FangJingYunner

请问TRT::infer推理其他模型可行吗？

3

您好，您的TensorRT开源项目中对TensorRT的封装我觉得对新手十分友好，上手非常简单，但是我还是有些疑问。我用TRT::infer加载其他任务如深度估计，语义分割的float32或者float16engine文件。我在推理结果保存成二进制文件后再python中再加载结果进可视化后发现推理出来的结果是错的。我onnx转engine使用trtexec转的。主要使用到的代码如下： ``` infer->input(0)->set_mat(0,img); infer1->forward(); auto out = infer1->output(0); infer1->output(0)->save_to_file(sttr); ``` onnx转trt是直接用的 ``` trtexec --onnx=XXXX --saveEngine=XXXX ``` 请问TRT::infer是不是不适用于其他模型的推理呢？

hello,the data organization doesn't show in readme.md. Would you please update the data organization part? 

About Deployment

1

Hi，It's a very great work.Can it be deployed on embed system such as Nvidia Jetson AGX?

hello，it's a very good work.And l have a question on it's precision with resnet18

hello，does the normnet result accuracy？ l have trained this normnet model on kitti dataset，all setting is same as monodepth2 except the decoder part ,and l have got a very bad result. the abs_rel is arround 0.35

1

hello，does the normnet result accuracy？ l have trained this normnet model on kitti dataset，all setting is same as monodepth2 except the decoder part ,and l have got a very bad...

As the title show,where to download kitti_split1?

训练损失出现nan的问题

2

您好，我用您的代码训练自己采集的旋转目标检测数据集。数据集中不一定是每一张图都有旋转目标检测的标注。但是网络在计算损失的时候有时候会出现一些变量是nan的情况，比如offset0，loss_fg,loss_neg是nan的情况。而且这个还时好时坏的，有的时候网络参数或者训练策略改改就不会有，有的时候在某些网络参数下训练了几百个step就会出现nan

请问代码可以给一个示例吗？谢谢！