I have encountered a problem in my code. While I cannot easily write a MWE for my particular case, I thought that perhaps dealing with the following issue would help me further.
Let us say, that we have some data y and parameters x. Using Zygote, we can calculate the gradient of the following:
using Zygote
x = [1.0;3.0;5.0] # parameters
y = [1.0;3.0;5.0].+ 10 # fake data
# calculate gradient wrt x
Zygote.gradient(x -> sum(abs2.(x-y)), x)
Now suppose that y contains missing values:
using Zygote
x = [1.0;3.0;5.0] # parameters
y = [1.0;3.0;missing].+ 10 # fake data
# calculate gradient wrt x
Zygote.gradient(x -> sum(abs2.(x-y)), x)
The above, as expected, will not work.
I tried using instead:
Zygote.gradient(x -> sum(abs2.(skipmissing(x-y))), x)
which gives me an error. I also tried:
Zygote.gradient(x -> sum(skipmissing(abs2.(x-y))), x)
which also doesn’t work, but, interestingly, gives me a different error.
Any advice appreciated.