propagate to add two stochastic triples together

[GuusAvis]

Hello!

I'm trying to use the propagate function to extract derivatives from some Monte Carlo simulations. However, I noticed that the function does not behave as I would expect in even some simple use cases. A minimal example is given by simply adding two stochastic triples together:

using StochasticAD

function add(x::Real, y)
    x + y
end

function add(x::StochasticAD.StochasticTriple, y)
    StochasticAD.propagate(add, x, y; keep_deltas=true)
end

# x = 1
x = stochastic_triple(1)
y = stochastic_triple(5)

add(x, y)  # 6 + 0ϵ
x + y  # 6 + 2ϵ
add(x, y) == x + y  # false

I expected that propagate would automatically ensure that this simple function would correctly propagate the perturbations but this does not seem to be the case. Am I doing something wrong or is this a bug?

Thanks!

[gaurav-arya]

At quick glance, I think it's because you need keep_deltas = Val(true), not just true. Although we should definitely catch the latter case and error.

[GuusAvis]

At quick glance, I think it's because you need keep_deltas = Val(true), not just true. Although we should definitely catch the latter case and error.

Many thanks for the quick reply and catching that silly mistake on my end! However I don't think replacing true with Val(true) fixes the issue. Moreover, I noticed that it persists even when one of the numbers is not a stochastic triple:

using StochasticAD

function add(x, y)
    x + y
end

function add(x::StochasticAD.StochasticTriple, y)
    StochasticAD.propagate(add, x, y; keep_deltas=Val(true))
end

x = stochastic_triple(1)
y = 5

add(x, y)  # 6 + 0ϵ
x + y  # 6 + 1ϵ
add(x, y) == x + y  # false

[gaurav-arya]

Just a hunch: can you try writing stochastic_triple(1.0)?

[GuusAvis]

Just a hunch: can you try writing stochastic_triple(1.0)?

Thanks, that does the trick! Any idea why the behavior for triples for floats is different than that for triples of ints?

[gaurav-arya]

Well, an integer shouldn't really be able to possess continuous perturbations. But again, there definitely shouldn't be a silent error here, I need to fix that!