2nd round of autograd fixes #1923
Conversation
tylerflex
commented
Aug 27, 2024
•
tylerflex
commented
- some adjoint source fixes for FieldData involving H fields.
- better gradients for Box shifting boundaries based on permittivity.
- re-introduce automatic permittivity grabbing for geometry group shifting boundaries.
- better error and warning handling across the board.
- adds a way to visualize adjoint fields (only for development purposes currently, no public API).
tylerflex
force-pushed
the
tyler/autograd_/cleanup2
branch
2 times, most recently
from
0a20211
to
b74f1a3
Compare
tylerflex
force-pushed
the
tyler/autograd_/cleanup2
branch
from
b74f1a3
to
b0b966c
Compare
tylerflex
force-pushed
the
tyler/autograd_/cleanup2
branch
from
b0b966c
to
ecc1e94
Compare
| @@ -471,7 +471,7 @@ def field_vol_postprocess_fn(sim_data, mnt_data): | |||
| def field_point_postprocess_fn(sim_data, mnt_data): | |||
| value = 0.0 | |||
| for _, val in mnt_data.field_components.items(): | |||
| value += abs(anp.sum(val.values)) | |||
| value += abs(anp.sum(abs(val.values))) | |||
There was a problem hiding this comment.
double abs? not that it matters i guess
There was a problem hiding this comment.
Could be nice to have private debug run function, or maybe a @debug decorator that could do things like this? Not really important but might be convenient to include in the future.
There was a problem hiding this comment.
which line? the adjoint field visualization?
There was a problem hiding this comment.
yeah exactly, that and the global toggle to trigger that code
|
|
||
| # make source go backwards | ||
| if "H" in name: | ||
| values *= -1 |
There was a problem hiding this comment.
Is this what we do in the adjoint plugin too?
I am trying to think about whether this is the correct thing to do in all cases. It's probably possible to see it somehow from the adjoint formulation if written for the H field. I think this is correct. I'm just wondering a little because the intuition of this minus sign "making source go backwards" makes sense for a real propagating mode, but it gets a bit murkier for evanescent or lossy modes. The other option is to take np.conj(values). If the E field is real and the mode is lossless and propagating, then the H field is imaginary and the two operations are the same. In general the np.conj(values) operation would be like time reversal while the values *= -1 operation is like reflection of the mode in the source plane, which produce different results in some cases.
I do think the -1 is correct here though but do you remember if it comes out of the adjoint math?
