Skip to content

TRAINING_REPORT.md

Latest commit

 

History

History
86 lines (77 loc) · 2.44 KB

TRAINING_REPORT.md

File metadata and controls

86 lines (77 loc) · 2.44 KB
Raw

KELPIE — Training Report (template)

This report is populated by anima_kelpie.train after each real training run. The fields below are placeholders — the autopilot will fill them in once GPUs 0/1 free up (blocked by the NIGHTHAWK HiRes mega dataset build).

Run Identity

Field Value
Project project_kelpie
Codename KELPIE
Wave 10 (WARDOG)
Paper Prototype-Based Low Altitude UAV Segmentation (PBSeg)
ArXiv 2604.01550
Run ID TBD
Date TBD
Commit TBD
GPU TBD
Backend TBD

Hyperparameters

Field Value
Backbone resnet50
Hidden dim 256
Num queries 100
PBCA layers 6
Num classes 8
Optimizer AdamW
LR 1e-3
Weight decay 1e-4
Epochs 100
Batch size auto via /gpu-batch-finder
Val ratio 0.10
Seed 42
Gradient clip 1.0

Data Splits

Split Samples Source Notes
train TBD UAVid + UDD6 80%
val TBD UAVid + UDD6 10%
test TBD UAVid + UDD6 10% held out

Loss Composition

Component Weight Final Value
CE semantic 1.0 TBD
Soft Dice 1.0 TBD
Per-class BCE 0.5 TBD
Query regularizer 0.25 TBD

Metrics (Test set)

Metric Value
mIoU TBD
mF1 TBD
OA TBD
Per-class IoU TBD

Artifacts

  • Checkpoints: /mnt/artifacts-datai/checkpoints/project_kelpie/
  • Logs: /mnt/artifacts-datai/logs/project_kelpie/
  • TensorBoard: /mnt/artifacts-datai/tensorboard/project_kelpie/
  • Exports: /mnt/artifacts-datai/exports/project_kelpie/
    • kelpie_pbseg.safetensors
    • kelpie_pbseg.onnx
    • kelpie_pbseg_fp16.plan (or trt_build_fp16.sh)
    • kelpie_pbseg_fp32.plan (or trt_build_fp32.sh)
  • HuggingFace: ilessio-aiflowlab/project_kelpie (private)

Reproduction

source .venv/bin/activate
nohup python -m anima_kelpie.train \
  --config configs/default.toml \
  --backend cuda \
  --epochs 100 \
  > /mnt/artifacts-datai/logs/project_kelpie/train_$(date +%Y%m%d_%H%M).log 2>&1 &
disown

Known Blockers

  • Upstream PBSeg repo is empty — reimplemented from paper (see prds/05_cto_review.md).
  • UAVid + UDD6 need to be mirrored into /mnt/forge-data/shared_infra/datasets/.
  • Training queued behind NIGHTHAWK HiRes mega dataset build (GPUs 2-7).