Skip to content

vision-dependency-stack.md

Latest commit

 

History

History
92 lines (71 loc) · 3.03 KB

vision-dependency-stack.md

File metadata and controls

92 lines (71 loc) · 3.03 KB
Raw

TR-HASH Vision dependency stack

The detector uses the same public third-party computer-vision primitives as the Ultralytics runtime where they are relevant to TR-HASH Vision:

  • PyTorch and matching Torchvision wheels;
  • OpenCV, Pillow, NumPy and Matplotlib;
  • Albumentations for detection augmentation;
  • pycocotools and faster-coco-eval for COCO evaluation;
  • psutil, Polars and nvidia-ml-py for runtime measurements;
  • ONNX, ONNX Runtime and onnxslim for portable export.

This is dependency-level interoperability only. The framework does not import or copy Ultralytics model code. ultralytics-platform is intentionally omitted because it is a hosted-platform client, and ultralytics-thop is unnecessary because detector FLOPs are measured with PyTorch's native flop counter.

Backend-safe installation

The installer resolves Torch and Torchvision together from the same backend index. This prevents a ROCm installation from being silently replaced by a CUDA wheel.

make install-vision-cuda
make install-vision-rocm
make install-vision-cpu

For an environment where the correct Torch stack is already installed:

pip install -e '.[detection,export]'

Verify a machine before training or export:

python scripts/check_vision_stack.py --strict

The reference dependency versions are tracked from the public Ultralytics pyproject.toml; backend selection and model implementation remain native to Complexity Framework.

Active detector pipeline

The COCO launchers use the installed primitives directly rather than only declaring them as optional packages:

  • OpenCV decodes random-access COCO and YOLO images;
  • Albumentations applies box-aware geometric and color transforms;
  • faster-coco-eval is selected automatically, with pycocotools as the official reference fallback;
  • distributed validation gathers each non-overlapping COCO shard and executes COCOeval once on rank zero;
  • best and best_nms_free are selected by official COCO mAP50-95.

Official comparable evaluation retains at most 100 detections per image. The framework's internal metric implementation remains available as a diagnostic, but it is no longer used to select or publish a COCO checkpoint.

The specialized COCO launcher enables this stack by default. The equivalent explicit training options are:

--image-backend opencv \
--augmentation-backend albumentations \
--eval-backend auto \
--eval-max-detections 100

Use the standalone evaluator to reproduce published metrics:

python scripts/evaluate_tr_hash_coco.py \
  artifacts/detector_coco_v06_native/best \
  --annotations artifacts/COCO/annotations/instances_val2017.json \
  --images artifacts/COCO/val2017 \
  --output artifacts/detector_coco_v06_native/evaluation \
  --eval-backend auto

Measure the decoder on the target machine before treating OpenCV as a speed claim:

python scripts/check_detection_io_performance.py \
  --annotations artifacts/COCO/annotations/instances_train2017.json \
  --images artifacts/COCO/train2017 \
  --samples 500 --repeats 3