Skip to content

feat(ipred): composable feature modules, classifier pipeline, and layered overlays - #2

Closed
phzwart wants to merge 30 commits into
mlexchange:mainfrom
phzwart:feat/ipred-composition-layers
Closed

feat(ipred): composable feature modules, classifier pipeline, and layered overlays#2
phzwart wants to merge 30 commits into
mlexchange:mainfrom
phzwart:feat/ipred-composition-layers

Conversation

@phzwart

@phzwart phzwart commented Jul 15, 2026

Copy link
Copy Markdown
Collaborator

Summary

  • Adds the ipred FastAPI compute service with modular feature compositions (skimage / CLAHE / SlimSAM / TomoJEPA / PCA), preprocess cache, and CatBoost + Mondrian conformal train/predict.
  • Hub workflow updates: Ipred composition window, Preprocess / Draw / Train stages, persistent Layers panel so image, features, viridis probability maps, predictions, and annotations can be toggled independently (including per-class).
  • Annotate backend proxies ipred; TomoJEPA weights/ONNX stay gitignored. Scaffold label-set / model-shelf UI is marked # REMOVE THIS AND USE YOUR OWN STUFF.

Test plan

  • start_all.sh brings up frontend, annotate (:8002), ipred (:8003), Tiled
  • Ipred: build/save a composition; health shows modules ready
  • Preprocess: compute feature bank; browse channels; no class manager
  • Draw: annotate labels; Train: train → predict; viridis proba overlay clips at threshold
  • Layers panel: toggle Image / Probability / Predictions / Annotations independently
  • pytest ipred/tests and frontend vitest for composition / layers / pixelClf helpers

Made with Cursor

davramov and others added 30 commits June 24, 2026 09:11
…ectly to Annotate page when a user uploads a dataset
…make it easier to not paint over an already annotated region with another class
…a new class, you can optionally prevent overlaps and automatically cut out holes to avoid two annotations for the same region
Add the ipred FastAPI engine with modular feature compositions (skimage,
CLAHE, SlimSAM, TomoJEPA, PCA), CatBoost conformal train/predict, and
viridis probability overlays with a persistent canvas Layers panel so
image, features, predictions, and annotations can be toggled together.

Co-authored-by: Cursor <cursoragent@cursor.com>
@phzwart

phzwart commented Jul 15, 2026

Copy link
Copy Markdown
Collaborator Author

Closing: keeping this work as a branch on phzwart/forge_Segementation_Finch, not submitting upstream as a PR.

@phzwart phzwart closed this Jul 15, 2026
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