Yolov5sv6 not coverted with tengine!

I want to convert the yolov5s v6 model to the uint8 tmfile model on the vim3 a311d NPU. I follow many
web link such as Yolov5s 6 版模型训练设置与Tengine npu推理部署_Zxialuoke的博客-CSDN博客_npu模型部署 , after I generated the unit8 tmfile successfully which partly like that ,

1 Then I do ./tm_yolov5s_timvx -m yolo_test_uint8.tmfile -i ssd_dog.jpg , error info shows below

 2  I tried anthnoer model yolov5s_uint8.tmfile which is from the tengine's model zoo , the right result came yet with some error info.

3 Is there any reliable method or document to transfer the yolov5 or any other new model to uinit8 tmfile? It is not a automic tool to convert new model to unit8 model on the vim3, so much detailed needed to be deal with carefully !
4 The cost of yolov5s uint8 runing on the vim3 a311d 5TOPS should be less than 50ms , yet this results is more than 90ms ,so what could I reduce the time cost ?
my platform : ubuntu vim3 tengine a311d NPU

@lcl2020 Tengine will have issues running on NPU 6.4.6, we will release a new version at the end of the month. At that time, you can compile and run directly according to the Tengine documentation.

There will be the feature including converting the yolov5s or other model to the VIM3 NPU format easily and running on the vim3 as we expect in the new version at the end of the month, right?
However ,could you have a look at these question one by one ,I think the new version may not reslove all the problems.

@lcl2020 We are only responsible for ensuring that Tengine can be compiled and run on our board. As for whether to support yolov5 or other models, this is a support issue of the Tengine operator, which is not within the scope of our maintenance.

Hi, Frank
As you are working with Tengine closely. Could you please help me to ask Tengine several question,as they will feed back to you quickly.

Are Tengine working at liteHrNet deloying on the a311d NPU?
1 When could you release the liteHrnet uint8 model on a311d NPU?
2 How about PicoDet model?
3 When I tried to convert tinypose to a311D NPU with the command ./convert_tool -f paddle -m *.pbmodel -p *.pdixxx -o tinpose.tmfile, it will core dump or some op not supported! I think the tinypose model has been supported by Tengine alreay, right? How could I solve this problem?

Looking forward to your reply! Tks

Are you free now?


@lcl2020 We will release new firmware at the end of this month, you can compile and use Tengine directly according to the Tengine documentation

Your reply is always so simply and efficetive! We will look forward to your the new release this month.
Yet, there is another important quesions as following.
1 As we know, the sota pose estimation net liteHRnet is so famous in the are of pose estimation.So could Tengine support this efficient liteHRnet model to vim3 ?It is a so fantastic and amazing thing if Tengine could do this sota pose model !
2 Any plan of the lite-hrnet coverting to vim3 a311d NPU?
3 How about the sota of pose model just like tinypose ? Could we covert tinypose to the vim3 NPU?
Could you please send this to Tengie ,because you have worked together for a while! Tengine will deal with this request more carefully!


@lcl2020 The operator’s support mainly depends on whether the technical staff of Tengine has covered it. Check out Tengine’s documentation for this. The same is true for the SDK provided by amlogic. The support of operators is not done by us, all supported operators are in the support document.

I haven’t tried these operators, not sure if they support conversion. I’ll test it if I have time