Canvas mask selection, multi-delete & visible selection styling (#75)#77
Closed
cofade wants to merge 27 commits into
Closed
Canvas mask selection, multi-delete & visible selection styling (#75)#77cofade wants to merge 27 commits into
cofade wants to merge 27 commits into
Conversation
Records the triaged open upstream issues (validated 2026-06-12) with size classification. Each row is deleted in the PR that resolves it; the whole section is removed once empty. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Captures the method of writing triaged work items into CLAUDE.md as a temporary table with a deletion hook: each resolving PR deletes its row, and the section removes itself once empty. Un-ignores .claude/skills/ so project skills are tracked like the senior-reviewer agent. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds a combo above the image list (All / Without annotations / With annotations). Rows are hidden via setRowHidden — never removed — so row-index iteration (DINO navigation, COCO import) stays valid and no spurious currentRowChanged/switch_image fires. The current row is never hidden. Re-apply hooks into ClassController.update_slice_list_colors, which every annotation-mutation site already calls. Multi-dim images count as annotated when any of their slices has annotations. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
bnsreenu#57) torch.cuda.is_available() is True for Pascal GPUs (e.g. GTX 1050, sm_61) even though torch>=2.8 wheels ship sm_70+ kernels only, so every launch died with 'CUDA error: no kernel image is available'. New core/torch_utils.resolve_torch_device() compares the device capability against torch.cuda.get_arch_list() and returns cpu + an actionable warning on mismatch (cached process-wide). SAM passes device= on every predict, DINO's _resolve_device delegates to the helper (DINO_DEVICE override kept), YOLO train/predict pass device=. One-time QMessageBox via maybe_warn_cpu_fallback at SAM model pick and DINO detect. README gains a Pascal/Maxwell note with the torch-downgrade command. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds the device-fallback concept and the hide-don't-remove image-list rule to crosscutting concepts, registers core/torch_utils and the filter responsibility in the building block view, and removes the #27 and bnsreenu#57 rows from the CLAUDE.md backlog per its deletion hook. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Comment/doc now state the real re-apply contract (direct update_slice_list_colors call OR annotationsBatchSaved route) instead of overclaiming that every mutation site calls it directly. - apply_image_filter early-outs on the default 'All images' mode so the per-refresh annotation scan only runs with an active filter. - Integration test exercising the real commit path (annotationsBatchSaved -> save -> color refresh -> filter re-apply). - torch_utils: document the deliberate device-0-only probe; note the _8bit-key caveat in the prefix fallback; register core.torch_utils in the import smoke list. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Manual testing feedback: under an active filter, a selected image that doesn't match (e.g. unannotated under 'With annotations') stayed in the list until the user navigated away. Drop the never-hide-current-row exemption — the current row is now hidden like any other non-match. Hiding doesn't touch current_image or fire currentRowChanged, so the canvas keeps showing the worked-on image while its row leaves the list, which is the requested behavior. New integration test asserts the canvas-unchanged / no-switch_image guarantee; unit test updated to the new expectation; arc42 updated. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…nd-pascal-gpu-fallback feat: Image-list annotation filter (bnsreenu#27) + CPU fallback for unsupported CUDA GPUs (bnsreenu#57)
model.export() converts to a deployment format and crashes when given a .pt save path; model.save() writes the checkpoint as intended. Reporter- verified fix. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…nsreenu#44) On Windows, open() defaults to the cp1252 code page, so any YAML/JSON/TXT containing non-ASCII (e.g. a unicode class name, or a non-ASCII path in an .iap project) crashed with 'charmap codec can't decode/encode'. Add encoding='utf-8' to every text-mode open across YAML (yolo_trainer, import/export_formats, dataset_splitter), JSON (project load/save, COCO import/export, augmenter, combiner, dino merge, project search), and YOLO label .txt writes. PIL Image.open (binary) left untouched. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…u#33) start_polygon_edit returned the first polygon containing the double-click, so an annotation fully nested inside another could never be edited. Scan all containing polygons and pick the smallest by area (reusing the existing shoelace calculate_area from utils.py). bbox-only annotations remain out of scope (that's bnsreenu#40). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New ImageController.sort_image_list orders all_images and the list widget together (case-insensitive), keeping the all_images[i] <-> image_list.item(i) invariant that COCO import relies on. Rebuilds with signals blocked rather than setSortingEnabled, since currentRowChanged is wired to switch_image and a live re-sort would fire spurious image switches. Wired into add_images_to_list (selects/switches to the first added image), update_image_list, and the COCO import path. Skipped per image during project load to avoid O(n^2) re-sorts. