Introduction and guide on how to use the KSNN conversion tool.

The conversion tool is integrated in the NPU SDK.

$ git lfs install
$ git lfs clone --recursive https://github.com/khadas/aml_npu_sdk.git

The KSNN conversion tool is under acuity-toolkit/python,

$ cd aml_npu_sdk/acuity-toolkit/python && ls
$ convert  data  outputs

Choose tensorflow MobileNet SSD as an example.

1. Get the frozen model,

$ cd aml_npu_sdk/acuity-toolkit/python
$ wget https://github.com/yan-wyb/models-zoo/raw/master/tensorflow/mobilenet_ssd/ssd_mobilenet_v1_coco_2017_11_17.pb

2. Convert,

# uint8
$ ./convert --model-name mobilenet_ssd \
>           --platform tensorflow \
>           --model /home/yan/yan/Yan/models-zoo/tensorflow/mobilenet_ssd/ssd_mobilenet_v1_coco_2017_11_17.pb \
>           --input-size-list '300,300,3' \
>           --inputs FeatureExtractor/MobilenetV1/MobilenetV1/Conv2d_0/BatchNorm/batchnorm/mul_1 \
>           --outputs "'concat concat_1'" \
>           --mean-values '127.5 127.5 127.5 0.007843137' \
>           --quantized-dtype asymmetric_affine \
>           --source-files ./data/dataset/dataset0.txt \
>           --kboard VIM3 --print-level 0
 
# int8
$ ./convert --model-name mobilenet_ssd \
>           --platform tensorflow \
>           --model /home/yan/yan/Yan/models-zoo/tensorflow/mobilenet_ssd/ssd_mobilenet_v1_coco_2017_11_17.pb \
>           --input-size-list '300,300,3' \
>           --inputs FeatureExtractor/MobilenetV1/MobilenetV1/Conv2d_0/BatchNorm/batchnorm/mul_1 \
>           --outputs "'concat concat_1'" \
>           --mean-values '127.5 127.5 127.5 0.007843137' \
>           --quantized-dtype dynamic_fixed_point \
>           --qtype int8 \
>           --source-files ./data/dataset/dataset0.txt \
>           --kboard VIM3 --print-level 0
 
# int16
$ ./convert --model-name mobilenet_ssd \
>           --platform tensorflow \
>           --model /home/yan/yan/Yan/models-zoo/tensorflow/mobilenet_ssd/ssd_mobilenet_v1_coco_2017_11_17.pb \
>           --input-size-list '300,300,3' \
>           --inputs FeatureExtractor/MobilenetV1/MobilenetV1/Conv2d_0/BatchNorm/batchnorm/mul_1 \
>           --outputs "'concat concat_1'" \
>           --mean-values '127.5 127.5 127.5 0.007843137' \
>           --quantized-dtype dynamic_fixed_point \
>           --qtype int16 \
>           --source-files ./data/dataset/dataset0.txt \
>           --kboard VIM3 --print-level 0

During conversion, if you need to view detailed information, you can modify –print-level 0 to –print-level 1.

3. Model file generated after conversion.

$ cd aml_npu_sdk/acuity-toolkit/python
$ ls outputs/mobilenet_ssd/
mobilenet_ssd.nb  libnn_mobilenet_ssd.so

1. KSNN Usage
2. KSNN API Documentation