RK3588部署yolov8量化精度对比

1. 准备文件# 配置区 ONNX_MODEL best.onnx # YOLOv8 ONNX 模型路径 DATASET ./COCO/coco_subset_20.txt # 量化校准集 TEST_IMG frame_000000.jpg # 用于精度分析的测试图片 TARGET_PLATFORM rk3588 # 目标芯片 # 2. 调用rknn.accuracy_analysis工具分析精度# 1. 创建 RKNN 对象 rknn RKNN(verboseTrue) # 2. 模型配置 print(-- Config model) rknn.config(mean_values[0, 0, 0], std_values[255, 255, 255], target_platformTARGET_PLATFORM) print(done) # 3. 加载 ONNX 模型 print(-- Loading model) ret rknn.load_onnx(modelONNX_MODEL) if ret ! 0: print(Load model failed!) exit(ret) print(done) # 4. 构建模型 print(-- Building model) ret rknn.build(do_quantizationTrue, datasetDATASET) if ret ! 0: print(Build model failed!) exit(ret) print(done) # 5. 精度分析 (核心步骤) print(-- Accuracy analysis) ret rknn.accuracy_analysis(inputs[TEST_IMG], output_dir./snapshot, targetNone) if ret ! 0: print(Accuracy analysis failed!) exit(ret) print(done) rknn.release()3. 在./snapshot/error_analysis.txt中可以查看余弦相似度报告# simulator_error: calculate the output error of each layer of the simulator (compared to the golden value). # entire: output error of each layer between golden and simulator, these errors will accumulate layer by layer. # single: single-layer output error between golden and simulator, can better reflect the single-layer accuracy of the simulator. layer_name simulator_error entire single cos euc cos euc ----------------------------------------------------------------------------------------------- [Input] images 1.00000 | 0.0 1.00000 | 0.0 [exDataConvert] images_int8 1.00000 | 1.7407 1.00000 | 1.7407 [Conv] /model.0/conv/Conv_output_0 0.99997 | 61.855 0.99997 | 61.855 [exSwish] /model.0/act/Mul_output_0 0.99958 | 96.026 0.99963 | 90.032 [Conv] /model.1/conv/Conv_output_0 0.99887 | 187.64 0.99893 | 181.58 …… [exSwish] /model.22/cv3.0/cv3.0.1/act/Mul_output_0 0.98225 | 77.593 0.99978 | 8.6863 [Conv] /model.22/cv3.0/cv3.0.2/Conv_output_0 [Sigmoid] onnx::ReduceMax_325_int8 0.99751 | 0.3333 0.99994 | 0.0683 [exDataConvert] onnx::ReduceMax_325 0.99751 | 0.3333 0.99999 | 0.0199 [ReduceMax] 326_int8 0.99752 | 0.3327 0.99999 | 0.0177 [exDataConvert] 326 0.99752 | 0.3327 0.99999 | 0.0177可以看到余弦相似度大部分在0.99以上但也存在0.99以下的反映在结果上会存在检测框有一点点偏移或者个别低置信度的目标丢失后续将进行改进。