Difference between revisions of "RK Platform NPU SDK"

From ESS-WIKI
Jump to: navigation, search
 
(21 intermediate revisions by the same user not shown)
Line 23: Line 23:
 
*One isolated voltage domain to support DVFS  
 
*One isolated voltage domain to support DVFS  
  
 +
 +
== AI Program Development ==
 +
 +
AI Program Development for Rockchip platform mainly has two stages:
 +
 +
*AI Model Transfer.
 +
*Program Development.
 +
 +
We provide PC tools (rknn-toolkit2) to do "AI Model Transfer" and provide "AI API (rknpu2)" and "Qt/Gcc toolchain" to develop AI applications.
 +
 +
Gcc toolchain can be gotton from released image.
 +
 +
Qt toolchain can be gotton from [[RK_ALL_Qt_Toolchain|Advantech WIKI]].
 +
 +
 +
To get more details about RKNN C API from "rknpu2/doc/Rockchip_RKNPU_UserGuide_RKNN_API_V*.ptf"
 +
 +
To get more details about RKNN Python API from "rknn-toolkit2/rknn_toolkit_lite2/docs/Rockchip_UserGuide_RKNN_Toolkit_Lite2_V*.ptf"
 +
 +
 +
[[File:RK NPU Framework.png|800px|RTENOTITLE]]
  
 
== RKNN SDK ==
 
== RKNN SDK ==
  
−
RKNN SDK （Password: a887）include two parts:
+
RKNN SDK （[https://pan.baidu.com/s/1R0DhNJU56Uhp4Id7AzNfgQ Baidu] Password: a887）include two parts:
  
 
*rknn-toolkit2  
 
*rknn-toolkit2  
 
*rknpu2  
 
*rknpu2  
−
 
 
<pre>
 
<pre>
−
 
 
├── rknn-toolkit2
 
├── rknn-toolkit2
 
│  ├── doc
 
│  ├── doc
Line 45: Line 64:
  
 
= '''rknpu2''' =
 
= '''rknpu2''' =
−
TBD
+
 
 +
'rknpu2' include documents (rknpu2/doc) and examples (rknpu2/examples) to help to fast develop AI applications using rknn model(*.rknn).
 +
 
 +
Other models (eg:Caffe、TensorFlow etc) can be translated to rknn model through 'rknn-toolkit2'.
 +
 
 +
RKNN API Library file librknnrt.so and header file rknn_api.h can be found in rknpu2/runtime.
 +
 
 +
Released BSP and images have already included NPU driver and runtime libraries.
 +
 
 +
Here are two examples built in released images:
 +
 
 +
=== rknn_ssd_demo ===
 +
<pre>cd /tools/test/adv/npu2/rknn_ssd_demo
 +
./rknn_ssd_demo model/ssd_inception_v2.rknn model/bus.jpg
 +
 
 +
 
 +
resize 640 640 to 300 300
 +
Loading model ...
 +
rknn_init ...
 +
model input num: 1, output num: 2
 +
input tensors:
 +
  index=0, name=Preprocessor/sub:0, n_dims=4, dims=[1, 300, 300, 3], n_elems=270000, size=270000, fmt=NHWC, type=INT8, qnt_type=AFFINE, zp=0, scale=0.007843
 +
output tensors:
 +
  index=0, name=concat:0, n_dims=4, dims=[1, 1917, 1, 4], n_elems=7668, size=7668, fmt=NCHW, type=INT8, qnt_type=AFFINE, zp=50, scale=0.090787
 +
  index=1, name=concat_1:0, n_dims=4, dims=[1, 1917, 91, 1], n_elems=174447, size=174447, fmt=NCHW, type=INT8, qnt_type=AFFINE, zp=58, scale=0.140090
 +
rknn_run
 +
loadLabelName
 +
ssd - loadLabelName ./model/coco_labels_list.txt
 +
loadBoxPriors
 +
person @ (106 245 216 535) 0.994422
 +
bus @ (87 132 568 432) 0.991533
 +
person @ (213 231 288 511) 0.843047
 +
 
