Differentiable local field projection #1822

yaugenst-flex · 2024-07-11T12:36:07Z

No description provided.

tylerflex · 2024-07-25T20:57:50Z

👀

tylerflex

a few comments. I'm mostly a little confused about the base.py stuff?

CHANGELOG.md

tidy3d/components/autograd/utils.py

tidy3d/components/base.py

tylerflex

Some comments.. Mostly things I thought could be refactored, renamed, simplified.

Otherwise I think it's consistent / correct. Only thing I'm not 100% sure on is in the DataArray.static property.

vals = self.values if self.values.size == 0 else np.vectorize(getval)(self.values)

will getval work if self.values is a np.ndarray of dtype=object?

otherwise feeling pretty good about it in its current state

tidy3d/components/autograd/utils.py

tidy3d/components/base.py

tidy3d/components/data/data_array.py

yaugenst-flex · 2024-09-04T09:12:29Z

@tylerflex There are some changes to DataArray properties here. The naming and mechanics of these are probably not ideal, but since it's not user facing let's address them in the DataArray refactor? We'll end up changing most of these things anyways.
Also, #1939 lists this for 2.8rc1, should I rebase on pre/2.8? Although I think it makes sense to include this only with the gradient fixes in develop. So either as part of 2.7.3 or rebase to pre/2.8 only after develop is merged.

tylerflex · 2024-09-04T12:27:07Z

Also, #1939 lists this for 2.8rc1, should I rebase on pre/2.8? Although I think it makes sense to include this only with the gradient fixes in develop. So either as part of 2.7.3 or rebase to pre/2.8 only after develop is merged.

@momchil-flex what do you think? basically there were some changes in my (recently merged) #1923 that Yannick used in this PR. So ideally we would do another develop -> pre/2.8 merge before Yannick rebases against 2.8?

momchil-flex · 2024-09-04T13:24:51Z

@yaugenst-flex what do you think will be simplest? I'm already running notebook tests for 2.7.3 and since it's hard to tell if this PR may not have various small effects, I'd rather not include it for now. However we could merge it to develop after 2.7.3 and then eventually into pre/2.8, or we can merge develop into pre/2.8 again and then you rebase the PR. Either way it will likely only come out in 2.8.0rc1 since we may not do any other 2.7. patches, but if we did the first approach would put it in there too.

yaugenst-flex · 2024-09-04T13:56:12Z

@momchil-flex I think it makes sense either way to merge develop into pre/2.8 after the 2.7.3 release so we keep them somewhat in sync. I'll rebase to pre/2.8 after that, shouldn't be a problem.

tylerflex

Looks really good! thanks @yaugenst-flex . I dont know if you want @QimingFlex to look at the field projection bits? I couldn't understand most of the physics of the operations, but did follow through the code and thought it looked good.

tylerflex · 2024-09-04T14:14:17Z

pyproject.toml

@@ -38,6 +38,7 @@ dask = "*"
 toml = "*"
 autograd = "1.6.2"
 scipy = "*"
+opt-einsum = "^3.3.0"


we need to be careful about this due to the danger of bloating our core dependencies. It became an issue recently and we agreed to let MC know if we add any dependencies. Is this absolutely necessary you think?

opt-einsum is a pretty tiny dependency (~200kb) and itself only depends on numpy, and it's already a dependency of jax (although that's not core). Otherwise I can also rewrite that part using regular numpy einsum, not a big deal.

@lei-flex what do you think about adding this dependency?

tylerflex · 2024-09-04T14:15:07Z

tests/ruff.toml

@@ -10,4 +10,5 @@ extend-ignore = [
 "E731", # lambda assignment
 "F841", # unused local variable
 "S101", # asserts allowed in tests
+ "NPY201", # numpy 2.* compatibility check


snuck in from another PR?

