View source for AI/RKNN-Toolkit

{{Languages|AI/RKNN-Toolkit}}

    [[Home | Home]] > [[AI | AI Development]]  > [[AI/RKNN-Toolkit | RKNN Toolkit]]

=== Introduction ===

Rockchip offers the '''RKNN-Toolkit''' development kit for model conversion, forward inference, and performance evaluation.

Users can easily perform the following functions through the provided Python interface:

1) '''Model conversion''': support Caffe、Tensorflow、TensorFlow Lite、ONNX、Darknet model, support RKNN model import and export, and so the models can be loaded and used on the hardware platform.

2) '''forward inference''': user can simulate running the model on the PC and get the inference results, and run the model on the specified hardware platform RK3399Pro/RK1808 and get the inference results.

3) '''performance evaluation''': user can simulate running the model on a PC to get both the total time spent on the model and the time-consuming information of each layer. User can also run the model on the specified hardware platform RK3399Pro/RK1808 by online debugging, and get both the total time of the model running on the hardware and the time-consuming information of each layer.

This chapter mainly explains how to perform model conversion on the RK3399Pro/RK1808 development board. For other function descriptions, please refer to the RKNN-Toolkit User Guide: "'''RKNN-Toolkit User Guide_V*.pdf'''".

=== Installation preparation ===

    sudo dnf install -y cmake gcc gcc-c++ protobuf-devel protobuf-compiler lapack-devel
    sudo dnf install -y python3-devel python3-opencv python3-numpy-f2py python3-h5py python3-lmdb  python3-grpcio
    pip3 install scipy-1.2.0-cp36-cp36m-linux_aarch64.whl
    pip3 install onnx-1.4.1-cp36-cp36m-linux_aarch64.whl
    pip3 install tensorflow-1.10.1-cp36-cp36m-linux_aarch64.whl

After installing the above basic package, install the rknn-toolkit wheel package.
RKNN wheel package and other Python wheel packages can be downloaded from [https://rockchips-my.sharepoint.com/:f:/g/personal/addy_ke_rockchips_onmicrosoft_com/ErFTw5A6uYtBvtg-1ksWIQEByi5eoqA4hIWHh7xPD8SRpg?e=q3kbhe OneDrive].

Since pip does not have a ready-made aarch64 version of the scipy and onnx wheel packages, we have provided a compiled wheel package.
If you want the latest version of the wheel package or find a problem with the pre-compiled wheel package, you can use pip to install it yourself. This will compile and install the wheel package. It will take a long time and you need to wait patiently.

    pip3 install scipy
    pip3 install onnx

If the installation encounters an error, please install the corresponding software package according to the error message.

=== Model Conversion ===

==== API call flow ====

[[File:Rknn-conv-call.png]]

==== Example ====

<syntaxhighlight lang="Python" line='line'>
from rknn.api import RKNN

INPUT_SIZE = 64

if __name__ == '__main__':
    rknn = RKNN()   # Create an RKNN execution object

    '''
    Configure model input for NPU preprocessing of input data
    channel_mean_value='0 0 0 255', when runing forward inference, the RGB data will be
    converted as follows (R - 0) / 255, (G - 0) / 255, (B - 0) / 255,
    The RKNN model automatically performs the mean and normalization.
    reorder_channel='0 1 2' , used to specify whether to adjust the image channel order, 
    set to 0 1 2, means no adjustment according to the input image channel order.
    reorder_channel='2 1 0' , indicates that 0 and 2 channels are exchanged.
    If the input is RGB, it will be adjusted to BGR. If it is BGR will be adjusted to RGB
    Image channel order is not adjusted
    '''

    rknn.config(channel_mean_value='0 0 0 255', reorder_channel='0 1 2')

    '''
    load TensorFlow model
    tf_pb='digital_gesture.pb' specify the TensorFlow model to be converted
    inputs specify the input node in the model
    outputs specify the output node in the model
    input_size_list specify the size of the model input
    '''

    print('--> Loading model')
    rknn.load_tensorflow(tf_pb='digital_gesture.pb',
                         inputs=['input_x'],
                         outputs=['probability'],
                         input_size_list=[[INPUT_SIZE, INPUT_SIZE, 3]])
    print('done')

    '''
    Create a parsing pb model
    do_quantization=False do not to be quantified
    Quantization will reduce the size of the model and increase the speed of the operation,
    but there will be loss of precision.
    '''

    print('--> Building model')
    rknn.build(do_quantization=False)
    print('done')
    rknn.export_rknn('./digital_gesture.rknn')  # Export and save rknn model file
    rknn.release()  # Release RKNN Context
</syntaxhighlight>

=== Model Inference ===

==== API call flow ====

[[File:Rknn-infe-call.png]]

==== Example ====

<syntaxhighlight lang="Python" line='line'>
import numpy as np
from PIL import Image
from rknn.api import RKNN


# Analyze the output of the model to get the most probable gesture and corresponding probability
def get_predict(probability):
    data = probability[0][0]
    data = data.tolist()
    max_prob = max(data)
    return data.index(max_prob), max_prob


def load_model():
    rknn = RKNN()  # Create an RKNN execution object
    print('-->loading model')
    rknn.load_rknn('./digital_gesture.rknn')  # Load RKNN model
    print('loading model done')
    print('--> Init runtime environment')
    ret = rknn.init_runtime(host='rk3399pro')  # Initialize the RKNN runtime environment
    if ret != 0:
        print('Init runtime environment failed')
        exit(ret)
    print('done')
    return rknn


def predict(rknn):
    im = Image.open("../picture/6_7.jpg")   # load image
    im = im.resize((64, 64), Image.ANTIALIAS)  # Image resize to 64x64
    mat = np.asarray(im.convert('RGB'))    # Convert to RGB format
    outputs = rknn.inference(inputs=[mat])   # Run forward inference and get the inference result
    pred, prob = get_predict(outputs)     # Transform the inference results into visual information
    print(prob)
    print(pred)


if __name__=="__main__":
    rknn = load_model()
    predict(rknn) 
    rknn.release()
</syntaxhighlight>