Autograd support for affine transformations #2490
Conversation
Thanks! Went through it. Looks good. Added an else condition. Ran the numerical tests. They pass! |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
There was a problem hiding this comment.
8 files reviewed, 3 comments
Edit PR Review Bot Settings | Greptile
| return (fun(x0 + δ) - fun(x0 - δ)) / (2 * δ) | ||
|
|
||
|
|
||
| def _assert_close(fd_val, ad_val, tag, axis=None, extra="", tol=0.35): |
There was a problem hiding this comment.
style: The tolerance value 0.35 (35%) seems quite high for numerical comparison. Consider tightening or documenting justification.
| return anp.sum(data.get_intensity("m").values) | ||
|
|
||
|
|
||
| def finite_diff_theta(center, size, eps, theta, axis, delta=1e-3): |
There was a problem hiding this comment.
style: Default delta (1e-3) differs from global delta (0.016). Consider using the global delta by default.
| @@ -66,7 +67,6 @@ def rotate_tensor(self, tensor: TensorReal) -> TensorReal: | |||
|
|
|||
| if self.isidentity: | |||
| return tensor | |||
|
|
|||
| return np.matmul(self.matrix, np.matmul(tensor, self.matrix.T)) | |||
There was a problem hiding this comment.
logic: rotate_tensor still uses np.matmul instead of anp.matmul - needs to be consistent with autograd usage.
| return np.matmul(self.matrix, np.matmul(tensor, self.matrix.T)) | |
| return anp.matmul(self.matrix, anp.matmul(tensor, self.matrix.T)) |
Supersedes #2475
@rahul-flex I rebased all the changes into this PR. Could you have a look and check if anything is missing?
Greptile Summary
Adds comprehensive autograd support for affine transformations in Tidy3D, enabling gradient-based optimization of geometric parameters like rotation, scaling, and translation.
tidy3d/components/transformation.pytidy3d/components/geometry/base.pyto use surface mesh integration instead of xarray-based approachtests/test_components/test_box_chained_derivatives.pyvalidating automatic differentiation against finite differences within 30% tolerancetidy3d/web/api/autograd/autograd.pyto unify permittivity computation across geometry types