Conversation
refactor: split drift into a package without behavior changes
* feat: add drift registration diagnostics * feat: add structured drift reports * feat: add static drift diagnostic plots * test: validate drift diagnostics on frozen results * fix: keep drift diagnostics consistent with fitted state
* feat: add reference-based spectrum-image correction * fix: reject ambiguous spectrum-image reference matches
* feat: add reference-based spectrum-image correction * fix: reject ambiguous spectrum-image reference matches * feat: propagate drift across 4D-STEM scan axes * test: protect 4D-STEM detector coordinates * test: compare full detector signature field * test: verify 4D-STEM coverage and output residency * fix: stream native 4D-STEM scan correction * test: compare native detector frames across backends * test: set float32 backend parity tolerance
* feat: add reference-based spectrum-image correction * fix: reject ambiguous spectrum-image reference matches * feat: propagate drift across 4D-STEM scan axes * test: protect 4D-STEM detector coordinates * test: compare full detector signature field * test: verify 4D-STEM coverage and output residency * fix: stream native 4D-STEM scan correction * test: compare native detector frames across backends * test: set float32 backend parity tolerance * feat: preserve calibration during drift downsampling * test: lock downsampled coordinate provenance * perf: add average-pooled affine pyramid search * test: compare native and pyramid affine rates * feat: add strip-wise residual correction primitives * fix: make nonrigid regularization portable across torch backends * feat: propagate residual fields to calibrated element maps * test: lock residual-correction publication contracts * test: recover known strip residual shifts * feat: complete automatic affine manuscript parity * feat: complete multi-knot manuscript parity * test: verify complete manuscript workflow parity * fix: keep 4D-STEM integration usable without optional GPU package * refactor: adopt the scientist-facing Drift API * test: verify manuscript output frame contracts * fix: preserve natural corrected output frames * feat: expose manual translation alignment
Collaborator
Author
|
@wwmills ready for review. If you agree with the API proposed and folder architecture, we can probably merge so that you can add your new algo as needed and refactor with upcoming PRs. This PR is to enable further scaling of features and reproduce manuscript results. |
bobleesj
marked this pull request as ready for review
August 26, 2026 17:41
…#24) * feat: center-out scanline drift correction, with low-pass alignment costs Adds a scipy non-rigid stage that solves scanlines one at a time, outward from the center of the slow axis, each row seeded from the neighbor one row closer to the center. Fixing the center row fixes the gauge, which is where most of the uncontrolled error in an unanchored per-row solve lives. core/centerout.py solve_rows_center_out() solves one scan's rows center-out; the prediction for each row is both the L-BFGS-B start point and the center of a step_cap_px box. correct_nonrigid_center_out() is the sweep loop: center-out solve against the mean of the other warped scans, residual knot smoothing, re-warp, translation alignment. It owns its own loop rather than branching correct_nonrigid, whose body differs at nearly every step, so the torch and scipy paths coexist as two methods. core/knots.py transform_row_numpy() maps one scanline's knots to canvas coordinates in numpy, for K=1 and K>=2. The per-row scipy cost needs one row without building the full (H, W) torch coordinate grid. Conventions match the torch path exactly: the fast axis spans num_rows - 1 / num_cols - 1, and num_rows is the full image height. A mismatch there would not raise -- the row cost and the warp would optimize different problems. core/warping.py, core/affine.py, core/nonrigid.py soften_and_lowpass() restricts each image to its valid-weight bounding box, applies a separable linear ramp, mean-subtracts inside the mask, and optionally Gaussian low-passes in Fourier space. Threaded through translate_align, cross_corr_batch, the affine grid search (where the ramp doubles as cost_taper) and correct_nonrigid. All defaults leave existing call sites unchanged. imaging/a_sites.py fit_a_sites() finds peaks, indexes them against the lattice and fits a 2-D Gaussian per site; refit_adaptive() re-fits within an n_sigma window. Also switches a few docstrings to US spelling. * feat: bond-cloud statistics in a_sites Moves the nearest-neighbor bond analysis out of the figure notebook and next to the A-site fitting that produces its input. bond_statistics() collects the six nearest-neighbor bond vectors per site by walking the fitted lattice indices, clips outliers at 4 sigma over three rounds, and returns the per-direction clouds with their 2-D scatter. cloud_ellipse() and cloud_half_extent() give the covariance ellipse geometry and the extent the widest magnified ellipse needs. All three are numpy-only and consume the dict fit_a_sites() returns, so the analysis is reusable without matplotlib. Plotting stays in the notebook. * refactor!: lowpass is a real-space sigma in pixels soften_and_lowpass took ``lowpass`` as a reciprocal-space scale: the kernel was exp(-0.5 * lowpass**2 * f**2) with f in cycles per pixel, so lowpass=32 blurred by sigma = 32 / (2*pi) = 5.1 px. Callers had to carry that factor in their heads, and papers reporting "a 32 pixel Gaussian" were describing a 5 px one. The kernel is now exp(-0.5 * (2*pi*sigma)**2 * f**2), so ``lowpass`` is the real-space sigma directly. BREAKING: existing callers must divide their old value by 2*pi -- lowpass=32 becomes lowpass=5.1. * fix: apply the low-pass settings to the affine's closing registration correct_affine ends each search pass by re-registering the pair with warp_and_translate, and that call received neither the low-pass, the taper, nor the subpixel mode used by the grid search. The final knot positions -- and the _knots_after_affine checkpoint taken from them -- therefore came from a plain DFT alignment on unfiltered images, so callers needed a second align_translation afterwards to get the registration their settings asked for, and plot_combined(stage="affine") disagreed with the corrected output. Also collapses lowpass_ramp into cost_taper. The two tapered different copies of the same images -- one for estimating the shift, one for scoring -- and no caller set them independently. correct_nonrigid and correct_nonrigid_center_out take cost_taper for the same reason. --------- Co-authored-by: wwmills <226772206+wwmills@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What problem this PR addreseses
Support 3D, 4DSTEM drift correction.
PRs broken down as shown below:
Before
After
Public API change
from_data(...)from_images(...)align_translation()align_affine()correct_affine()align_nonrigid()correct_nonrigid()generate_corrected()corrected()number_knots,stepnum_knots,max_drift_rateAffine speed:
The pyramid downsampling strategy enables real-time drift correction on any personal laptops.
What should the reviewer(s) do
Tutorials are available here: https://github.com/bobleesj/quantem-tutorials/tree/drift-tutorial/tutorials/imaging/drift
.