NPU SDK 6.4.3 release

Announcements
#1

New NPU SDK 6.4.3 base on Ubuntu 201126 Rlease.

Upgrade

  1. SUpport more Neural Newwork Layer

  2. Fix the bug that the conversion code does not match aml_npu_app in version 6.4.2.1.

  3. Fix the bug that the yolo model recognition result is incorrect in version 6.4.2.1.

  4. Update the aml_npu_app warehouse, fix the bug of low recognition accuracy rate of yolov4

New

  1. Added amlogic SDK api . You can use these interfaces to call nb files without using the converted vnn interface.

  2. Added amlogic SDK api documentation and AMLNN documentation.

  3. Added opencv4.2 support. Please refer to khadas Docs to compile or run the version of opencv4.

Docs

We rewrote all documents on the NPU section

  1. Remove the training document. Please refer to the official website of the model for this part.

  2. Added aml_app_npu and SDK and aml_npu_demo_binaries instructions for use.

  3. Rewrite the conversion example document.

Note

If you want to use openvc4 version with Ubuntu 20.04. Please follow my step to install opencv4 .

$ sudo aupt update
$ sudo apt upgrade
$ sudo apt remove opencv3
$ sudo apt install libopencv-dev python3-opencv
7 Likes
VIM3_Pie_V201112的安卓固件是使用v6.4.2.1版本的NPU SDK么？
Npu_sdk issue from new npu_sdk released yesterday
VIM3 yolov3 convert issue
#2

Nice! Are there plans to support MaskRCNN sometime in the future?

#3

@prozor Maybe , I am not sure . You can try to convert it

1 Like
#4

I’m working on it, I will share if I make any progress. Right now, convertkeras complains about unknown layers so I am trying to figure out how to specify Mask_RCNN’s custom layers to it. I have managed to get the Acuitylib package so I will try to play with it directly instead of using compiled binaries like convertkeras. Why is the Acuitylib Python package not provided anywhere, but compiled into the convertkeras, convertensorflow, convertpytorch… binaries?

#5

@prozor

Because this part is not open source, we don’t have source code either

1 Like
#6

But the source can be easily obtained from the binaries, no?

#7

@prozor Is not available

1 Like