A drop-in performance layer for RF-DETR. Replace one import. Keep your code. Get optimized fused kernels and GPU-level performance improvements automatically.
Illustrative benchmark. Actual speedups vary by GPU, model, batch size and workload.
from rfdetr import RFDETRBasefrom rfdetr_ghost import RFDETRBaseThat's it.
No rewrite. No new API. No migration project.
Just an optimized implementation underneath.
from rfdetr import RFDETRBase
model = RFDETRBase()
model.train(...)
predictions = model.predict(...)from rfdetr_ghost import RFDETRBase
model = RFDETRBase()
model.train(...)
predictions = model.predict(...)Ghost handles the low-level work so you can stay in Python. Expensive paths are swapped for implementations built for the GPU.
Measured on representative workloads. Your mileage may vary.
RF-DETR users interact with the same familiar Python API. Ghost handles the low-level optimization layer underneath.
See how it worksInstall it next to your existing project. No migration guide required.
$ pip install rfdetr-ghost
# before
from rfdetr import RFDETRBase
# after
from rfdetr_ghost import RFDETRBase
$ python your_existing_rfdetr_script.pyGhost operates beneath the API surface, replacing expensive paths with optimized implementations while keeping the developer experience familiar.
Drop-in API · Custom kernels · Open source