Why is this not working?
root@Khadas:~# sudo do-fenix-full-upgrade
sudo: do-fenix-full-upgrade: command not found
I don’t remember which image version I tried mid of January. . . My guess it was https://github.com/khadas/fenix/suites/1854204686/artifacts/36010770
With help from discord i tried this sdcard images as well.
But as soon as i shifted this to emmc i got a Bootloop as well
Tengine looks really interesting. I love that you’re not just updating, you’re also putting out new stuff. Thanks!
1 Like
Any feedback why my updated V20200530 ubuntu image is missing fenix-full-upgrade?
When I try to remove gnome-desktop, I find it will remove khadas apps together.
@numbqq @Frank About NN API
* add amlogic nnsdk libraries & headers
This API would help for easier deployment. In the SDK documentation, there is SDK integration demo like
./nnSDK mobilenetv1_a1.nb 0 224x224.jpeg
However, I could not find the executed scripts and their related code for running this in the NPU toolkit. Where can I get the code “nnSDK” for using the API?
Thanks.
But except image classification, all the demos in aml_nnsdk_app has been removed citing they are unusable.
And after following the instructions,
person_detect_640x384_camera
person_detect_640x384_picture
does not work as well.
@Vignesh_Raja You can reset to this commit 789ee9a7ef3e6b31f60f205b14fcf2a0fece9f47
If you think the recognition effect is average, you can modify our demo, our demo is open source. But I think this is the problem of the accuracy of the model after the conversion. A large part of it is not open source, and we can’t handle it.
1 Like
I reverted to that commit that you mentioned and I made the code work. I get output like this
The code publishes like
Num:4
i:0 x:0.614844 w:0.376563 y:0.0390625 h:0.927083 score:0
i:1 x:0.464063 w:0.115625 y:0.428385 h:0.361979 score:0
i:2 x:0.624219 w:0.376563 y:0.0690104 h:0.882812 score:0
i:3 x:0.170312 w:0.01875 y:0.539062 h:0.0729167 score:0
“score” is not exposed in the person_detect_t API. It seems bit slower and less accurate than aml_demo_binaries.
I would like to know the score of each object to eliminate the false positives.
Thanks in advance.
@Vignesh_Raja There is also a relationship between accuracy and pre-processing, and the nb file uses the converted nb file. I don’t know the parameters of the conversion and the parameter configuration of the original model. These are not open source, and it is difficult for me to make a comparison.
Yes, that make sense.
What about the “score” parameter? It seems to be not exposed in the API. It is very important to set the right threshold for detections.
Btw, NNSDK API is very user friendly than other ones. It would be nice to see more demos with this API.