Sparsity and sparse execution at inference accelerate object detection models like YOLOv8 fivefold, with minimal impact on accuracy. Neural Magic's sparse-quantized YOLOv8m model (top) achieves 25.5 FPS and recovers to 99% of the baseline accuracy, compared to a dense model (bottom) that achieves 5.5 FPS. Both models are running on a 4-core laptop CPU.
You can download free sparse-quantized models from Neural Magic's SparseZoo: sparsezoo.neuralmagic.com
To achieve real-time speeds on low CPU cores, use DeepSparse, Neural Magic's inference runtime: neuralmagic.com/deepsparse. You can use it for free via PIP, or in production for 90 days via a free trial. DeepSparse can be deployed in the cloud, at the edge, or anywhere in between on x86 and ARM CPUs.
Негізгі бет 5x Faster YOLOv8 on CPUs
Пікірлер: 2