Actually no, this got flagged (np.trapz usage) and I disabled it. We should enable it again once we support numpy 2.

tidy3d/components/autograd/utils.py

tidy3d/components/data/data_array.py

tylerflex · 2024-09-04T14:22:32Z

tidy3d/components/data/data_array.py

 return super().conj(*args, **kwargs)

 @property
 def real(self):
 """Return the real part of this DataArray."""
- if self.tracers is not None:
- return self.__array_wrap__(anp.real(self.tracers))
+ if isbox(self.values.flat[0]) or AUTOGRAD_KEY in self.attrs:


perhaps we can refactor this logic out (eg. the one in real, conj, etc).

def apply_traced(self, fn): if isbox(self.values.flat[0]) or AUTOGRAD_KEY in self.attrs: tracers = anp.array(self.values.tolist())) new_values = fn(tracers) return self.insert_tracers(new_values) return fn(self) @property def real(self): return self.apply_traced(anp.real)

also it seems the anp.array(self.values.tolist()) and if statement both could use the convenience methods / functions used to check these things?

Not really sure how to handle the super() call this way? But yeah I can at least partially refactor this..

do we need super or can np.real(self) suffice?

wouldn't that be infinite recursion? pretty sure np.real(self) will call self.real

tylerflex · 2024-09-04T14:24:06Z

tidy3d/components/data/data_array.py

I am blanking on a lot of the discussions we had regarding the data array handing, but will mostly assume that this logic follows what we already talked about before?

Is there anything you think I should take a more detailed look at?

Yeah I think it's fine for the scope of this PR, so no need to dive deep for now, I vote that we put together a proper DataArray refactor soonish though.

tylerflex · 2024-09-04T14:24:53Z

tidy3d/components/data/monitor_data.py

- power_phi = 0.5 * np.real(-self.Ephi.values * np.conj(self.Htheta.values))
+ conj = np.vectorize(np.conj)
+ real = np.vectorize(np.real)
+ power_theta = 0.5 * real(self.Etheta.values * conj(self.Hphi.values))


do we need to do this treatment in other parts of this class? I think this kind of operation does pop up from time to time?

If we want to differentiate w.r.t. those methods then yeah this needs to be done. Although I'm hoping that we can get rid of this "hack" with a DataArray refactor...

tylerflex · 2024-09-04T14:33:52Z

tidy3d/plugins/autograd/README.md

@@ -144,7 +143,12 @@ The following components are traceable as outputs of the `td.SimulationData`
 - `SimulationData.get_intensity`
 - `SimulationData.get_poynting`

+- Local field projections using `td.FieldProjector.project_fields()` for:


was this mentioned as a "to do" item? if so would be good to remove it.

not seeing that anywhere in this PR, but I know I've seen it somewhere...

yaugenst-flex · 2024-09-04T16:03:10Z

Hey @QimingFlex, would be great if you could have a quick look at the field projection bits in this PR, I reworked some of it to more easily accomodate autograd. I did try to make sure that all of the field projection values that come out of this are still exactly the same as before, but it wouldn't hurt to have another pair of eyes on it.

….project_fields()`

yaugenst-flex self-assigned this Jul 11, 2024

yaugenst-flex force-pushed the yaugenst-flex/autograd-farfield branch 4 times, most recently from 9e2bc35 to cecf530 Compare July 17, 2024 14:32

yaugenst-flex marked this pull request as ready for review July 31, 2024 20:38

yaugenst-flex requested a review from tylerflex July 31, 2024 20:40

tylerflex reviewed Aug 1, 2024

View reviewed changes

CHANGELOG.md Outdated Show resolved Hide resolved

tidy3d/components/autograd/utils.py Outdated Show resolved Hide resolved

tidy3d/components/base.py Show resolved Hide resolved

yaugenst-flex force-pushed the yaugenst-flex/autograd-farfield branch 4 times, most recently from 260293e to b387268 Compare August 2, 2024 11:48

tylerflex mentioned this pull request Aug 2, 2024

adjoint Field projection monitor support #1114

Open

tylerflex linked an issue Aug 2, 2024 that may be closed by this pull request

adjoint Field projection monitor support #1114

Open

tylerflex added 2.7 will go into version 2.7.* .2 labels Aug 2, 2024

tylerflex reviewed Aug 2, 2024

View reviewed changes

yaugenst-flex force-pushed the yaugenst-flex/autograd-farfield branch 3 times, most recently from 65ff9c3 to 5542fff Compare August 5, 2024 11:56