 +
</pre>
 +
 
 +
[[File:Rknn ssd demo output.jpg|RTENOTITLE]]
 +
 
 +
 
 +
=== rknn_mobilenet_demo ===
 +
<pre>cd /tools/test/adv/npu2/rknn_mobilenet_demo
 +
./rknn_mobilenet_demo model/mobilenet_v1.rknn model/cat_224x224.jpg
 +
 
 +
model input num: 1, output num: 1
 +
input tensors:
 +
  index=0, name=input, n_dims=4, dims=[1, 224, 224, 3], n_elems=150528, size=150528, fmt=NHWC, type=INT8, qnt_type=AFFINE, zp=0, scale=0.007812
 +
output tensors:
 +
  index=0, name=MobilenetV1/Predictions/Reshape_1, n_dims=2, dims=[1, 1001, 0, 0], n_elems=1001, size=1001, fmt=UNDEFINED, type=INT8, qnt_type=AFFINE, zp=-128, scale=0.003906
 +
rknn_run
 +
--- Top5 ---
 +
283: 0.468750
 +
282: 0.242188
 +
286: 0.105469
 +
464: 0.089844
 +
264: 0.019531
 +
 
 +
</pre>
  
 
= '''rknn-toolkit2''' =
 
= '''rknn-toolkit2''' =
−
TBD
+
 
 +
== Tool Introduction ==
 +
 
 +
RKNN-Toolkit2 is a development kit that provides users with model conversion, inference and performance evaluation on PC platforms. Users can easily complete the following functions through the Python interface provided by the tool:
 +
 
 +
#<span style="color:#0000ff;">'''Model conversion'''</span>: support to convert Caffe / TensorFlow / TensorFlow Lite / ONNX / Darknet / PyTorch model to RKNN model, support RKNN model import/export, which can be used on Rockchip NPU platform later.
 +
#<span style="color:#0000ff;">'''Quantization'''</span>: support to convert float model to quantization model, currently support quantized methods including asymmetric quantization (asymmetric_quantized-8). and support hybrid quantization.
 +
#<span style="color:#0000ff;">'''Model inference'''</span>: Able to simulate NPU to run RKNN model on PC and get the inference result. This tool can also distribute the RKNN model to the specified NPU device to run, and get the inference results.
 +
#<span style="color:#0000ff;">'''Performance & Memory evaluation'''</span>: distribute the RKNN model to the specified NPU device to run, and evaluate the model performance and memory consumption in the actual device.
 +
#<span style="color:#0000ff;">'''Quantitative error analysis'''</span>: This function will give the Euclidean or cosine distance of each layer of inference results before and after the model is quantized. This can be used to analyze how quantitative error occurs, and provide ideas for improving the accuracy of quantitative models.
 +
#<span style="color:#0000ff;">'''Model encryption'''</span>: Use the specified encryption method to encrypt the RKNN model as a whole.
 +
 
 +
 
 +
 
 +
== System Dependency ==
 +
 
 +
'''OS Version'''&nbsp;: Ubuntu18.04(x64) / Ubuntu20.04(x64)
 +
 
 +
'''Python Version'''&nbsp;: 3.6 / 3.8
 +
 
 +
'''Python library dependencies'''&nbsp;:
 +
 
 +
'''Python 3.6'''
 +
<pre>cat rknn-toolkit2/doc/requirements_cp36*.txt
 +
# if install failed, please change the pip source to 'https://mirror.baidu.com/pypi/simple'
 +
 
 +
# base deps
 +
numpy==1.19.5
 +
protobuf==3.12.2
 +
flatbuffers==1.12
 +
 
 +
# utils
 +
requests==2.27.1
 +
psutil==5.9.0
 +
ruamel.yaml==0.17.4
 +
scipy==1.5.4
 +
tqdm==4.64.0
 +
bfloat16==1.1
 +
opencv-python==4.5.5.64
 +
 