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…enu#56) Reading an LZW-compressed TIFF without the optional imagecodecs package raised ValueError and crashed the app. Add imagecodecs to install_requires (fixes new installs) and catch the codec ValueError in add_images_to_list: show an actionable 'pip install imagecodecs' dialog and skip the file instead of crashing or leaving a half-added entry. Non-codec ValueErrors still propagate. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ard test, execute backlog hook Adds the image-list sort and TIFF-codec concepts to crosscutting docs, updates the ImageController building-block row, adds a UTF-8 round-trip regression test for bnsreenu#44, and removes the five resolved quick-win rows (bnsreenu#30/bnsreenu#44/bnsreenu#33/bnsreenu#60/bnsreenu#56) from the CLAUDE.md backlog per its deletion hook. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Removes unused/duplicate imports, placeholder-less f-strings, and one unused local in files modified by this PR, so ruff check --select F,E9 passes cleanly. No behavior change; full suite still green (156 passed). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- _is_missing_codec_error: drop the over-broad 'compression' substring arm that could swallow unrelated ValueErrors behind a misleading 'install imagecodecs' dialog; match only the reliable 'imagecodecs' token (bnsreenu#56). - sort_image_list: use info.get('file_name','') so the sort key is total even if an entry ever lacks file_name (bnsreenu#60). - image_label: hoist 'from ..utils import calculate_area' to module scope per ADR-016 (no inline imports that rot after module moves). - Add test_project_load_path_ends_sorted pinning the load -> update_ui rebuild-sorts contract. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…nsreenu#56) Locks in the P1 narrowing from the prior commit so a future broadening of _is_missing_codec_error is caught. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Five upstream quick wins: bnsreenu#30, bnsreenu#44, bnsreenu#33, bnsreenu#60, bnsreenu#56
Ultralytics ships no SAM trainer (SAM.train() -> NotImplementedError), so
this adds a custom decoder (optionally encoder) fine-tuning loop that reuses
Ultralytics' own SAM2Predictor forward methods (get_im_features /
prompt_inference) under autograd, with focal+dice loss + AdamW. Fine-tuned
checkpoints save as {"model": state_dict} and reload through the unchanged
SAM(path) inference path, appearing in the SAM model selector.
- training/sam_trainer.py: SAMFineTuner engine + geometry/loss/naming helpers
- training/sam_dataset.py + io/export_formats.export_sam_dataset: project and
prepared-folder dataset producers
- dialogs/sam_trainer_dialog.py: config dialog (reuses TrainingInfoDialog)
- controllers/sam_train_controller.py: menu, GPU gate, SAMTrainingThread,
selector registration; wired in annotator_window
- inference/sam_utils.py: custom-model registry so fine-tuned models load by path
- tests/integration/test_sam_finetuning.py: geometry/loss/dataset/API-drift
guards + opt-in end-to-end (SAM_TRAIN_E2E=1)
- docs: ADR-021, building-block, runtime, glossary; CLAUDE.md structure
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Encoder path: backprop once per image (sum instance losses) so the shared encoder feature graph isn't freed mid-loop; batch_size now accumulates over images. Fixes "backward through the graph a second time" crash on any image with >1 instance when fine-tuning the image encoder. - Concurrency: lock SAM inference UI (tools, model selector, fine-tune menu) for the duration of a run; restore in training_finished on success AND error. The trainer uses its OWN SAM instance — corrected the false "drives the resident model" narrative in code docstrings, ADR-021, and runtime view; the real hazard is two models on one CUDA context. - Threading: convert in-memory slice QImages to numpy on the GUI thread before handing to the worker (matches _qimage_to_numpy cross-thread contract). - Harden export_sam_dataset path components with os.path.basename. - Add opt-in encoder-path multi-instance regression test. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- _launch wraps setup-through-start() in try/except so a failure before the worker thread starts re-enables the SAM inference UI (otherwise tools stay disabled until restart — training_finished was the only other unlock site). - export_sam_dataset slice lookup uses `is not None` (consistency with sam_dataset.py). - Add controller tests: UI lock toggle + lock-recovery on setup failure. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The Unicode → in the progress print raises UnicodeEncodeError on a Windows cp1252 console. Use -> so a console-launched run can't crash on the export log. