Skip to content

NEXT_STEPS.md

Latest commit

 

History

History
59 lines (52 loc) · 2.7 KB

NEXT_STEPS.md

File metadata and controls

59 lines (52 loc) · 2.7 KB
Raw

NEXT_STEPS.md

Last updated: 2026-04-11 MVP Readiness: 65%

Done

  • Paper verification + red flag documentation
  • PRD, ASSETS.md, PIPELINE_MAP.md, architecture docs
  • CAWM-Mamba backbone: WAPM, CFIM, WSSB, FreqSSM, CDSM modules
  • Dual backend device layer (MLX/CUDA/CPU auto-detect)
  • CLI with smoke/parity/train-step commands
  • IR-Visible paired dataset loader + synthetic fallback
  • Train/val/test split (90/5/5) with reproducible indices
  • Paired augmentations preserving spatial alignment
  • Full training loop: cosine warmup LR, early stopping, checkpointing
  • bf16 mixed precision, gradient clipping, TensorBoard logging
  • nohup+disown launch script for GPU server
  • Image quality metrics: MAE, PSNR, SSIM
  • Export pipeline: pth -> safetensors -> ONNX
  • Training report generator
  • Dockerfile.cuda + Dockerfile.mlx + docker-compose.yml
  • HTTP serve.py with /health, /predict, /info
  • anima_module.yaml manifest
  • Comprehensive test suite (data, metrics, export, config, serve)
  • GPU training configs (train_cuda.toml, debug.toml)

In Progress

  • Code review pass (pending)

TODO — Requires GPU (deferred: NIGHTHAWK mega dataset building on GPUs 2-7)

  • Acquire/locate IR-Visible UAV dataset pairs on server
  • Set up module .venv on GPU server (uv venv + uv sync)
  • Run /gpu-batch-finder on L4 for optimal batch size
  • Full training run (GPU 0 or 1 only during NIGHTHAWK)
  • Test set evaluation with SSIM/PSNR/MAE report
  • TensorRT FP16 + FP32 exports from best checkpoint
  • Push trained model to HuggingFace

TODO — Integration (post-training)

  • ROS2 fusion node implementation
  • Gazebo simulator integration
  • YOLO26 downstream detection evaluation
  • Docker build + health check
  • Demo recording for Shenzhen Robot Fair

Blocking

  • GPUs 2-7 occupied: NIGHTHAWK HiRes mega dataset generating. Use GPU 0 or 1 only.
  • AWMM-100K unavailable: Paper benchmark dataset not directly downloadable. Proceed with UAV datasets.
  • Upstream code gaps: ori_version_526_version missing from reference repo. Using clean-room implementation.

Downloads Needed

  • IR-Visible UAV pair dataset — check /mnt/train-data/datasets/ for DroneVehicle or similar
  • YOLO26m weights — ultralytics hub or shared models dir

Session Log

Date Agent What happened Next
2026-04-10 Research Agent Project scaffolded, docs created Verify paper
2026-04-10 WEIMARANER Codex Verification, PRD, initial scaffold Data + train
2026-04-11 WEIMARANER Opus Full build: data pipeline, training loop, metrics, export, Docker, serve, tests Train when GPUs free