NPU Model Convert

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

$ git clone https://gitlab.com/khadas/vim4_npu_sdk.git
$ cd vim4_npu_sdk
$ ls
adla-toolkit-binary  adla-toolkit-binary-1.2.0.9  convert-in-docker.sh  Dockerfile  docs  README.md

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

Convert the demo model in docker:

./convert-in-docker.sh

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

~/vim4_npu_sdk$ ls adla-toolkit-binary/demo/
caffe_output     darknet_output  dataset.txt   libstdc++_so  mxnet_output  paddle_output   quantized_tflite_output  tflite_output
convert_adla.sh  data            keras_output  model_source  onnx_output   pytorch_output  tensorflow_output

Converted model is xxxx.adla.

• model-type - Model type used in the conversion.
• model(model files)/weights(model weights file) - model parameter must be set. weights needs to be set when model-type is caffe/darknet/mxnet.
• inputs/input-shapes - inputs Input node names. input-shapes The sizes of input nodes. Need to be set when model-type is pytorch/onnx/mxnet/tensorflow/ tflite/paddle.
• outputs - Output node names. When model-type is tensorflow, outputs needs to be set.
• dtypes - Input type, set the type information corresponding to the input (optional).Default float32.
• quantize-dtype - Quantization type. Currently, it supports int8, int16 and uint8 quantification types.
• source-file - dataset.txt. The txt file contains paths to the images that need to be quantized. It supports images and npy files.
• channel-mean-value - The pre-processing parameters are set according to the pre processing methods used during model training. It includes four values, m1, m2, m3, and scale. The first three are mean-value parameters and the last one is the scale parameter. For the input data with three channels (data1, data2, data3), the pre-processing steps are
  • Out1 = (data1-m1)/scale
  • Out2 = (data2-m2)/scale
  • Out3 = (data3-m3)/scale
• batch-size - Set the batch-size for the adla file after conversion. Currently, the default value is 1.
• iterations - Optional parameter, if dataset.txt provides data for multi-groups, and needs to use all data provided for quantization, set iterations and make sure iterations*batch size = Number of data in the multi-groups.
• outdir - Directory for the generated. The default value is the current directory.
• target-platform - Specify the target platform for the adla file, it should be PRODUCT_PID0XA003 for VIM4.

For more information, please check docs/model_conversion_user_guide_1.2.pdf.