Shopping Cart | 中文网站
 
 
Events
Trade Shows
Seminars
Product Updates
 
          Home > News Center > Product Updates
 

Build Your Applications with FZ3 Card and Baidu Brain PaddlePaddle

 
2020-7-29   16:24:38     Click: 697
 

FZ3 Card is a deep learning accelerator card launched by MYIR while cooperating with Baidu. It is based on Xilinx Zynq UltraScale+ ZU3EG MPSoC which integrates 64-bit quad-core ARM Cortex-A53, GPU and FPGA, thus comes with multi-core processing capability, FPGA programmable capability and hardware decoding capability for video stream.



The FZ3 Card has built-in deep learning soft core based on Linux OS and Baidu PaddlePaddle deep learning AI (Artificial Intelligence) framework which is fully compatible to use Baidu Brain’s model resources and AI development tools like EasyDL, AI Studio and EasyEdge to enable developers and engineers to quickly train-deploy-reasoning models. Provided with these hardware capabilities and software resources, the FZ3 Card reduces the threshold of development validation, product integration, scientific research and teaching significantly.



How to build your applications with FZ3 Card and Baidu Brain PaddlePaddle? Below there are some guidances.

1. Obtaining models

Currently Paddle-Mobile only supports Paddle trained models. If the models in your hands are different types of models, you need to perform model conversion before running. Verified networks include resnet, Inception, ssd, mobilenet, etc.

1.1 Training model

If you don't have a model, you can use the model in sample or train the model by yourself.

1. Train model through Paddlepaddle open source deep learning framework. Reference for detailed use: PaddlePaddle

2. Through model through AI studio platform training model, reference for detailed use: AI Studio

3. You may upload labeled data on EasyDL platform to train the model. Reference for detailed use: EasyDL

1.2 Converting model

1. If you already have a Caffe model, MYIR provides a corresponding conversion tool to help you convert it to a Paddle model. Reference for detailed use: X2Paddle_caffe2fluid

2. If you already have a TensorFlow model, MYIR provides a corresponding conversion tool to help you convert it to a Paddle model. Reference for detailed use: X2Paddle_tensorflow2fluid


2. Connecting video data source

2.1 Video data input via USB protocol video

You can choose the UVC USB camera as the video source. Insert USB camera into USB interface of FZ3.

2.2 Video data input via BT1120 protocol

You can select the webcam with BT1120 video data output from Hisilicon. Connect the BT1120 interface of FZ3 through FPC cable. For the specific pin-description, please refer to the hardware description.

2.3 Video data input via Mipi protocol

You can choose a suitable Mipi camera as the video source and connect to the Mipi interface of FZ3 through FPC.

2.4 Video data input via GigE protocol

You can choose the GIGE camera that supports the Linux system, and contact our company to adapt the official SDK of the camera. The hardware is connected to the FZ3 network port.

3. Load device driver

Using FZ3's acceleration function, the prediction library will calculate the op with a large amount of calculation through the driver to call

FPGA to calculate. The driver needs to be loaded before running your own application. The compiled driver is in the directory

/home/root/workspace/driver, providing two versions: the first version with no log output and the second version with log output.

Load driver

insmod /home/root/workspace/driver/fpgadrv.ko

Uninstall driver (In normal circumstances, you don’t need to uninstall the driver. If you need to load the version with log output, you can uninstall it with the following command and then load this version)

rmmod /home/root/workspace/driver/fpgadrv.ko

Set the driver to be automatically loaded

1) Add a self-starting script to the system

// Open the startup directory

cd /etc/init.d/

// Create a new startup script and edit it, the name can be customized

vim eb.sh

      Script content

chmod +x /home/root/workspace/driver/fpgadrv.ko

insmod /home/root/workspace/driver/fpgadrv.ko

2) Establish soft links

cd /etc/rc5.d/

ln -s /etc/init.d/eb.sh S99eb

3) Change script permissions

chmod +x /etc/init.d/eb.sh

reboot


4. Using the prediction library

FZ3 supports Paddle-Moblie prediction library. The compiled prediction library is in

/home/root/workspace/paddle-mobile.

How to use it: Copy the header file and dynamic library of the prediction library to your own application. You can also refer to the sample MYIR provides.

Paddle-Moblie source code: https://github.com/PaddlePaddle/paddle-mobile

5. Create applications

5.1 Add prediction library

Copy the dynamic libraries and header files in /home/root/workspace/paddle-mobilie/ to your project. Add a reference to the Paddle-Mobile library in CmakeLists.txt
set(PADDLE_LIB_DIR "${PROJECT_SOURCE_DIR}/lib" )
set(PADDLE_INCLUDE_DIR "${PROJECT_SOURCE_DIR}/include/paddle-mobile/" )
include_directories(${PADDLE_INCLUDE_DIR}) LINK_DIRECTORIES(${PADDLE_LIB_DIR})
...
target_link_libraries(${APP_NAME} paddle-mobile)

5.2 Add model

Copy your trained model to your project.

