Relational matcher - #5
Merged
Merged
Conversation
Sliding-window tiling fragments small label boxes at tile boundaries and runs the detector away from its 1280px training resolution. On real_test (100 pages), full-image @1280 raises label recall 53%->80%, structure recall 93%->99%, and cuts label false-positives ~5x vs tiling. detect_full now defaults imgsz=1280; add scripts/finetune/lps/diag_label_recall.py (conf x mode x class recall sweep + miss characterization). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
New geometry-only matcher (SetMatcher, ~105K params): transformer self- attention over all page detections + log-space Sinkhorn optimal transport with learnable dustbins for unmatched structures (SuperGlue-style), adapted to structure-label pairing. A third BaseMatcher alongside HungarianMatcher and the visual LPS. On real_test it matches distance on clean-box assignment (99.2%), has the best rejection of unlabelled structures (100%), and beats the LPS end-to-end (F1 0.804 det-trained vs 0.772). Now the ChemPipeline default (lazy import avoids relmatch<->pipeline circular import). - structflo/cser/relmatch/: features, model, dataset (GT + DetMatch), matcher, train - sf-train-relmatch entry point; cser-relmatcher in weights registry (v0.1, HF Hub) - scripts/finetune/relmatch/: eval_compare(_all), prepare_det_data, sweep_margin - publish_weights.py: cser-relmatcher support + create_repo, --filename, --no-registry, --yes - ChemPipeline also threads imgsz=1280 to full-image detection Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
CLAUDE.md: relational matcher documented as the pipeline default in the Matching strategies section. 01-quickstart notebook: uses the default matcher and adds a 3-matcher comparison section, using a synthetic example image only. (The reframed paper drafts under docs/*draft*.md are updated on disk but remain untracked per the repo's existing .gitignore rule.) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… score diag
diag_e2e_decompose.py: add --mode {tiled,full640,full1280,full2048} to compare
inference strategies end-to-end. diag_lps_scores.py: dump LPS score
distributions on GT boxes (threshold-miss vs ranking-error decomposition).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
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
1 participant
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.
No description provided.