 +
# base
 +
onnx==1.9.0
 +
onnxoptimizer==0.2.7
 +
onnxruntime==1.10.0
 +
torch==1.10.1
 +
torchvision==0.11.2
 +
tensorflow==2.6.2
 +
</pre>
 +
 
 +
'''Python3.8'''
 +
<pre>cat rknn-toolkit2/doc/requirements_cp38*.txt
 +
# if install failed, please change the pip source to 'https://mirror.baidu.com/pypi/simple'
 +
 
 +
# base deps
 +
numpy==1.19.5
 +
protobuf==3.12.2
 +
flatbuffers==1.12
 +
 
 +
# utils
 +
requests==2.27.1
 +
psutil==5.9.0
 +
ruamel.yaml==0.17.4
 +
scipy==1.5.4
 +
tqdm==4.64.0
 +
bfloat16==1.1
 +
opencv-python==4.5.5.64
 +
 
 +
# base
 +
onnx==1.9.0
 +
onnxoptimizer==0.2.7
 +
onnxruntime==1.10.0
 +
torch==1.10.1
 +
torchvision==0.11.2
 +
</pre>
 +
 
 +
&nbsp;
 +
 
 +
 
 +
== Installation ==
 +
 
 +
Create virtualenv environment. If there are multiple versions of the Python environment in the system, it is recommended to use virtualenv to manage the Python environment. Take Python3.6 for example:
 +
 
 +
*Install virtualenv 、Python3.6 and pip3
 +
<pre>sudo apt-get install virtualenv
 +
sudo apt-get install python3 python3-dev python3-pip
 +
</pre>
 +
 
 +
*Install dependencies
 +
<pre>sudo apt-get install libxslt1-dev zlib1g zlib1g-dev libglib2.0-0 libsm6 libgl1-mesa-glx libprotobuf-dev gcc
 +
</pre>
 +
 
 +
*Install requirements_cp36-*.txt
 +
<pre>virtualenv -p /usr/bin/python3 venv
 +
source venv/bin/activate
 +
sed -i 's|bfloat16==|#bfloat16==|g' rknn-toolkit2/doc/requirements_cp36-*.txt
 +
pip3 install -r rknn-toolkit2/doc/requirements_cp36-*.txt
 +
sed -i 's|#bfloat16==|bfloat16==|g' rknn-toolkit2/doc/requirements_cp36-*.txt
 +
pip3 install -r rknn-toolkit2/doc/requirements_cp36-*.txt
 +
</pre>
 +
 
 +
*Install RKNN-Toolkit2
 +
<pre>pip3 install rknn-toolkit2/packages/rknn_toolkit2*cp36*.whl
 +
</pre>
 +
 
 +
*Check whether RKNN-Toolkit2 install successfully，press ctrl+d to exit
 +
<pre>
 +
python3
 +
>>> from rknn.api import RKNN
 +
>>>
 +
</pre>
 +
If install successfully, there is no error information.
 +
 
 +
Here is one of the failed informations:
 +
<pre>
 +
>>> from rknn.api import RKNN
 +
Traceback (most recent call last):
 +
  File "<stdin>"，line 1，in <module>
 +
ImportError: No module named 'rknn'
 +
</pre>
 +
 
 +
== Model Conversion Demo ==
 +
 
 +
Here gives an example to show how to convert tflite model(mobilenet_v1_1.0_224.tflite) to RKNN model (mobilenet_v1.rknn) ON PC.
 +
 