5.3 Add prediction data sources

You can select pictures and camera data as the source of prediction data. To use the camera, you need to insert the corresponding camera.

5.3.1 USB camera

1) After plugging in the camera, check the device access through ls /dev/video*. Display as below means pass:

config.dev_name = "/dev/video2";

/dev/video2 outputs YUV data for USB camera. When the application prompts that the device cannot be found, you can modify src/video_classify.cpp or src/video_detection.cpp. Check the camera connectivity through the video tool in /home/root/workspace/tools.

// src/video_classify.cpp line 169

config.dev_name = "/dev/video2";

2) The camera resolution can be modified

// src/video_classify.cpp line 170

config.width = 1280;

config.height = 720;

3) Run the video tool

// Read the USB camera data, collect a picture and save it to the local

cd /home/root/workspace/tools/video

./v4l2demo -i /dev/video2 -j -n 1

// If in doubt, check the help
./v4l2demo -h


After executing the program, a .jpg file will be generated in the directory build, you can check whether the picture is correct or not. If no picture is generated, please check if the USB device is recognized.

5.3.2 BT1120 ipc camera

After FZ3 receives the original data through BT1120 protocol and reasonig, it can transmit the result back to ipc through the serial port or spi (For interface description of BT1120, serial port, spi please refer to the hardware description). The frame number of the image can be carried in the pixel data.

After inserting the camera, use the video tool in /home/root/workspace/tools to check the camera connectivity.

1) View the device, the directory is /dev/video1 in normal circumstances

ls /dev/video*

/dev/video0 /dev/video1

2) Set the camera parameters

media-ctl -v --set-format '"a0010000.v_tpg":0 [RBG24 1920x1080 field:none]'


3) Run the video tool

// Read BT1120 camera data, collect a picture and save to local

cd /home/root/workspace/tools/video

./v4l2demo -i /dev/video1 -j -n 1

// If in doubt, check the help

./v4l2demo -h

After executing the program, a .jpg file will be generated in the directory build, you can check if the picture is correct. If no picture is generated, check whether the BT1120 cable is connected correctly.

5.4 Call the prediction library to load the model and use the prediction data

5.4.1 Initialize the model

Predictor _predictor_handle = new Predictor();

_predictor_handle->init(model, {batchNum, channel, input_height, input_width}, output_names);


5.4.2 Prepare data

1) Scale the picture to the specified size. If the neural network requires a fixed size, the picture needs to be scaled to that fixed size.

2) Image preprocessing (Minus mean value, floating point conversion, normalization, etc.).

3) Output data. Because FZ3 uses NHWC format, usually the data from the video is in NHWC format, so NHWC->NCHW conversion is not required.


5.4.3 Predict data

bool predict(const float* inputs, vector &outputs,vector&output_shapes);


Do you have any idea to use MYIR’s FZ3 Card with Baidu Paddle to build your applications now? Some typical applications are as below for reference.



Please get more information about the FZ3 Card from MYIR’s website:

http://www.myirtech.com/list.asp?id=630


About MYIR

MYIR Tech Limited is a global provider of ARM hardware and software tools, design solutions for embedded applications. We support our customers in a wide range of services to accelerate your pace from project to market.

We sell products ranging from board level products such as development boards, single board computers and CPU modules to help with your evaluation, prototype, and system integration or creating your own applications. MYIR also provide our customers charging pile billing control units, charging control boards and relative solutions inside of China. Our products are used widely in industrial control, medical devices, consumer electronic, telecommunication systems, Human Machine Interface (HMI) and more other embedded applications. MYIR has an experienced team and provides custom services based on many processors (especially ARM processors) to help customers make your idea a reality.


More information about MYIR can be found at: www.myirtech.com





 
Prev:  MYIR Introduces FZ5 EdgeBoard AI Box and Accelerator Card Next:  Evaluation of MYIR's i.MX 8M Mini based SOM and EVM
 

Room 04, 6th Floor, Building No.2, Fada Road, Yunli Smart Park, Bantian, Longgang District, Shenzhen, Guangdong, China 518129
E-mail: sales@myirtech.com or myirtech@yahoo.com  Phone: +86-755-22984836  Fax: +86-755-25532724
Copyright © MYIR Tech Limited 2011-2020 all rights reserved.
myirtech_facebook myirtech_Twitter myirtech_linkedin myirtech_youtube