Skip to content

Add slmsuite Hologram algorithm for BeamShaping - #3

Merged
seb5g merged 3 commits into
5.3.xfrom
feature/slmsuite-algorithms
Oct 25, 2026
Merged

seb5g merged 3 commits into
5.3.xfrom
feature/slmsuite-algorithms

Conversation

@seb5g

@seb5g seb5g commented Oct 25, 2026

Copy link
Copy Markdown
Contributor

Summary

  • New SlmsuiteHologram algorithm (GS, WGS-Leonardo/Kim/Nogrette/Wu/tanh) wrapping slmsuite.holography.algorithms.Hologram, usable in the BeamShaping extension. The optimizer state is kept from step to step and restarts when the input, target or initial phase changes. CG is not exposed yet.
  • Discovered by BeamShaping through the pymodaq_plugins_beam_shaping.algorithms entry point, added in hatch_build.py after the template hook (the template generates entry-points, so it can't be declared statically in pyproject.toml; this deviates from the "leave as is" part of the template).
  • beamshaping extra, README section, tests (methods vs slmsuite defaults, convergence on a 4-spot target for each method, restart behaviour, entry-point discovery).

Stacked on #2 (base is the phase 1 branch). Needs PyMoDAQ/pymodaq_plugins_beam_shaping#57 (entry-point discovery) to be found at runtime.

Tested with 42 passed locally (simulated SLM, offscreen Qt).


Generated by Claude Code

Registered through the pymodaq_plugins_beam_shaping.algorithms entry point
(added by the metadata hook), with a beamshaping extra and tests.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CAhZwmbyCSpSgfTyicUxuS
Base automatically changed from feature/slmsuite-daq-move to 5.3.x October 25, 2026 21:57
@seb5g
seb5g marked this pull request as ready for review October 25, 2026 21:58
@seb5g
seb5g merged commit 7234677 into 5.3.x Oct 25, 2026
8 checks passed
@seb5g
seb5g deleted the feature/slmsuite-algorithms branch October 25, 2026 22:00
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants