Closed dpo closed 5 years ago
TRON is now failing to converge, so those tests fail. @abelsiqueira
Using @view
fixed a problem. There is some other error that I'm looking into.
Here is the fix for the other errors: https://github.com/JuliaSmoothOptimizers/Optimize.jl/pull/68/commits/85242ee9657cd889806f820e1b514cd3f11040c7 I've pushed to #68 to test.
The error is that (Preallocated) LinearOperators and ForwardDiff are not working together, because the memory used by the linear operator can't store Dual numbers.
Thank you. Does that mean we have to use operators without preallocation when using AD models?
We can't have the objective function or constraints using LinearOperators. e.g. f(x) = dot(x, A * x)
and c(x) = A * x - b
only work if A
is a matrix or non-allocating linear operators. Its a ForwardDiff
limitation.
However, we can have operators with preallocation after the differentiation, such as hess_op!
and jac_op
. Those aren't broken if f
is accepted by ForwardDiff
.
ERROR: The following package names could not be resolved:
* Krylov (not found in project, manifest or registry)
Please specify by known `name=uuid`.
The new Pkg3
is really frustrating.
Looks like we'll need a Krylov release.
Supersedes #68.