v1.3.0

Changes Since v1.2.0

Summary

• Major additions: multi-process support (esp. LSQML), new constraints (object hard limits, probe support mask), probe power constraint behavior,
  diffraction blur, large-tensor offload.
• Several fixes in multiprocessing/LSQML and performance/memory improvements.
• Docs/build/test updates and movie tooling improvements.

Change List (by theme)

Features

1. Multi-process LSQML reconstruction and supporting infrastructure.
2. Hard limits constraint on object magnitude/phase.
3. Probe support constraint with geometric mask (ellipse/rectangle).
4. Diffraction pattern blur option in forward model.
5. Large tensor CPU offload/reload helper on Task.
6. Position affine update flexibility override.
7. Auto-rescaling of DIP outputs, and half precision support.

Fixes

1. Multiple multi-GPU/multiprocess sync issues in LSQML and reconstruction hooks.
2. Correct device assignment for phase unwrap in multiprocessing.
3. Correct FFT normalization in probe power constraint.
4. Fix handling of remove_grid_artifacts direction.
5. Fix multiscan tester circular data handoff.
6. Logging/tqdm edge fixes.

Performance/Memory

1. Uniform batch sampler builds/syncs indices on GPU and rank 0 only.
2. Chunk processing uses expand over repeat.
3. Replicate tensors on-the-fly to reduce memory.
4. DDP instead of DataParallel for AD model.

Docs/Build/Test

1. Added multiprocess docs and task docs, updated index/citations.
2. Added pytest, uv.lock, python pin; adjusted matplotlib dependency for docs build.
3. New tests for multiprocess LSQML, object hard limits, probe support, blur, offload.

Movies/IO

1. Movie functionality extended to record probe positions and settings cleanup.
2. I/O additions and documentation.

API Changes

1. Dtypes enum now includes FLOAT16.
   • File: src/ptychi/api/enums.py
2. New enums: MagPhaseComponents, ProbeSupportMethods.
   • File: src/ptychi/api/enums.py
3. New object constraint option: ObjectHardLimitsMagnitudePhase and ObjectOptions.hard_limits_magnitude_phase.
   • File: src/ptychi/api/options/base.py
4. RemoveGridArtifactsOptions adds component (magnitude/phase/both).
   • File: src/ptychi/api/options/base.py
5. ProbePowerConstraintOptions adds scale_object and clarifies behavior (probe power based on probe itself).
   • File: src/ptychi/api/options/base.py
6. ProbeSupportConstraintOptions adds fixed support mask options and params.
   • File: src/ptychi/api/options/base.py
7. PositionAffineTransformConstraintOptions adds override_update_flexibility with validation.
   • File: src/ptychi/api/options/base.py
8. LBFGS optimizer is now explicitly restricted to Autodiff reconstructors (raises ValueError in multiple option classes).
   • File: src/ptychi/api/options/base.py
9. ForwardModelOptions adds diffraction_pattern_blur_sigma.
   • File: src/ptychi/api/options/base.py
10. Task now mixes in multiprocessing and changes behavior in multi-process mode:
    - Device init uses process group, rank-based device selection.
    - run() now has reset_timer_globals arg.
    - get_data() triggers DIP generate() for object/probe.
    - New set_large_tensor_device() helper.
    - Multi-process selects reconstructor via get_multiprocess_reconstructor_by_enum.
    - Files: src/ptychi/api/task.py, src/ptychi/maps.py, src/ptychi/parallel.py
11. New public class export: MultiprocessLSQMLReconstructor.
    - File: src/ptychi/reconstructors/__init__.py
12. New type alias Numeric in API types.
    - File: src/ptychi/api/types.py

v1.2.0

What's changed

New features

• Allowing using total intensity for probe centering constraint
• PIE now supports multislice
• Added device module and symbol wrapper to support Intel GPUs

API changes

• Users can now export/import settings from/to Options objects as JSON
• OptimizationPlan is now a subclass of Options and is moved to api.options.base
• Removed depreciated Options classes
• OPRModeWeightsSmoothingOptions.method now has a default value

v1.1.0

What's changed

Bug fixes

• Fixed affine transformation matrix fitting.
• Corrected sign in get_max_batch_size.
• Fixed an in-place update in propagators which caused an issue in AD backpropagation.
• Fixed AD's device transfer bug when using multiple GPUs with DataParallel.

Algorithm changes

• Probe update magnitude limit is further clipped by mean absolute deviation of update.

UI changes

• Default setting of image differentiation method for probe position correction was changed to FOURIER_DIFFERENTIATION; step size default was changed to 0.3; update magnitude limit default was changed to 0.1.
• Spelling of ASYMMETRY in affine transformation constraint's degrees of freedom was corrected.
• Added LSQML preconditioning damping factor to API.

Build-related changes

• Added requirements.txt generated by uv.
• Use uv for environment management in CI.

Experimental features

• Synthetic sparse dictionary learning for probe optimization (#33)

v1.0.0

What's Changed

New fetures

• Save movies of arrays by @hruth in #31
• Deep image prior (experimental) by @mdw771 in #32
• Adding the Bilinear Hessian method for reconstruction by @nikitinvv in #27
• Non-square pixel size support
• Slice spacing optimization (AD only)
• Affine transformation constraint on probe positions​
• Allow choosing the slice for probe position correction​

Bug fixes

• Allow choosing alternative ways of defining object-frame pixel coordinates of position origin
• Fixed performance issue in uniform batching
• Fixed performance issue in automatic differentiation

Full Changelog: v0.1.0...v1.0.0

v0.1.0

What's changed

• The reconstruction quality of LSQML is significantly improved on data with low overlap and large illumination variation.
• New difference map reconstructor.
• Reconstruction speed is now 2--3x faster; further speedup can be obtained by setting reconstruction_options.use_double_precision_for_fft to False (less tested; use with caution).
• New profiling utility.
• Added new constraint routines including remove_object_probe_ambiguity and opr_weight_smoothing, and improved multislice_regularization.
• Added documentation.
• Allow switching between faster yet non-deterministic and slower yet deterministic algorithms.

v0.0.1-alpha

Features

Reconstruction engines

• LSQML
• PIE (including ePIE and rPIE)
• Difference map
• Automatic differentiation

Major correction features

• Probe position correction: gradient-based and cross correlation-based
• Multislice (only in LSQML and AD)
• Orthogonal probe relaxation (only in LSQML and AD)
• Mixed-state probe
• Compact batching

Other features

• Multi-GPU (only in AD)