tylerflex added .3 and removed .2 labels Aug 5, 2024

yaugenst-flex force-pushed the yaugenst-flex/autograd-farfield branch 3 times, most recently from 5ec140c to 2fe0315 Compare August 7, 2024 19:57

yaugenst-flex force-pushed the yaugenst-flex/autograd-farfield branch from 2fe0315 to 5bb4a0a Compare August 14, 2024 11:11

yaugenst-flex force-pushed the yaugenst-flex/autograd-farfield branch 3 times, most recently from de9a4ba to 3cd6ae9 Compare August 30, 2024 19:02

yaugenst-flex force-pushed the yaugenst-flex/autograd-farfield branch 2 times, most recently from 52b3b7f to 77c5973 Compare September 3, 2024 15:25

tylerflex mentioned this pull request Sep 3, 2024

Autograd 2.8 plan #1939

Open

11 tasks

yaugenst-flex force-pushed the yaugenst-flex/autograd-farfield branch from c770e70 to c5a5caa Compare September 4, 2024 09:04

yaugenst-flex requested a review from tylerflex September 4, 2024 09:09

yaugenst-flex force-pushed the yaugenst-flex/autograd-farfield branch from c5a5caa to 350d1db Compare September 4, 2024 09:40

yaugenst-flex mentioned this pull request Sep 4, 2024

Inverse design demo for field projection flexcompute/tidy3d-notebooks#165

Open

tylerflex reviewed Sep 4, 2024

View reviewed changes

yaugenst-flex requested a review from QimingFlex September 4, 2024 16:01

yaugenst-flex force-pushed the yaugenst-flex/autograd-farfield branch from 0e92e95 to c064c42 Compare September 5, 2024 17:22

yaugenst-flex added 2 commits September 9, 2024 08:09

feat: Differentiable local field projections using `td.FieldProjector…

1fdc847

….project_fields()`

Tyler comments 1

862d5b8

yaugenst-flex force-pushed the yaugenst-flex/autograd-farfield branch from c064c42 to 862d5b8 Compare September 9, 2024 06:09

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Differentiable local field projection #1822

Differentiable local field projection #1822

yaugenst-flex commented Jul 11, 2024 •

edited

Loading

tylerflex commented Jul 25, 2024

tylerflex left a comment

tylerflex left a comment

yaugenst-flex commented Sep 4, 2024 •

edited

Loading

tylerflex commented Sep 4, 2024

momchil-flex commented Sep 4, 2024

yaugenst-flex commented Sep 4, 2024

tylerflex left a comment

tylerflex Sep 4, 2024

yaugenst-flex Sep 4, 2024 •

edited

Loading

tylerflex Sep 4, 2024

tylerflex Sep 4, 2024

yaugenst-flex Sep 4, 2024

tylerflex Sep 4, 2024

tylerflex Sep 4, 2024

yaugenst-flex Sep 4, 2024

tylerflex Sep 4, 2024

yaugenst-flex Sep 4, 2024

tylerflex Sep 4, 2024

yaugenst-flex Sep 4, 2024

tylerflex Sep 4, 2024

yaugenst-flex Sep 4, 2024

tylerflex Sep 4, 2024

yaugenst-flex Sep 4, 2024

yaugenst-flex commented Sep 4, 2024

Differentiable local field projection #1822

Are you sure you want to change the base?

Differentiable local field projection #1822

Conversation

yaugenst-flex commented Jul 11, 2024 • edited Loading

tylerflex commented Jul 25, 2024

tylerflex left a comment

Choose a reason for hiding this comment

tylerflex left a comment

Choose a reason for hiding this comment

yaugenst-flex commented Sep 4, 2024 • edited Loading

tylerflex commented Sep 4, 2024

momchil-flex commented Sep 4, 2024

yaugenst-flex commented Sep 4, 2024

tylerflex left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

yaugenst-flex Sep 4, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

yaugenst-flex commented Sep 4, 2024

yaugenst-flex commented Jul 11, 2024 •

edited

Loading

yaugenst-flex commented Sep 4, 2024 •

edited

Loading

yaugenst-flex Sep 4, 2024 •

edited

Loading