 +
<pre>
 +
cd rknn-toolkit2/examples/tflite/mobilenet_v1
 +
python3 test.py
 +
 
 +
 
 +
W __init__: rknn-toolkit2 version: 1.4.0-22dcfef4
 +
--> Config model
 +
W config: 'target_platform' is None, use rk3566 as default, Please set according to the actual platform!
 +
done
 +
--> Loading model
 +
...  ...
 +
done
 +
--> Building model
 +
Analysing : 100%|█████████████████████████████████████████████████| 58/58 [00:00<00:00, 2739.15it/s]
 +
Quantizating : 100%|████████████████████████████████████████████████| 58/58 [00:00<00:00, 95.72it/s]
 +
... ...
 +
I rknn buiding done
 +
done
 +
--> Export rknn model
 +
done
 +
--> Init runtime environment
 +
W init_runtime: Target is None, use simulator!
 +
done
 +
--> Running model
 +
Analysing : 100%|█████████████████████████████████████████████████| 60/60 [00:00<00:00, 2740.54it/s]
 +
Preparing : 100%|██████████████████████████████████████████████████| 60/60 [00:00<00:00, 136.33it/s]
 +
mobilenet_v1
 +
-----TOP 5-----
 +
[156]: 0.9345703125
 +
[155]: 0.0570068359375
 +
[205]: 0.00429534912109375
 +
[284]: 0.003116607666015625
 +
[285]: 0.00017178058624267578
 +
 
 +
done
 +
 
 +
</pre>

Latest revision as of 01:26, 19 September 2023

Preface

NPU Introduce

RK3568

  • Neural network acceleration engine with processing performance up to 0.8 TOPS
  • Support integer 4, integer 8, integer 16, float 16, Bfloat 16 and tf32 operation
  • Support deep learning frameworks: TensorFlow, Caffe, Tflite, Pytorch, Onnx NN, Android NN, etc.
  • One isolated voltage domain to support DVFS

RK3588

  • Neural network acceleration engine with processing performance up to 6 TOPS
  • Include triple NPU core, and support triple core co-work, dual core co-work, and work independently
  • Support integer 4, integer 8, integer 16, float 16, Bfloat 16 and tf32 operation
  • Embedded 384KBx3 internal buffer
  • Multi-task, multi-scenario in parallel
  • Support deep learning frameworks: TensorFlow, Caffe, Tflite, Pytorch, Onnx NN, Android NN, etc.
  • One isolated voltage domain to support DVFS


AI Program Development

AI Program Development for Rockchip platform mainly has two stages:

  • AI Model Transfer.
  • Program Development.

We provide PC tools (rknn-toolkit2) to do "AI Model Transfer" and provide "AI API (rknpu2)" and "Qt/Gcc toolchain" to develop AI applications.

Gcc toolchain can be gotton from released image.

Qt toolchain can be gotton from Advantech WIKI.


To get more details about RKNN C API from "rknpu2/doc/Rockchip_RKNPU_UserGuide_RKNN_API_V*.ptf"

To get more details about RKNN Python API from "rknn-toolkit2/rknn_toolkit_lite2/docs/Rockchip_UserGuide_RKNN_Toolkit_Lite2_V*.ptf"


RTENOTITLE

RKNN SDK

RKNN SDK （Baidu Password: a887）include two parts:

  • rknn-toolkit2
  • rknpu2 
├── rknn-toolkit2
│   ├── doc
│   ├── examples
│   ├── packages
│   └── rknn_toolkit_lite2
└── rknpu2
    ├── doc
    ├── examples
    └── runtime

rknpu2

'rknpu2' include documents (rknpu2/doc) and examples (rknpu2/examples) to help to fast develop AI applications using rknn model(*.rknn).

Other models (eg:Caffe、TensorFlow etc) can be translated to rknn model through 'rknn-toolkit2'.

RKNN API Library file librknnrt.so and header file rknn_api.h can be found in rknpu2/runtime.

Released BSP and images have already included NPU driver and runtime libraries.

