Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

examples: Boundary reconstruction #1926

Open
wants to merge 4 commits into
base: master
Choose a base branch
from

Conversation

fffarias
Copy link

Hello everybody.

This PR tries to suggest reconstructing the source wave field from the one saved in the border. This should be an option for the gradient computation, in which the forward field is reconstructed inside the backwards loop. It was necessary to put a boolean argument inside the SparseTimeFunction's inject method, since I needed the injected field not to be incremented. Given that I changed an important part of the code, I added a test to validate the option of not incrementing. As an example, I also add a demonstration tutorial and application of the method in FWI.

Thanks a lot

@review-notebook-app
Copy link

Check out this pull request on  ReviewNB

See visual diffs & provide feedback on Jupyter Notebooks.


Powered by ReviewNB

implicit_dims=self.sfunction.dimensions)
for b, vsub in zip(self._interpolation_coeffs, idx_subs)]
else:
eqns = [Eq(field.xreplace(idx_subs[0]), _expr.xreplace(idx_subs[0]),
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

why are you picking idx_subs[0]? should you not traverse the whole list of idx_subs like in the increment case?

Could you do something like:

if increment:
   cls: Inc
else:
   cls: Eq
eqns = [cls(field.xreplace(....... .....) for b, vsub in zip(....)]

?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

because I thought in this case very specifically to apply boundary conditions, not source injection, for example. The ` increment=False option as I suggested should not be used for off-grid points.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

While this is true for your case, the generic case is still a standard potentially off the grid point injection without increment. So there is two option

  • Use the injection as is knowing the coefficient will still be zero for the interpolation if the point is on the grid
  • Define somehow an "on-the-grid" sparse function that ignores the interpolation.

The first case is definitely easier and should still work for your case and would boil down to what @FabioLuporini wrote earlier

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I believe the first option is not possible for the case increment = False. Because even if the point is exactly on the grid, interpolation coefficients equal to zero will vanish the fields at points that should not be vanished. Which is not the case for injecting the source (when increment = True), for example, since accumulating zeros does not change the field at the other points involved in the interpolation.

image

The way I see it, we are left with the second option that @mloubout suggested, or use subdomains that are under development in another PR if I'm not mistaken.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Indeed good point I missed that gotta think about it

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@mloubout prodding

Copy link
Contributor

@mloubout mloubout Jun 28, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Getting back here since been looking into sparse, I think that the "best" way to go for increment=Flase is to do what interpolation does i.e

  • Define sum and do the standard sum += inject(expr)
  • assign sum afterward field = sum

This would cover the off-the-grid case as well

((10, 10, 10), 1., False),
((10, 10, 10), 5., True)
])
def test_inject_time_increment(shape, result, increment):
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thank you for attaching tests to this PR!

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Would it be worth having another test to check that increment=False specified on SparseFunction with non-grid-aligned points is caught and handled appropriately?

@FabioLuporini FabioLuporini changed the title Boundary reconstruction examples: Boundary reconstruction May 24, 2022
@FabioLuporini FabioLuporini requested review from mloubout and EdCaunt May 24, 2022 07:51
@FabioLuporini
Copy link
Contributor

Thank you, Fernanda! We will review this PR carefully. Looks nice!

Meanwhile, note that some of the commit messages would have to be edited, via rebasing, to use valid tags: https://github.com/devitocodes/devito/wiki/Tags-for-commit-messages-and-PR-titles

@FabioLuporini FabioLuporini added the examples examples label May 24, 2022
@fffarias fffarias force-pushed the BoundaryReconstruction branch from ad7916e to 4952720 Compare May 24, 2022 14:21
@fffarias fffarias force-pushed the BoundaryReconstruction branch from 4952720 to 67d3f34 Compare May 24, 2022 17:42
Copy link
Contributor

@mloubout mloubout left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks so much for the contribution @fffarias this is very much appreciated. I left a couple comments for the tutorial to make it a bit more pedagogic, happy to discuss what's doable without adding too much on your plate.

implicit_dims=self.sfunction.dimensions)
for b, vsub in zip(self._interpolation_coeffs, idx_subs)]
else:
eqns = [Eq(field.xreplace(idx_subs[0]), _expr.xreplace(idx_subs[0]),
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

While this is true for your case, the generic case is still a standard potentially off the grid point injection without increment. So there is two option

  • Use the injection as is knowing the coefficient will still be zero for the interpolation if the point is on the grid
  • Define somehow an "on-the-grid" sparse function that ignores the interpolation.

The first case is definitely easier and should still work for your case and would boil down to what @FabioLuporini wrote earlier

@@ -0,0 +1,1069 @@
{
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nice tutorial. Some generic comments.

I think it would be beneficial to split and/or extend this tutorial for readability. I would

  • first focus purely on the wavefield itself and show snapshots/movie/... of the forward wavefield and its reconstruction (i.e check the snapshot tutorial for ploting)
  • Second detail the gradient computation

And on the into side, It would be beneficial to use slightly more explicit variable names, such as u and u_boundary or such. I would also completely skip the illumination part. While this is indeed very useful in practice, it adds to the complexity of the tutorial.

a = SparseTimeFunction(name=name, grid=grid, npoint=npoints, nt=nt)
dims = tuple([np.linspace(0., 1., d) for d in grid.shape])
for i in range(len(grid.shape)):
a.coordinates.data[:,i] = np.meshgrid(*dims)[i].flatten()
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

can probably compute meshgrid only once before the for loop

@@ -0,0 +1,1069 @@
{
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Might be good to have this presented as percentage of max wavefield value?


Reply via ReviewNB

@mloubout
Copy link
Contributor

mloubout commented Jun 8, 2022

Generic question. Would it be easier to use 1D Function instead of Sparse functions for this?

@fffarias
Copy link
Author

fffarias commented Jun 9, 2022

Generic question. Would it be easier to use 1D Function instead of Sparse functions for this?

possibly. I'm not sure how to get just the necessary points from and into the wavefield. Would I have to use subdomains you think? Also it may be more practical to use 2D function in space, to accommodate more layers according to the operator's spatial order.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
examples examples
Projects
None yet
Development

Successfully merging this pull request may close these issues.

4 participants