Skip to content

NEXT_STEPS.md

Latest commit

 

History

History
81 lines (67 loc) · 4.62 KB

NEXT_STEPS.md

File metadata and controls

81 lines (67 loc) · 4.62 KB
Raw

KELPIE — Execution Ledger

Last updated: 2026-04-11 (autopilot round 2) MVP Readiness: 78%

Resume rule: Read this file completely before writing code. This project covers exactly one paper: Prototype-Based Low Altitude UAV Semantic Segmentation (PBSeg).

1. Working Rules

  • Work only inside project_kelpie/
  • Prefix every commit with [KELPIE]
  • Stage only project_kelpie/ files
  • Training is blocked — GPUs 2-7 are running NIGHTHAWK HiRes mega dataset build
  • When training resumes, use GPU 0 or 1 only

2. Paper

3. Done this session (2026-04-11, full-build autopilot)

  • PRD-02 — MaskClassificationLoss (CE+Dice+BCE+cls), SegmentationMeter, HistoryLogger
  • PRD-03 — Real UAV dataset adapters: UAVid (8-class RGB palette), UDD6 (gray), VisDrone (binary vehicle merge)
  • PRD-04 — Export pipeline: safetensors + ONNX opset 17 (deform-disabled) + TRT fp16/fp32 build scripts
  • PRD-05 — FastAPI serve endpoint + anima_module.yaml + docker/Dockerfile.serve + docker-compose.serve.yml
  • PRD-06 — Hero page (assets/hero.html + assets/hero.png 1280x640), README rewrite, TRAINING_REPORT.md template
  • Tests: 26/26 passing (added 20 new)
  • CLI now includes kelpie export subcommand
  • pyproject.toml: new [export] and [serve] extras

4. Verification Snapshot

  • Read paper completely
  • Check reference repo existence (empty)
  • Check reported datasets availability (partial — VisDrone only)
  • Check claimed metrics plausibility
  • Check independent reproductions/citations (none yet)
  • Document red flags and prepare CTO review
  • Phase-1 code scaffold + Phase-2 hardening (losses, metrics, adapters)

Red Flags

  • Repo contains no implementation code (single README-only commit).
  • Exact reproduction from author code currently impossible.

5. Datasets

Shared volume: /Volumes/AIFlowDev/RobotFlowLabs/datasets/

Found:

  • wave10_staging/visdrone
  • repos/DepthPrior/datasets/visdrone

Not found in current scan:

  • uavid, udd6, uavdt, dronevehicle, seadronessee

Adapters are already coded for all of the above (src/anima_kelpie/data/uav_datasets.py) — they just need the data to arrive.

6. TODO (in priority order)

  • Acquire/normalize UAVid and UDD6 for reproduction phase
  • Wait for NIGHTHAWK mega dataset build to finish (GPUs 2-7)
  • Start Phase 2 reproduction runs on GPU 0/1 once free
  • Run /gpu-batch-finder on KELPIE before first real training launch
  • Fill in TRAINING_REPORT.md after first full run
  • Push safetensors + ONNX to ilessio-aiflowlab/project_kelpie after training
  • Add Hungarian matcher (currently using flat query loss as Phase-1 approx)
  • Wire YOLO26 detection-guided inference once vehicle detection is stable

7. Blocking

  • GPUs 2-7 occupied by NIGHTHAWK HiRes mega dataset build (ETA mid next week).
  • UAVid + UDD6 datasets not mirrored on GPU server yet.

8. Downloads Needed

  • UAVid — ~30GB — huggingface-cli download dvlab-research/UAVid --local-dir /mnt/forge-data/shared_infra/datasets/uavid (or official site)
  • UDD6 — ~3GB — mirror from github.com/MarcWong/UDD
  • YOLO26 checkpoint — ~50MB — ultralytics fetch or HF mirror

9. Session Log

Date Agent What happened Next
2026-04-10 ANIMA Research Agent Project scaffolded with initial docs Read paper and verify
2026-04-10 KELPIE agent Completed verification; found major red flag (empty upstream repo); authored full PRD/task docs; implemented Phase 1 PBSeg scaffold; validated smoke/train/eval scripts and pytest Acquire UAVid/UDD6 and start controlled reproduction runs
2026-04-11 KELPIE agent (autopilot) Built full solution without training: losses + metrics + history, UAVid/UDD6/VisDrone adapters, export pipeline (safetensors/ONNX/TRT), FastAPI serve + manifest + Docker.serve, hero page, 26/26 tests passing Wait for GPUs 0/1 to free up, then train on real data
2026-04-11 KELPIE agent (autopilot r2) 5 small commits (settings attrib, NIGHTHAWK notice, Wave10 data guide + MAP, vendored .claude/rules, hero refresh + PNG regen), ran /code-review, fixed 5 findings (mIoU inflation on absent classes, ephemeral checkpoint path, serve.py pickle-RCE, dead FocalLoss, _resize_np (H,W) assertion). 26/26 tests still pass Push to origin main; training still blocked on NIGHTHAWK