaches the subsasgn method. Following various help files I found an example to create your own type int64p:

classdef int64p < int64

methods

function obj = int64p(data)

if nargin < 1, data = int64([]); end

data(isnan(data)) = [];

obj = obj@int64(data); % Store data on superclass

end

end

end

So I define my own version of int64 as int64p and then I can have my class as:

classdef myclass < matlab.mixin.SetGet

% CLASSDEF myclass class to test add NaN

%

properties

a(1,:) int64p = []

b(1,:) string = []

c(1,:) string = []

end % methods

methods

function set.a(obj, val)

if isnan(val), val = []; end

obj.a = int64(val);

end

end % methods

end % class

Answer 1

[Matt J]

One possibility:

>> obj=myclass; obj.a=1:5, obj.a(2)=nan

obj =

myclass with properties:

a: [1 2 3 4 5]

obj =

myclass with properties:

a: [1 3 4 5]

classdef myclass < matlab.mixin.SetGet
% CLASSDEF TestClass class to test add NaN 
%
    properties 
	    a
    end % methods

    methods

        function  set.a(obj, val)

            val(isnan(val))=[];
    	    obj.a = int64(val);
            
        end

        function val=get.a(obj)
             val=double(obj.a);
        end
        
    end % methods
end % class

Answer 2

[Matt J]

Another approach, using subsasgn(). This one avoids conversion of obj.a to double.

>> obj=myclass; obj.a=1:5, obj.a(2)=nan

obj =

myclass with properties:

a: [1 2 3 4 5]

obj =

myclass with properties:

a: [1 3 4 5]

classdef myclass < matlab.mixin.SetGet
% CLASSDEF TestClass class to test add NaN 
%
    properties 
	    a int64
    end % methods

    methods

        function  obj=subsasgn(obj,S,val)

              
              obj=builtin('subsasgn',obj,S,val);
              if strcmp(S(1).type,'.') && strcmp(S(1).subs,'a')
 
                  b=false(size(obj.a));
                  if numel(S)>1
                   b(S(2).subs{:})=isnan(val);
                  else
                   b=isnan(val);
                  end

                  obj.a(b)=[];

              end
        end
    end % methods
end % class