Wiring killswitch predator 670

W600k-r50.onnx Info

Example with ONNX Runtime GPU:

sess_options = ort.SessionOptions()
sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_EXTENDED
providers = ['CUDAExecutionProvider', 'CPUExecutionProvider']
sess = ort.InferenceSession("w600k-r50.onnx", sess_options, providers=providers)

python -m onnxruntime.tools.quantize --input w600k-r50.onnx --output w600k-r50-quant.onnx --mode dynamic

Output: [1, 512] (A 512-dimensional embedding vector).

aligned_face = cv2.imread("aligned_face.jpg") embedding = get_face_embedding(aligned_face) print(f"Embedding shape: embedding.shape") # (512,) w600k-r50.onnx

Convert the ONNX to TensorRT for 0.5ms inference latency. Example with ONNX Runtime GPU: sess_options = ort

trtexec --onnx=w600k-r50.onnx --saveEngine=w600k-r50.engine --fp16

How does w600k-r50.onnx compare to other popular face recognition models? python -m onnxruntime

| Model | Size (FP32) | LFW Accuracy | CPU Inference (Intel i7) | GPU (RTX 3060) | | :--- | :--- | :--- | :--- | :--- | | w600k-r50.onnx | 96 MB | 99.78% | 35 ms | 3 ms | | FaceNet (Inception) | 180 MB | 99.65% | 85 ms | 7 ms | | MobileFaceNet | 4 MB | 99.48% | 8 ms | 1 ms | | VGG-Face (16) | 500 MB | 98.95% | 120 ms | 9 ms |

Key Takeaway: The R50 model offers state-of-the-art accuracy (99.78% on Labeled Faces in the Wild benchmark) while being compact enough to run on a CPU at 30 FPS.

No model is perfect. The w600k-r50.onnx has specific weaknesses:

Powered by
Graphic Details Digital Media