This article mainly introduces how to call NPU through Python API in VIM3/VIM3L.

1. Clone code to local.

$ git clone --recursive https://github.com/khadas/ksnn.git

2. Installation dependencies.

$ pip3 install matplotlib

3. Install KSNN library.

$ cd ksnn/ksnn
$ pip3 install ksnn-1.3-py3-none-any.whl

Demos are all concentrated in the examlpes directory,

$ cd ksnn/examples/ && ls
caffe  darknet  keras  onnx  pytorch  tensorflow  tflite

Take Inception V3 as an example, other demos are similar.

$ cd tensorflow && ls
README.md  box_priors.txt  data  inceptionv3.py  libs  mobilenet_ssd_picture.py  models

The running commands and conversion parameters are in the README file in the corresponding directory.

$ ~/ksnn/examples/tensorflow$ cat README.md
 
# run
 
$ python3 inceptionv3.py --model ./models/VIM3/inceptionv3.nb --library ./libs/libnn_inceptionv3.so --picture ./data/goldfish_299x299.jpg --level 0
$ python3 mobilenet_ssd_picture.py --model ./models/VIM3/mobilenet_ssd.nb --library ./libs/libnn_mobilenet_ssd.so --picture data/1080p.bmp --level 0
 
# Convert
 
$ ./convert \
--model-name inception \
--platform tensorflow \
--model inception_v3_2016_08_28_frozen.pb \
--input-size-list '299,299,3' \
--inputs input \
--outputs InceptionV3/Predictions/Reshape_1 \
--mean-values '128,128,128,128' \
--quantized-dtype asymmetric_affine \
--kboard VIM3 --print-level 1
 
$ ./convert \
--model-name mobilenet_ssd \
--platform tensorflow \
--model 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 '127.5,127.5,127.5,127.5' \
--quantized-dtype asymmetric_affine \
--kboard VIM3 --print-level 1
 
If you use VIM3L , please use `VIM3L` to replace `VIM3`

Run Inception V3:

$ python3 inceptionv3.py --model ./models/VIM3/inceptionv3.nb --library ./libs/libnn_inceptionv3.so --picture ./data/goldfish_299x299.jpg --level 0
 |--- KSNN Version: v1.3 +---| 
Start init neural network ...
Done.
Get input data ...
Done.
Start inference ...
Done. inference :  0.042353153228759766
----- Show Top5 +-----
     2: 0.93457
   795: 0.00328
   408: 0.00158
   974: 0.00148
   393: 0.00093

The –level parameter can be used to adjust the level of printed information. The following command sets the printing level to the highest.

$ python3 inceptionv3.py --model ./models/VIM3/inceptionv3.nb --library ./libs/libnn_inceptionv3.so --picture ./data/goldfish_299x299.jpg --level 2
 |--- KSNN Version: v1.3 +---| 
Start init neural network ...
#productname=VIPNano-QI, pid=0x88
Create Neural Network: 283ms or 283181us
Done.
Get input data ...
Done.
Start inference ...
Start run graph [1] times...
generate command buffer, total device count=1, core count per-device: 1, 
current device id=0, AXI SRAM base address=0xff000000
---------------------------Begin VerifyTiling -------------------------
AXI-SRAM = 1048576 Bytes VIP-SRAM = 522240 Bytes SWTILING_PHASE_FEATURES[1, 1, 0]
  0 NBG [(   0    0    0 0,        0, 0x(nil)(0x(nil), 0x(nil)) ->    0    0    0 0,        0, 0x(nil)(0x(nil), 0x(nil))) k(0 0    0,        0) pad(0 0) pool(0 0, 0 0)]
 
 id IN [ x  y  w   h ]   OUT  [ x  y  w  h ] (tx, ty, kpc) (ic, kc, kc/ks, ks/eks, kernel_type)
   0 NBG DD 0x(nil) [   0    0        0        0] -> DD 0x(nil) [   0    0        0        0] (  0,   0,   0) (       0,        0, 0.000000%, 0.000000%, NONE)
 
PreLoadWeightBiases = 1048576  100.000000%
---------------------------End VerifyTiling -------------------------
layer_id: 0 layer name:network_binary_graph operation[0]:unkown operation type target:unkown operation target.
uid: 0
abs_op_id: 0
execution time:             20552 us
[     1] TOTAL_READ_BANDWIDTH  (MByte): 67.540481
[     2] TOTAL_WRITE_BANDWIDTH (MByte): 18.245340
[     3] AXI_READ_BANDWIDTH  (MByte): 30.711348
[     4] AXI_WRITE_BANDWIDTH (MByte): 15.229973
[     5] DDR_READ_BANDWIDTH (MByte): 36.829133
[     6] DDR_WRITE_BANDWIDTH (MByte): 3.015367
[     7] GPUTOTALCYCLES: 94344921
[     8] GPUIDLECYCLES: 78109663
VPC_ELAPSETIME: 118090
*********
Run the 1 time: 118.00ms or 118636.00us
vxProcessGraph execution time:
Total   118.00ms or 118996.00us
Average 119.00ms or 118996.00us
Done. inference :  0.1422710418701172
----- Show Top5 +-----
     2: 0.93457
   795: 0.00328
   408: 0.00158
   974: 0.00148
   393: 0.00093

You can see all relevant information.

1. The Demos that currently support cameras include the Yolo series and OpenPose. Take Yolov3 as an example,

$ cd ksnn/examples/darknet
$ python3 hand-cap.py --model ./models/VIM3/hand.nb --library ./libs/libnn_hand.so --device X

2. Currently, the only demo that supports RTSP is the yolo series. Take Yolov3 as an example,

$ cd ksnn/examples/darknet
$ python3 flask-yolov3.py --model ./models/VIM3/yolov3.nb --library ./libs/libnn_yolov3.so --device X

• KSNN Convert tool Usage
• KSNN API Documentation