Here are two examples built in released images:

rknn_ssd_demo

cd /tools/test/adv/npu2/rknn_ssd_demo
./rknn_ssd_demo model/ssd_inception_v2.rknn model/bus.jpg


resize 640 640 to 300 300
Loading model ...
rknn_init ...
model input num: 1, output num: 2
input tensors:
  index=0, name=Preprocessor/sub:0, n_dims=4, dims=[1, 300, 300, 3], n_elems=270000, size=270000, fmt=NHWC, type=INT8, qnt_type=AFFINE, zp=0, scale=0.007843
output tensors:
  index=0, name=concat:0, n_dims=4, dims=[1, 1917, 1, 4], n_elems=7668, size=7668, fmt=NCHW, type=INT8, qnt_type=AFFINE, zp=50, scale=0.090787
  index=1, name=concat_1:0, n_dims=4, dims=[1, 1917, 91, 1], n_elems=174447, size=174447, fmt=NCHW, type=INT8, qnt_type=AFFINE, zp=58, scale=0.140090
rknn_run
loadLabelName
ssd - loadLabelName ./model/coco_labels_list.txt
loadBoxPriors
person @ (106 245 216 535) 0.994422
bus @ (87 132 568 432) 0.991533
person @ (213 231 288 511) 0.843047

RTENOTITLE


rknn_mobilenet_demo

cd /tools/test/adv/npu2/rknn_mobilenet_demo
./rknn_mobilenet_demo model/mobilenet_v1.rknn model/cat_224x224.jpg

model input num: 1, output num: 1
input tensors:
  index=0, name=input, n_dims=4, dims=[1, 224, 224, 3], n_elems=150528, size=150528, fmt=NHWC, type=INT8, qnt_type=AFFINE, zp=0, scale=0.007812
output tensors:
  index=0, name=MobilenetV1/Predictions/Reshape_1, n_dims=2, dims=[1, 1001, 0, 0], n_elems=1001, size=1001, fmt=UNDEFINED, type=INT8, qnt_type=AFFINE, zp=-128, scale=0.003906
rknn_run
 --- Top5 ---
283: 0.468750
282: 0.242188
286: 0.105469
464: 0.089844
264: 0.019531

rknn-toolkit2

Tool Introduction

RKNN-Toolkit2 is a development kit that provides users with model conversion, inference and performance evaluation on PC platforms. Users can easily complete the following functions through the Python interface provided by the tool:

  1. Model conversion: support to convert Caffe / TensorFlow / TensorFlow Lite / ONNX / Darknet / PyTorch model to RKNN model, support RKNN model import/export, which can be used on Rockchip NPU platform later.
  2. Quantization: support to convert float model to quantization model, currently support quantized methods including asymmetric quantization (asymmetric_quantized-8). and support hybrid quantization.
  3. Model inference: Able to simulate NPU to run RKNN model on PC and get the inference result. This tool can also distribute the RKNN model to the specified NPU device to run, and get the inference results.
  4. Performance & Memory evaluation: distribute the RKNN model to the specified NPU device to run, and evaluate the model performance and memory consumption in the actual device.
  5. Quantitative error analysis: This function will give the Euclidean or cosine distance of each layer of inference results before and after the model is quantized. This can be used to analyze how quantitative error occurs, and provide ideas for improving the accuracy of quantitative models.
  6. Model encryption: Use the specified encryption method to encrypt the RKNN model as a whole.


System Dependency

OS Version : Ubuntu18.04(x64) / Ubuntu20.04(x64)

Python Version : 3.6 / 3.8

Python library dependencies :

Python 3.6 

cat rknn-toolkit2/doc/requirements_cp36*.txt
# if install failed, please change the pip source to 'https://mirror.baidu.com/pypi/simple'

# base deps
numpy==1.19.5
protobuf==3.12.2
flatbuffers==1.12

# utils
requests==2.27.1
psutil==5.9.0
ruamel.yaml==0.17.4
scipy==1.5.4
tqdm==4.64.0
bfloat16==1.1
opencv-python==4.5.5.64

