Skip to content

TODO.md

Latest commit

 

History

History
68 lines (49 loc) · 3.92 KB

TODO.md

File metadata and controls

68 lines (49 loc) · 3.92 KB
Raw

TODO

Standalone list of known remaining work. This file is intentionally not linked from the main README or workflow docs.

Preprocessing

  • Audit src/preprocessing/extract_crops_and_coords_LR.py and shared contour cleanup in src/preprocessing/generate_masks_utils.py with known failing LR slices.
  • Verify remaining LR preprocessing edge cases not covered by the current polygon repair, including bbox/image-bound assumptions, crop/GeoJSON frame consistency, degenerate surface intersections, and multiprocessing logging.

Documentation And Config Defaults

  • Align default configs and downstream code where they still reference older high_res_aligned or misc/high_res_affines names.
  • Align code and default configs where they still reference old input/wsis, input/masks, output/masks, density_maps, experiments_results, lr_preds, or misc paths.
  • Make all YAML configs support a configurable data root.
  • Finish aligning legacy configs and classification defaults with the canonical data-root layout.
  • Align code/config defaults where they still expect older CA* or R-CA*-adj region names; the maintained docs use RCA1-RCA4.

Segmentation

  • Add a script to run segmentation for all maintained areas and merge the outputs.
  • Update legacy segmentation configs outside experiments/segmentation/allmodels/ to canonical paths.
  • Smoke-test full all-model segmentation with HoverNet enabled because it is GPU-memory sensitive.

Classification

  • Release classification ground-truth annotations or document their project-artifact handoff path.
  • Release training GT and implement the maintained classification training pipeline.
  • Train and release a scikit-learn classifier based on pretrained ResNet18 features if a public non-gated classification path is needed.
  • Handle the released sklearn classifier compatibility warning by pinning scikit-learn==1.7.2 for inference or re-exporting the artifacts.
  • Refactor the classification pipeline.
  • Make classification inference automatic across RCA1, RCA2, RCA3, and RCA4.
  • Align classification training defaults with the canonical folder layout.
  • Adjust the classification dataloader for canonical GT naming and region folder conventions.
  • Add resnet18 classifier release support to the maintained classification training/release workflow.

Density Dataset

  • Revise dataset creation with the new simplified HR-LR mapping.
  • Fix the issue with using merged OOR Daniela's data.

Density Training And Experiments

  • Test whether a fourth total-count ground truth could provide context and/or gradient backpropagation improvements.
  • Re-test NAE-only loss and NAEPixelCount loss.
  • Implement recent SOTA density-estimation losses.
  • Implement unsupervised pretraining, including data-preparation pipeline changes.

LR GT Evaluation

  • Add a script to create output/test_lr_density_gt/test_set_gt_allCA_128_96_smooth_b05_k5_roi/ from the density dataset test split.
  • Run GT evaluation only after the GT preparation script creates data/output/test_lr_density_gt/test_set_gt_allCA_128_96_smooth_b05_k5_roi/.
  • Verify whether using the full LR crop folder as --input-dir is the right default once the GT preparation script defines the exact test subset.
  • Resolve the dependency of data/output/lr_gt_eval/allCA_best_model_128_96_smooth_b05_k5_roi/ on test_lr_density_gt creation.

Point Cloud Reconstruction

  • Add optional rotating GIF and STL/volume export.
  • Add future surface visualization outputs where needed.
  • Add future 3D spinning visualization with class colors where needed.
  • Add future surface mesh output where needed.

Data Layout And Artifacts

  • Finalize segmentation train/test split files under input/train_test_splits/segmentation/.
  • Finalize classification train/test split files under input/train_test_splits/classification/.
  • Finalize input/classification_gt/ layout and contents.
  • Clarify or remove models/original_weights/ for pre-finetuning weights.