Khadas Docs

Amazing Khadas, always amazes you!

User Tools

Site Tools

Home / products / sbc / vim4 / npu / npu-sdk
products:sbc:vim4:npu:npu-sdk

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
Next revision 		Previous revision
Last revision Both sides next revision
products:sbc:vim4:npu:npu-sdk [2023/06/07 01:53]
louis 		products:sbc:vim4:npu:npu-sdk [2024/04/18 04:33]
nick [Get Convert Tool]
Line 1: Line 1:
-====== VIM4 NPU SDK Usage ======+~~tag>VIM4 NPU Docker ~~ 
 + 
 +====== NPU Model Convert ====== 
 + 
 +{{indexmenu_n>2}}
  
 <WRAP important > <WRAP important >
-Only **New VIM4** supports NPU, you can check the version of your VIM4 here: [[products:sbc:vim4:configurations:identify-version|]]+Only **New VIM4** supports NPU, you can [[../configurations/identify-version|check VIM4 the version]]
 </WRAP> </WRAP>
  
Line 9: Line 13:
 </WRAP> </WRAP>
  
-===== Install Docker =====+===== Build Docker Environment =====
  
 Follow Docker official docs to install Docker: [[https://docs.docker.com/engine/install/ubuntu/|Install Docker Engine on Ubuntu]]. Follow Docker official docs to install Docker: [[https://docs.docker.com/engine/install/ubuntu/|Install Docker Engine on Ubuntu]].
  
-===== Get NPU SDK =====+Follow the script below to get Docker image:
  
 ```shell ```shell
-$ git clone https://gitlab.com/khadas/vim4_npu_sdk.git +docker pull numbqq/npu-vim4
-$ cd vim4_npu_sdk +
-$ ls +
-bin  demo  docs  README.md+
 ``` ```
  
-  * docs: SDK documentations +===== Get Convert Tool =====
-  * bin: SDK tools required for model conversion +
-  * demo: Conversion examples +
- +
-===== Convert model in Docker ===== +
- +
-Get Docker:+
  
 ```shell ```shell
 +$ git clone https://gitlab.com/khadas/vim4_npu_sdk.git --depth=1
 $ cd vim4_npu_sdk $ cd vim4_npu_sdk
-$ docker pull yanwyb/npu:v1 +$ ls 
-$ docker run -it --name vim4-npu1 -v $(pwd):/home/khadas/npu \ +adla-toolkit-binary  adla-toolkit-binary-1.2.0.9  convert-in-docker.sh  Dockerfile  docs  README.md
- -v /etc/localtime:/etc/localtime:ro \ +
- -v /etc/timezone:/etc/timezone:ro \ +
- yanwyb/npu:v1 +
 ``` ```
 +
 +  * ''adla-toolkit-binary/docs'' - SDK documentations
 +  * ''adla-toolkit-binary/bin'' - SDK tools required for model conversion
 +  * ''adla-toolkit-binary/demo'' - Conversion examples
 +
 +===== Convert Model =====
  
 Convert the demo model in docker: Convert the demo model in docker:
  
 ```shell ```shell
-khadas@2655b6cbbc01:~/npu$ cd demo/ +./convert-in-docker.sh
-khadas@2655b6cbbc01:~/npu/demo$ bash convert_adla.sh+
 ``` ```
  
-If everything works well you will find the converted files below:+If everything works well you will find the converted files below.
  
 ```shell ```shell
-khadas@2655b6cbbc01:~/npu/demo$ ls+~/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 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 convert_adla.sh  data            keras_output  model_source  onnx_output   pytorch_output  tensorflow_output
 ``` ```
  
-===== Parameter =====+Converted model is ''xxxx.adla''.
  
-  - ''model-type'': Model type used in the conversion. +<WRAP tip > 
-  - ''model(model files)/weights(model weights file)'':''model'' parameter must be set. ''weights'' needs to be set when model-type is caffe/darknet/mxnet. +If you want to convert your own model, just modify script ''adla-toolkit-binary/demo/convert_adla.sh'' and  then execute ''./convert-in-docker.sh'' to convert your model. 
-  - ''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. +</WRAP> 
-  - ''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 quantization types. --source-file: dataset.txt. The txt file contains paths to the images that need to be quantitated. it supports image and npy files. +==== Important parameters ==== 
-  - ''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:+ 
 +  * ''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     * Out1 = (data1-m1)/scale
     * Out2 = (data2-m2)/scale     * Out2 = (data2-m2)/scale
     * Out3 = (data3-m3)/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. +  * ''batch-size'' - Set the batch-size for the adla file after conversion. Currently, the default value is 1. 
-  - ''outdir'': Directory for the generated. The default value is the current directory:”./”. +  * ''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. 
-  - ''target-platform'': specify the target platform for the adla file. Currently, the default value is PRODUCT_PID0XA001.  +  * ''outdir'' - Directory for the generated. The default value is the current directory. 
-    * There are currently three valid values: PRODUCT_PID0XA001、PRODUCT_PID0XA002 and PRODUCT_PID0XA003. +  * ''target-platform'' - Specify the target platform for the adla file, it should be ''PRODUCT_PID0XA003'' for VIM4.
-    * You can use the following command to confirm the value that should be set.Execute: +
-```shell +
-khadas@Khadas:~$ cat /proc/cpuinfo +
-``` +
-    * View the first four characters of the sequence code corresponding to Serial in the second last line (the part in bold after Serial). Serial number and parameters PRODUCT_PID0X? with the relationship are as follows: +
- +
-{{:products:sbc:vim4:npu:npu-pid.webp?1000|}}+
  
 ===== See Also ===== ===== See Also =====
  
 For more information, please check ''docs/model_conversion_user_guide_1.2.pdf''. For more information, please check ''docs/model_conversion_user_guide_1.2.pdf''.
 +
Last modified: 2024/04/18 04:33 by nick

Page Tools

Khadas Site

Khadas Community

Khadas GitHub

Khadas Downloads