Richard Fang
Richard Fang
Sorry for the late reply. Yes, the reason is exactly what you said. Because I didn't realize that the TransformerBlock module already exists in the yolov5 repository when I first...
Sorry, I don't understand the meaning of replacing the input source path, whether it is exchanging the input of RGB and IR images？ If yes, there is indeed a huge...
> 我分析原因是你用的上面分支的label作为最后的label，IR的视野相对大一些，所以换过来效果会好。不知道理解的对不对。 我觉有可能是这个原因，但确定则还需要进一步测试。 感谢你提的建议。
这个画图是训练结束时候，用来画检测效果的函数并以马赛克的方式拼接，你可忽略它，或者你可以找到train中画图的代码部分，将其注销。 如果你需要展示检测效果，可以直接使用`detect_twostream.py`.
Have you tested it on the PC？
System: Ubuntu 16.04 torch version: 1.4 (source code install) conda python envs: 3.6.9 cuda version: 10.0.130 protobuf version: 3.11 tensorrt version: 5.1.0.5
还没尝试，之后打算xavier上试下，TX2这类前一代产品不支持半精度运算，在上面跑半精度模型提速非常有限。
你可以查看一下这个torch2trt下面这个issue [#166](https://github.com/NVIDIA-AI-IOT/torch2trt/issues/166)，这是我之前碰到和你一样问题的参考方法。
a) method one(If you install pytorch by pip or conda) Lower down your gcc and g++ version from 5 to 4.8.(As the change about the ABI issue) build the protobuf...
Hi @baheytharwat Sorry for long time no reply，I have been busy with other work in the past half a year and rarely check this repo. For your problem，you should check...