import onnxruntime as ort import numpy as np import time # Update this with your ONNX model path MODEL_PATH = r"C:\Users\upaya\Downloads\fasterrcnn_mobilenet_v3_large_320_fpn_Opset16.onnx" def summarize_provider(session): try: providers = session.get_providers() print("ONNXRuntime Providers available:", providers) prefer = session.get_provider_options() print("Provider options:", prefer) except Exception as e: print("Provider info unavailable:", e) def get_input_details(session): input_meta = session.get_inputs()[0] name = input_meta.name shape = [d if isinstance(d, int) else 1 for d in input_meta.shape] # Set undefined dims to 1 dtype = np.float32 if input_meta.type == "tensor(float)" else None return name, shape, dtype def benchmark(session, input_name, dummy_input, runs=10): # Warmup for _ in range(3): session.run(None, {input_name: dummy_input}) # Timed runs start = time.perf_counter() for _ in range(runs): output = session.run(None, {input_name: dummy_input}) elapsed = time.perf_counter() - start print(f"Avg inference time ({session.get_providers()[0]}): {elapsed/runs:.4f}s") return output def main(): providers = [("DmlExecutionProvider", {"device_id": 0}), "CPUExecutionProvider"] for provider in providers: print(f"\nTrying provider: {provider}") try: session = ort.InferenceSession( MODEL_PATH, providers=[provider] if isinstance(provider, str) else [provider[0]], provider_options=[provider[1]] if isinstance(provider, tuple) else None, ) summarize_provider(session) input_name, input_shape, input_dtype = get_input_details(session) dummy_input = np.random.rand(*input_shape).astype(input_dtype) output = benchmark(session, input_name, dummy_input) print( f"Output type/len: {type(output)}, {[o.shape if hasattr(o, 'shape') else type(o) for o in output]}" ) except Exception as e: print(f"Error with provider {provider}: {e}") if __name__ == "__main__": main()