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Found during real GUI testing on non-square images. - Mask shift (correctness): training loss mapped decoder logits to the image with a naive F.interpolate, but SAM2 letterboxes (pad bottom/right) and inference crops that padding via ops.scale_masks(padding=False). The decoder learned padding-shifted targets -> masks shifted down on non-square images. Fix: run logits through the SAME ops.scale_masks(padding=False) inference uses. Validated on a landscape image: fine-tuned mask centroid dY=-0.1px, IoU=0.994 (was visibly shifted). Square test images had hidden the bug. - Progress dialog: clear the reused log on launch; disable Stop on completion so a post-finish click can't strand it on "Stopping...". SAM instance only; shared TrainingInfoDialog class untouched (YOLO unaffected). - Device label: show GPU name (cuda:0 read as "not detected"). - Light mode: drop inline color:palette(text) on the config-dialog note and the pre-existing DINO sidebar caption (No Hardcoded Colors Rule). - Tests: ops.scale_masks padding=False crop-geometry guard + opt-in landscape shift regression. ADR-021 coordinate-frame consequence documented. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The loaded canvas + updated window title already make a successful open obvious; the modal was just an extra click. Error dialogs unchanged. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
feat: Fine-tune SAM 2 / 2.1 on user annotations (bnsreenu#73)
The ClassThresholdTable header styled itself inline with background-color: palette(mid); color: palette(text). Those palette() references resolve against the OS palette (dark on Windows), so in the app's light mode the header painted dark-on-light, out of step with the rest of the UI. Light mode (default_stylesheet) also had no QHeaderView::section rule at all, while dark mode did. Strip the inline palette() colours (keep only structural font-weight / padding) and add a light QHeaderView::section rule to default_stylesheet, symmetric with the dark sheet. Header colour now follows the active theme. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…bnsreenu#75) Give the canvas the same selection power the annotation list already had, and make selection legible. Selection (idle mode — no drawing/SAM tool active): - single click selects the smallest mask under the cursor (segmentation or bbox); click on empty space clears - Shift+click toggles a mask; drag draws a rubber band that box-selects; Shift+drag adds. Esc cancels an in-progress band. Ctrl stays pan. - Delete removes the selected mask(s). Canvas selection is unified with the annotation list via AnnotationController.apply_canvas_selection, which mirrors onto the QListWidget (signals blocked) and matches by dict value-equality; Delete/Merge/Change Class are reused unchanged. Canvas Delete gates on the list selection (CanvasContext) so a stale highlight can't pop a spurious warning. Selection rendering — no longer recolours the mask red (invisible on a red-class mask). The mask keeps its class colour; a class-colour- independent overlay is drawn last: a dashed selection-blue bounding box plus bright handle squares at corners/edge-midpoints (open-garden-planner CAD style; handles are visual-only, resize is upstream bnsreenu#40). Class colours: default palette no longer starts on red — core/constants default_class_color (curated tab10-style, red last, muted) replaces the Qt.GlobalColor(n % 16 + 7) formula at the 4 assignment sites. Default fill opacity 0.3 -> 0.2 so masks don't bury the image. Saved projects keep their persisted colours. Tests: unit (hit-testing, gesture resolution, Esc-cancel, render-no- recolor, palette) + integration (controller selection mirror, canvas- delete gating). arc42: ADR-022 (+amendment), runtime, crosscutting, glossary, CLAUDE.md. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This was referenced Jun 21, 2026
Owner
|
Conflicts |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #75.
Addresses #75 — "single click should select a mask … draw a selection box around several … pick several with a modifier … delete several at once."
What this adds (the subject of #75)
Selection on the canvas (idle mode — only when no drawing/SAM tool is active):
The canvas selection is unified with the existing annotation-list selection, so Delete / Merge / Change Class work identically from either surface.
Visible, class-colour-independent selection styling. Selected masks are no longer recoloured red (which was invisible on a red-class mask — and red was the default first class colour). The mask keeps its class colour; selection is drawn as a dashed blue bounding box + handle squares at the corners/edge-midpoints. The default class palette no longer starts on red, and the mask fill opacity is lowered so the image stays legible.
Validation
This is a cross-fork PR from
cofade, and because the fork has diverged substantially from upstream (it has been refactored into per-controller modules, plus SAM fine-tuning and other features), the diff is not just the #75 change — it stacks on the fork's prior work and therefore overlaps the open SAM PR #76. Cherry-picking only the #75 commits onto a clean upstream branch isn't feasible against the current upstream layout.Bundled commits include: image-list status filter (#27), CPU compute-capability fallback (#57), quick-wins (#30/#33/#44/#56/#60), and SAM 2/2.1 fine-tuning (#73, also in #76), then the #75 canvas-selection work on top.
If you'd prefer to take #75 in isolation, the cleanest path is to merge #76 first (or a from-scratch reimplementation against upstream's monolithic
ImageLabel) — happy to help either way. The canvas-selection commit itself isfeat: Canvas mask selection, multi-delete & visible selection styling.