yijiang1 / fold_slice Goto Github PK

Currently the best way to correct large position errors in ptychography is to do low-resolution reconstruction using cropped data. However, sometimes reconstruction may fail if diffraction patterns are too small, preventing position correction at large pixel size.
I think it's possible to do coarse (or any arbitrary pixel size) during the position correction step by resizing the object and the probe. Thus one can use large diffraction patterns and still perform efficient position correction.

The TV minimization in GPU engines currently uses the Chambolle algorithm, which seems to be too aggressive with the default parameters. It's worth testing some other TV algorithms.
Here is a good review: https://doi.org/10.1016/j.softx.2019.04.003

1. position errors not rotated but overlaid on rotated object phase

In ptycho_solver.m, the loop "ll = 1:max(par.object_modes, par.Nscans)" for object constraints (e.g. apply_smoothness_constraint) will result in error if multiple scans share the same object. In this case, the number of object modes is less than par.Nscans.
Need to find a way to determine the correct number of objects based on eng. share_object.

A collection of Matlab bugs people may encounter when using fold_slice:

1. Matlab 2023b: Sometimes, the h5write function may throw out a "File or folder not found" error depending on the system.

if using 1 layer and GPU_MS engine, the probe is not in the center but the corner. The object is basically random.
of course we can use GPU instead of GPU_MS, but I think it's good to make single slice work in GPU_MS engine too.

This could be useful for resolving layer structures in multi-slice ptychographic reconstruction
https://arxiv.org/abs/2102.00869

If diffraction patterns are removed based on p.avg_photon_threshold or p.avg_photon_threshold_ub, loading initial scan positions from .mat files will not work because the number of positions is different from the number of diffraction patterns.

Add a feature to try loading a ptychographic reconstruction several times before throwing out an error. This could be useful for large-scale tomography processing and when the data drive is unstable.

Record the pixel size of each ptychographic reconstruction when loading them for tomography.
This is useful if ptycho reconstructions have different pixel sizes.

I think it's possible to make use of existing code to optimize sample thickness (delta_z) during GPU_MS reconstruction.

Functions worth checking out:
+GPU_MS/LSQML.m
+GPU_MS/private/gradient_NF_propagation_solver.m
+ML_MS/gradient_ptycho_MS.m
+ML_MS/calc_ms_grad.c

It might be useful to do position correction on multiple object layers in multi-slice ptychography.
One way is to calculate position updates using each object layer and then "average" them at the end of an iteration.

Add wrapper functions and an example script for an automatic parameter tuning workflow based on Bayesian optimization with Gaussian processes.
These functions can be used with external python libraries and enable automatic optimization of reconstruction and experimental parameters in electron/X-ray ptychography. For more details, see https://arxiv.org/abs/2204.11815.

the matrix for storing bad pixels, fmask is always the same size as all the diffraction patterns, 3D. For data as large as 256^4, the fmask can use memory (GPU!) up to 16 GB! For most of the data, fmask may not change for different diffraction, then fmask should only be a single 2D matrix instead of 3D. Even if fmask is pattern dependent, bad pixels are only a small fraction. Using the indexes of bad pixels instead of a 3D matrix with many ones can save a lot of memory.

it seems some diffraction patterns are never used if group size is too large.

https://www.osapublishing.org/oe/fulltext.cfm?uri=oe-26-10-12585&id=386092

Michal's current implementation has very limited use. Need to improve the GPU engine for applications at the APS.

Checkpoints:
Scan positions

• Linear trajectory: perfectly rectangular grid
• Linear trajectory: rectangular grid w. missing scan points
• Linear trajectory: rotated rectangular grid
• Continuous trajectory
• Load external positions

Algorithm:

• Probe initialization
• Mixed-state + arbitrary-path fly-scan
• Constant vs. varying probe intensity
• Position correction
• Variable probe correction
• Online FSC estimation

all modes show as probe 1

The online FSC estimation code in +engines/+GPU_MS/+analysis/ does not work because the engine no longer supports multiple-scan reconstruction.
Need to use external scripts to calculate FSC score for multi-slice reconstructions.

Functions worth checking out:
ptycho/+core/+analysis/calc_FSC.m and aligned_FSC_template.m