Hessian adjoint tangent projection using ForwardDiff over adjoint residual function #99
Conversation
michel2323
changed the title
michel2323
changed the title
src/autodiff.jl
Outdated
| else | ||
| error("Unsupported Jacobian structure") | ||
| end | ||
| uncompress_kernel!(arrays.J, arrays.compressedJ, arrays.coloring) |
There was a problem hiding this comment.
We have an issue here: the type of arrays.J is inferred as SparseMatrixCSC, instead of SparseMatrixCSC{Float64, Int}. That leads to some inefficiencies in uncompress_kernel!
There was a problem hiding this comment.
In fact, on line 80 of src/autodiff.jl, we should replace:
SMT = SparseMatrixCSC
```julia
by
```julia
SMT = SparseMatrixCSC{Float64, Int}
(yeah, type inference is really my new dada)
michel2323
changed the title
|
Sorry @frapac, a lot rewiring. In the end it's much less code. I removed everything we don't need. That's also why I might want to keep the history this time. In case I want to do that seeding again. I did a merge back from develop and resolved a few minor conflicts. |
| # Seeding | ||
| nmap = length(H.map) | ||
| t1sseedvec = zeros(Float64, length(x)) | ||
| for i in 1:nmap |
There was a problem hiding this comment.
I would maybe add an @inbounds macro there
There was a problem hiding this comment.
Ok. The bounds checking actually creates a real overhead. So I'm not sure whether to use it everywhere or not.
There was a problem hiding this comment.
In fact, even if bounds checking creates an overhead, it is actually negligible on modern architecture, as far as I observed
src/polar/kernels.jl
Outdated
| F[npq + i] += coef_cos * sin_val - coef_sin * cos_val | ||
| end | ||
| if i > npv | ||
| F[npq + i] += coef_cos * sin_val - coef_sin * cos_val |
There was a problem hiding this comment.
I don't understand the following code. It looks like to me that this line is redundant with the line 146 above?
| adj_v_a[to] -= adj_aij | ||
| end | ||
| # qinj is not active | ||
| # if i > npv |
There was a problem hiding this comment.
Maybe remove dead comment? Or explain why we are keeping it from now?
michel2323
force-pushed
the
ms/adjoint
branch
from
dc6cd20
to
fe8d992
Compare
No description provided.