EngineConfig

EngineConfig

• Description: High-level ONNXEngine configuration (dataclass) to centralize commonly used onnxruntime session/provider/run options.

• Dependencies

  • Requires onnxruntime.

• Common fields

  • graph_optimization: "disable" | "basic" | "extended" | "all" (or onnxruntime.GraphOptimizationLevel).
  • execution_mode: "sequential" | "parallel" (or onnxruntime.ExecutionMode).
  • intra_op_num_threads / inter_op_num_threads: thread counts.
  • provider_options: options keyed by provider name (merged into engine defaults).
  • fallback_to_cpu: whether to fallback to CPU when the requested provider is unavailable.
  • enable_io_binding: enable IO binding (may reduce copies; behavior depends on provider).
  • enable_profiling: enable ORT profiling.

• Example

  import numpy as np
  from capybara.onnxengine import EngineConfig, ONNXEngine
  
  cfg = EngineConfig(
      graph_optimization="all",
      enable_io_binding=False,
      fallback_to_cpu=True,
  )
  engine = ONNXEngine("model.onnx", backend="cpu", config=cfg)
  
  outputs = engine.run({"input": np.ones((1, 3, 224, 224), dtype=np.float32)})
  print(outputs.keys())