halcon深度学习, tensorRT 问题请教

训练了一个Resnet50的分类模型.  使用如下代码转换模型到TensorRT:

1. HTuple hv_ConversionReport, hv_OptimizeForInferenceParams;
   
2. SetDlModelParam(hv_DLModelHandle, "device", hv_DLDeviceCPU);
   
3. GetDlDeviceParam(hv_DLDeviceTrt03, "optimize_for_inference_params", &hv_OptimizeForInferenceParams);
   
4. OptimizeDlModelForInference(hv_DLModelHandle, hv_DLDeviceTrt03, hv_Precision,HTuple(), hv_OptimizeForInferenceParams, &m_hv_DLModelHandle[2], &hv_ConversionReport);

复制代码

将加载的同一个模型转换3次:

1. OptimizeDlModelForInference(hv_DLModelHandle, hv_DLDeviceTrt01, hv_Precision, HTuple(), hv_OptimizeForInferenceParams, &m_hv_DLModelHandle[0], &hv_ConversionReport);
   
2. OptimizeDlModelForInference(hv_DLModelHandle, hv_DLDeviceTrt01, hv_Precision, HTuple(), hv_OptimizeForInferenceParams, &m_hv_DLModelHandle[1], &hv_ConversionReport);
   
3. OptimizeDlModelForInference(hv_DLModelHandle, hv_DLDeviceTrt01, hv_Precision, HTuple(), hv_OptimizeForInferenceParams, &m_hv_DLModelHandle[2], &hv_ConversionReport);

复制代码

然后对同一张图片进行推理, 发现不同模型置信度有少许偏差:

1. 0.967595<==>0.0324055
   
2. 0.968262<==>0.03158570.96802<==>0.0319796

复制代码

请问有人遇到同样的问题吗?  怎么解决?  

另外就是模型加载和转换过程很慢, 耗时很长.  请问是不是有解决办法.