Skip to content

backends.md

Latest commit

 

History

History
60 lines (42 loc) · 1.64 KB

backends.md

File metadata and controls

60 lines (42 loc) · 1.64 KB
Raw

Backend Selection and Checks

Choosing DML vs CUDA

Use DML When

  • You want the least complicated Windows GPU path.
  • You are testing compatibility.
  • You do not want to install CUDA and TensorRT.
  • You are using an ONNX model.

Use CUDA + TensorRT When

  • You have an NVIDIA GPU.
  • CUDA Toolkit and TensorRT are installed.
  • You want the highest-performance backend.
  • You are using a TensorRT .engine model, or you have an .onnx ready for first-time engine generation.

DML Runs But Does Not Detect Anything

Check these first:

backend = DML
ai_model = your_model.onnx
confidence_threshold = 0.10
class_player = 0
class_head = 1

Common causes:

  • The selected model is a TensorRT .engine instead of .onnx.
  • confidence_threshold is too high for the model.
  • The model class IDs do not match class_player and class_head.
  • dml_device_id points to the wrong GPU.
  • The capture source is not actually showing the target content.

CUDA Runs But GPU Usage Spikes

Start by checking which features force CPU-readable frames:

  • Debug/preview window: show_window = true.
  • Data collection.
  • Screenshots.
  • Any feature that needs pixels on the CPU for display or saving.

For the current CUDA path, the recommended FOV limiter is:

circle_fov_enabled = true

If the spike disappears when the GUI or overlay is open, compare the capture diagnostics in both states. The GUI/preview can change whether the app requests CPU copies, which can make the runtime path different from the closed-GUI path.

Related docs: