English|中文

Classification Network Application (Python)

This application can run on the Atlas 200 DK to implement the inference function of the ResNet-18 network and output images with inference result labels and confidence.

The current application adapts to DDK&RunTime of 1.3.0.0 as well as 1.32.0.0 and later versions.

Prerequisites

Before deploying this sample, ensure that:

• Mind Studio has been installed.

• The Atlas 200 DK developer board has been connected to Mind Studio, the SD card has been created, and the build environment has been configured.

• The developer board is connected to the Internet over the USB port by default. The IP address of the developer board is 192.168.1.2.

Sample Deployment

You can use either of the following methods:

1. Quick deployment: visit https://github.com/Atlas200dk/faster-deploy.

   NOTE:

   • The quick deployment script can be used to quickly deploy multiple cases. Select classification-python.
   • The quick deployment script automatically completes code download, model conversion, and environment variable configuration. To learn about the detailed deployment process, select the common deployment mode. Go to 2. Common deployment.

2. Common deployment: visit https://github.com/Atlas200dk/sample-README/tree/master/sample-classification-python.

   NOTE:

   • In this deployment mode, you need to manually download code, convert models, and configure environment variables. After that, you will have a better understanding of the process.

Environment Deployment

1. Copy the application code to the developer board.

   Go to the root directory of the classification network application (python) code as the Mind Studio installation user, for example, $HOME/sample-classification-python, and run the following command to copy the application code to the developer board:

   scp -r ../sample-classification-python/ HwHiAiUser@192.168.1.2:/home/HwHiAiUser/HIAI_PROJECTS

   Enter the password of the developer board as prompted. The default password is Mind@123, as shown in Figure 1.

   Figure 1 Copying application code

   

   Log in to the host side as the HwHiAiUser user in SSH mode on Ubuntu Server where Mind Studio is located.

   ssh HwHiAiUser@192.168.1.2

   Switch to the root user. The default password of the root user on the Atlas DK developer board is Mind@123.

   su root

2. Configure the network connection of the developer board.

   Configure the network connection of the Atlas DK developer board by referring to https://github.com/Atlas200dk/sample-README/tree/master/DK_NetworkConnect.

3. Install the environment dependency.

   Configure the environment dependency by referring to https://github.com/Atlas200dk/sample-README/tree/master/DK_Environment.

Application Running

1. Switch to the HwHiAiUser user and go to the directory where the classification network application code is located.

   su HwHiAiUser

   cd ~/HIAI_PROJECTS/sample-classification-python/classificationapp

2. Run the application.

   Run the classify.py script to print the inference result on the execution terminal.

   Command example:

   python classify.py

   Figure 2 shows the inference result after the execution is successful.

   Figure 2 Successful inference

   

3. Query the execution result.

   The execution result is stored in resent18Result of the current directory. You need to run the following command on the Atlas200 DK to copy the result to the Ubuntu server to view the inference result image:

   scp -r username@host_ip:/home/username/HIAI_PROJECTS/sample-classification-python/classificationapp/resent18Result ~

   • username: user name of the developer board. The default value is HwHiAiUser.
   • host_ip: IP address of the developer board. Generally, the IP address is 192.168.1.2 for USB connection and 192.168.0.2 for network cable connection.

   Command example:

   scp -r HwHiAiUser@192.168.1.2:/home/HwHiAiUser/HIAI_PROJECTS/sample-classification-python/classificationapp/resent18Result ~

   This command copies the inference result to the home directory of the Mind Studio installation user. You can view the inference result directly.

Remarks

• The process of the classification network application (Python) is as follows:

  1. Read a JPEG image from the ImageNetRaw directory.
  2. Call OpenCV to resize the read JPEG image to 256 x 224 and convert it to YUV420SP.
  3. Send the converted YUV420SP image data to Matrix for inference. The demo uses the ResNet-18 network, and the inference result is the confidence of 1000 classes.
  4. Post-process the image. 1000 classification confidence levels are sorted, the highest confidence level and its classification label are selected, and the image is labeled. A labeled image is stored in the resnet18Result directory.

• The directory structure of the classification network application (Python) is described as follows:

  • ImageNetRaw: storing input images
  • classify.py: main program
  • imageNetClasses.py: labeling for 1000 classes of ImageNet
  • jpegHandler.py: JPEG image processing, such as resizing, color space conversion (CSC), and text labeling
  • models: storing model networks
  • resnet18Result: storing labeled images