FinFunction does not constrain assignments to be in codomain

[corybrunson]

I'm experimenting with Catlab and wondering why i don't get an error with the following:

n = FinSet(4)
m = FinSet(2)
f = FinFunction([3,2,3,1], n, m)

I'm able to pretty_table() the result, and it shows that 1, 2, 3, and 4 are mapped to 3, 2, 3, and 1, respectively; but what does this mean when 3 is not in m?

This seems like it could cause unexpected composition errors, like this:

o = FinSet(1)
g = FinFunction([1, 1], m, o)
fg = compose(f, g)
#> BoundsError: attempt to access 2-element Vector{Int64} at index [[1, 2, 3, 3]]

Thank you for any help!

[lukem12345]

After asking around, the consensus is essentially that there are some engineering trade-offs in determining when to check whether a function is well-defined. Here, the trade-off is having fast construction of FinFunctions over explicitly checking at construction-time whether the provided codomain fits.

As you note, this mismatch is indeed caught at composition time.

The current behavior might be improved by:

1. Providing a function that checks that this condition is met.
2. Check at construction that this condition is met, unless some flag is set, like what happens when you use the @inbounds hint.

These are some suggestions raised by @epatters and paraphrased by me.

[corybrunson]

Thank you! While i'm not a power-user of AJ, my impression is that (1) would not be much help, as most users would not deviate from default settings in order to perform such checks.

Having a flag à la (2) that could be set globally (for a session) makes sense to me, but perhaps not if this is the only constructor for which the check would be relevant. Is FinFunction one of a larger class of constructors with this behavior?

I would still defer to y'all on which, if any, change is best.