Follow Docker official docs to install Docker: Install Docker Engine on Ubuntu.

Follow the script below to get Docker image:

docker pull numbqq/npu-vim4

The conversion tool is integrated in the NPU SDK.

$ git lfs install
$ git lfs clone https://gitlab.com/khadas/vim4_npu_sdk.git

$ cd vim4_npu_sdk
$ ls
adla-toolkit-binary  adla-toolkit-binary-3.1.7.4  convert-in-docker.sh  Dockerfile  docs  ksnn_args.txt  README.md

• adla-toolkit-binary/docs - SDK documentations
• adla-toolkit-binary/bin - SDK tools required for model conversion
• adla-toolkit-binary/python - KSNN conversion examples

Before converting your model, you should modify ksnn_args.txt. Those in text are the parameters which converting needs.

Finish modification, run following command to convert model.

$ bash convert-in-docker.sh ksnn

If everything works well you will find the converted files below.

~/vim4_npu_sdk$ ls adla-toolkit-binary/python/
convert  data  dataset.txt  onnx_output  yolov8n.onnx

Converted model is xxxx.adla in the parameter outdir you set and dynamic library is also in here.

• model-name – The name of generated model.
• model-type - Model type used in the conversion.
• model – The path of model file.
• weights – The path of model weight file. Caffe, DarkNet and MxNet model need set. Such as xxx.caffemodel, xxx.weights, xxx..params.
• inputs – The input name of model.
• input-shapes – The input shape of model.
• outputs – The output name of model. Only TensorFlow model must set.
• dtypes – Describe the type of input.
• quantize-dtype – Quantization type. Currently, it supports int8, int16 and uint8.
• outdir – The path of folder which saves converted model and dynamic library.
• channel-mean-value – The normalization of model. Args “m1,m2,m3,scale”. Image (data1, data2, data3).
  • Out1 = (data1-m1)/scale
  • Out2 = (data2-m2)/scale
  • Out3 = (data3-m3)/scale
• source-file – Dataset path and filename.
• iterations – Number of iterations to run.
• batch-size – Batch size of quantification model.
• kboard – Choose khadas board. Now only support VIM4.
• disable-per-channel – Disable per-channel quantitation. Default True.
• print-level – Information printing level. Default 0. Option 0, 1.
• inference-input-type – After converting to adla, the input data type of the model.
• inference-output-type – After converting to adla, the output data type of the model.