# base
onnx==1.9.0
onnxoptimizer==0.2.7
onnxruntime==1.10.0
torch==1.10.1
torchvision==0.11.2
tensorflow==2.6.2

Python3.8 

cat rknn-toolkit2/doc/requirements_cp38*.txt
# if install failed, please change the pip source to 'https://mirror.baidu.com/pypi/simple'

# base deps
numpy==1.19.5
protobuf==3.12.2
flatbuffers==1.12

# utils
requests==2.27.1
psutil==5.9.0
ruamel.yaml==0.17.4
scipy==1.5.4
tqdm==4.64.0
bfloat16==1.1
opencv-python==4.5.5.64

# base
onnx==1.9.0
onnxoptimizer==0.2.7
onnxruntime==1.10.0
torch==1.10.1
torchvision==0.11.2

 


Installation

Create virtualenv environment. If there are multiple versions of the Python environment in the system, it is recommended to use virtualenv to manage the Python environment. Take Python3.6 for example:

  • Install virtualenv 、Python3.6 and pip3 
sudo apt-get install virtualenv
sudo apt-get install python3 python3-dev python3-pip
  • Install dependencies 
sudo apt-get install libxslt1-dev zlib1g zlib1g-dev libglib2.0-0 libsm6 libgl1-mesa-glx libprotobuf-dev gcc
  • Install requirements_cp36-*.txt 
virtualenv -p /usr/bin/python3 venv 
source venv/bin/activate 
sed -i 's|bfloat16==|#bfloat16==|g' rknn-toolkit2/doc/requirements_cp36-*.txt
pip3 install -r rknn-toolkit2/doc/requirements_cp36-*.txt
sed -i 's|#bfloat16==|bfloat16==|g' rknn-toolkit2/doc/requirements_cp36-*.txt
pip3 install -r rknn-toolkit2/doc/requirements_cp36-*.txt
  • Install RKNN-Toolkit2 
pip3 install rknn-toolkit2/packages/rknn_toolkit2*cp36*.whl
  • Check whether RKNN-Toolkit2 install successfully，press ctrl+d to exit 
python3 
>>> from rknn.api import RKNN 
>>>

If install successfully, there is no error information.

Here is one of the failed informations: 

>>> from rknn.api import RKNN 
Traceback (most recent call last):
  File "<stdin>"，line 1，in <module>
ImportError: No module named 'rknn'

Model Conversion Demo

Here gives an example to show how to convert tflite model(mobilenet_v1_1.0_224.tflite) to RKNN model (mobilenet_v1.rknn) ON PC. 

cd rknn-toolkit2/examples/tflite/mobilenet_v1
python3 test.py


W __init__: rknn-toolkit2 version: 1.4.0-22dcfef4
--> Config model
W config: 'target_platform' is None, use rk3566 as default, Please set according to the actual platform!
done
--> Loading model
...  ...
done
--> Building model
Analysing : 100%|█████████████████████████████████████████████████| 58/58 [00:00<00:00, 2739.15it/s]
Quantizating : 100%|████████████████████████████████████████████████| 58/58 [00:00<00:00, 95.72it/s]
... ...
I rknn buiding done
done
--> Export rknn model
done
--> Init runtime environment
W init_runtime: Target is None, use simulator!
done
--> Running model
Analysing : 100%|█████████████████████████████████████████████████| 60/60 [00:00<00:00, 2740.54it/s]
Preparing : 100%|██████████████████████████████████████████████████| 60/60 [00:00<00:00, 136.33it/s]
mobilenet_v1
-----TOP 5-----
[156]: 0.9345703125
[155]: 0.0570068359375
[205]: 0.00429534912109375
[284]: 0.003116607666015625
[285]: 0.00017178058624267578

done
Retrieved from "https://ess-wiki.advantech.com.tw/wiki/index.php?title=RK_Platform_NPU_SDK&oldid=30912"