Skip to content

Edge-case image RI: target-accepted batching, scale switch, config guides, full-scale outputs - #77

Merged
arvinAM17 merged 8 commits into
mainfrom
docs/edge-case-image-generation-polish
Sep 29, 2026
Merged

Edge-case image RI: target-accepted batching, scale switch, config guides, full-scale outputs#77
arvinAM17 merged 8 commits into
mainfrom
docs/edge-case-image-generation-polish

Conversation

@arvinAM17

@arvinAM17 arvinAM17 commented Sep 28, 2026 •
edited
Loading

Copy link
Copy Markdown
Collaborator

Summary

  • NB2 generates until each class reaches TARGET_ACCEPTED accepted images, with a per-class MAX_ATTEMPTS budget and 10% oversampling. It uses unused scenes before reusing old ones, and reports honestly when a target isn't reached.
  • scale in configs/config.yaml shrinks the NB0 download and the NB2/NB3 counts together (0.2 = bootcamp small mode).
  • Short "where to change it" notes throughout the notebooks for bringing your own dataset or local images, anomalies, editor, detector, and judge/API, plus a config map in the README.
  • Full-scale outputs: 1110 images, 200 accepted synthetic images in about 157 minutes. On the held-out real test, traffic-cone AP50 goes 0.245 → 0.321 at +50% synth and 0.253 at +100%.
    Clickup Ticket(s): Link(s) if applicable.

Type of Change

  • 🐛 Bug fix (non-breaking change that fixes an issue)
  • ✨ New feature (non-breaking change that adds functionality)
  • 💥 Breaking change (fix or feature that would cause existing functionality to not work as expected)
  • 📝 Documentation update
  • 🔧 Refactoring (no functional changes)
  • ⚡ Performance improvement
  • 🧪 Test improvements
  • 🔒 Security fix

Changes Made

  • NB2 generates until each class reaches TARGET_ACCEPTED accepted images, with a per-class MAX_ATTEMPTS budget and 10% oversampling. It uses unused scenes before reusing old ones, and reports honestly when a target isn't reached.
  • scale in configs/config.yaml shrinks the NB0 download and the NB2/NB3 counts together (0.2 = bootcamp small mode).
  • Short "where to change it" notes throughout the notebooks for bringing your own dataset or local images, anomalies, editor, detector, and judge/API, plus a config map in the README.
  • Full-scale outputs: 1110 images, 200 accepted synthetic images in about 157 minutes. On the held-out real test, traffic-cone AP50 goes 0.245 → 0.321 at +50% synth and 0.253 at +100%.
    Clickup Ticket(s): Link(s) if applicable.

Testing

  • Tests pass locally (uv run pytest tests/)
  • Type checking passes (uv run mypy <src_dir>)
  • Linting passes (uv run ruff check src_dir/)
  • Manual testing performed (describe below)

Manual testing details:

Screenshots/Recordings

Related Issues

Deployment Notes

Checklist

  • Code follows the project's style guidelines
  • Self-review of code completed
  • Documentation updated (if applicable)
  • No sensitive information (API keys, credentials) exposed

Arvin Azarmina and others added 8 commits September 23, 2026 13:26
Arvin Azarmina @cursoragent Arvin Azarmina
Edge-case image RI: target accepted images per class in NB2, fidelity…
2390e60 
…/novelty plot, docs polish

- NB2 batch now aims for TARGET_ACCEPTED accepted images per class with a
  per-class MAX_ATTEMPTS budget; classes below target revisit their seed pool
  with new variations, and the run reports classes that miss the target.
- Record every judged edit (checkpoint/judged.jsonl) and add a
  fidelity-vs-novelty plot on the CLIP embedding proxies.
- Batch summary (target, accepted, attempts, retries, rejects, rate, elapsed).
- README conceptual overview; NB0.5 callout; NB1 multi-signal wording.
- CLI: --target-accepted / --max-attempts.

Co-authored-by: Cursor <cursoragent@cursor.com>
Arvin Azarmina @cursoragent Arvin Azarmina
NB2: draw fresh scenes before reusing, 10% oversample, larger scene pool
7526b44 
- Seed pools are drawn by cursor: first round target x (1 + oversample)
  fresh scenes; later rounds (1 + oversample) x deficit / rate, filled by
  retries then unused scenes, reusing scenes only when the pool runs out.
- Summary reports scenes new/reused/pool per class; resume restores
  in-flight items instead of re-queuing the whole pool.
- extract_max_generic 320 -> 1000 with an NB0 top-up (only missing scenes
  are downloaded); NB2 pools 300 scenes per class.
- pick_synth_seeds shrinks pools to the clean scenes available instead of
  seeding from scenes that already contain a rare object.
- CLI: --oversample, --n-synth-seeds default 300.

Co-authored-by: Cursor <cursoragent@cursor.com>
Arvin Azarmina @cursoragent Arvin Azarmina
Replace fidelity-vs-novelty scatter with embedding histograms and CLI…
887f5f9 
…P PCA

The scatter collapsed onto the diagonal because the neighbor bank includes
the real images (neighbor_sim >= real_sim). Show per-class histograms of each
gate signal with its threshold, plus a 2-D PCA of real vs judged CLIP vectors.
Judged-edit and real-bank vectors are now persisted in the checkpoint.

Co-authored-by: Cursor <cursoragent@cursor.com>
Arvin Azarmina @cursoragent Arvin Azarmina
Add workshop scale switch and config customization guides
b34d560 
- configs/config.yaml: scale multiplies the NB0 Mapillary download and the
  NB2/NB3 counts (0.2 = bootcamp small mode); scaled() helper in config.
- Notebooks: short 'where to change it' notes above config-driven cells
  (dataset/local images, anomalies, editor, detector, judge/API, NB3 knobs).
- NB2: drop the embedding plot cells for now (library code kept).

Co-authored-by: Cursor <cursoragent@cursor.com>
@arvinAM17
Add updated outputs to the notebooks.
40efdde
Arvin Azarmina @cursoragent
Notebooks: measured runtimes and full-run results note for publishing
47eca29 
- NB2: replace runtime estimate with the measured full-scale run (484 edits, ~157 min).
- NB3: runtime note for 200 synth; wrap-up states what the committed run shows.
- NB0.5/NB1: tidy runtime callouts; drop internal 'slice B/C' labels.

Co-authored-by: Cursor <cursoragent@cursor.com>
Arvin Azarmina
Merge remote-tracking branch 'origin/main' into docs/edge-case-image-…
d011251 
…generation-polish
Arvin Azarmina @cursoragent
Keep cv2.resize results typed under numpy 2.2 stubs from main's lock
8f62718 
Co-authored-by: Cursor <cursoragent@cursor.com>
@arvinAM17
arvinAM17 requested a review from fatemetkl September 28, 2026 03:35
@coderabbitai

coderabbitai Bot commented Sep 28, 2026

Copy link
Copy Markdown

Important

  • 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Repository: VectorInstitute/synthetic-data-bootcamp/.coderabbit.yaml

Review profile: CHILL

Plan: Advanced

Run ID: a5f375e3-7b7a-4d4f-9382-5a23221c7056

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

fatemetkl
fatemetkl approved these changes Sep 28, 2026
View reviewed changes

@fatemetkl fatemetkl left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM! Notebooks run with no issues!

@arvinAM17
arvinAM17 merged commit d9b1a99 into main Sep 29, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@fatemetkl fatemetkl fatemetkl approved these changes

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@arvinAM17 @fatemetkl