The root contains only the model runner and the demo. All other executable utilities live here.
| Tool | Purpose |
|---|---|
compare.py |
Generate and validate the speed comparison table from benchmark JSON |
evaluate_gt.py |
Score model outputs against the configured ground-truth manifest |
verify_accuracy.py |
Compare an engine with its ONNX reference and FP32 output |
models.py |
Report registered, built, measured, and stale models |
package_artifacts.py |
Assemble reproducible deployment artifacts |
| Tool | Purpose |
|---|---|
profile_model.py |
Attribute runtime to preprocessing, inference, and postprocessing layers |
tune_build.py |
Build repeated TensorRT variants while sweeping builder settings |
size_sweep.py |
Measure the accuracy impact of input resolution |
make_eval_manifest.py |
Create an explicit evaluation subset |
prepare_eval_inputs.py |
Regenerate evaluation crops for the supported aspect ratios |
sync_desktop.sh |
Synchronize source and reports with the remote GPU desktop |
check_diode_convention.py and check_pointmap.py validate coordinate, scale, and point-map conventions. Run them after changing preprocessing or postprocessing.
tools/retired/ contains drivers for questions that have already been answered. They are retained for provenance and are not part of the normal build path. In particular, old video-frame helpers expect a user-provided local video/video2.mp4; that input is intentionally not tracked.
python tools/models.py --stale
python tools/compare.py --check
python -m pytest tests -q
git diff --check