This article is shared from Huawei Cloud Community "VSCode One-key Access Notebook Experience Algorithm Kit to Quickly Complete Water Meter Reading", author: HWCloudAI.

This example focuses on the real AI demand scenario and introduces the use process of VSCode's one-click access to the Notebook experience algorithm suite to quickly complete the water meter reading.

The algorithm development kit currently provides two sets of algorithm assets, self-developed (ivg series) and open source (mm series), which can be applied to tasks such as classification, detection, segmentation, and OCR. In this example, the self-developed segmentation algorithm (ivgSegmentation) and the open source OCR algorithm (mmOCR) will be combined to complete the water meter reading recognition project, and the algorithm development kit will be used to deploy it as an online service on HUAWEI CLOUD.

illustrate:

This case tutorial is only applicable to the "North China-Beijing 4" region, the new version of Notebook.

obs://{OBS bucket name}/water_meter_segmentation water meter dial segmentation data set

obs://{OBS bucket name}/water_meter_crop Water meter dial reading OCR recognition data set

illustrate:

Downloading datasets from AIGallery is free, but a small fee will be charged for storing datasets in OBS buckets. For details, please refer to the OBS price details page. Please clear resources and data in time after the use of the case is completed.

On the "ModelArts Console > Development Environment > Notebook (New)" page, create a Notebook based on the pytorch1.4-cuda10.1-cudnn7-ubuntu18.04 image and type GPU. For details, see the section Creating Notebook Instances.

In this case, you need to use VS Code to remotely connect to Notebook, and you need to enable SSH remote development.

Figure 1 Create a Notebook instance

1. The key file of the instance needs to be downloaded to the following local directory or its subdirectory:

Windows: C:\Users{{user}}

Mac/Linux: Users/{{user}}

2. In the ModelArts Console -> Development Environment Notebook, click "More > VS Code Access" in the "Operation" column.

If VS Code has been installed locally, click "Open" to enter the "Visual Studio Code" page.

If VS Code is not installed locally, please select "win" or "other" to download and install VS Code. For VS Code installation, please refer to Installing VS Code Software.

If the user has not installed the ModelArts VS Code plug-in before, an installation prompt will pop up at this time, please click "Install and Open" to install; if the plug-in has been installed before, there will be no such prompt, please skip this step and directly follow the next steps

The installation process is expected to take 1 to 2 minutes. After the installation is complete, a dialog box will pop up in the lower right corner. Please click "Reload Window and Open".

In the prompt that pops up, check "Don't ask again for this extension", and then click "Open".

3. Connect to the Notebook instance remotely.

1. After successful access, click File -> Open Folder on the VS Code page, and select the following folder to open

2. Create a new terminal

3. Execute in the work directory

command to create a project, and enter the project name according to the prompt, for example: water_meter. Then press Enter directly to select the default parameters, and choose to skip the asset installation step (choose 6).

4. Execute the following command to enter the project directory.

5. Execute the following command to copy the project data to Notebook.

illustrate:

The {obs_dataset_path} path is the dataset downloaded to OBS in Step1 Prepare Data, such as "obs://{OBS bucket name}/water_meter_segmentation" and "obs://{OBS bucket name}/water_meter_crop"

1. First install the ivgSegmentation suite.

If it prompts that the version of ivgSegmentation is incorrect, you can query the version through the command python manage.py list algorithm.

2. After installing the ivgSegmentation suite, enter the "./algorithms/ivgSegmentation/config/sample" folder in the project directory on the left side of the interface to view the currently supported segmentation models. Take sample as an example (the default algorithm of sample is deeplabv3), The folder includes config.py (algorithm shell configuration) and deeplabv3_resnet50_standard-sample_512x1024.py (model structure).

3. Dial segmentation only needs to distinguish the background and reading area, so it belongs to two classifications, and the configuration file needs to be modified according to the data set required by the project, as follows:

Modify the ./algorithms/ivgSegmentation/config/sample/config.py file.

Press Ctrl+S to save after modification.

4. Modify the ./algorithms/ivgSegmentation/config/sample/deeplabv3_resnet50_standard-sample_512x1024.py file.

5. Press Ctrl+S to save after modification.

6. In the water_meter project directory, install the deeplabv3 pre-training model.

7. Train the segmentation model. (GPU is recommended for training)

The trained model will be saved in the specified location, which is output/deeplabv3_resnet50_standard-sample_512x1024/checkpoints/ by default.

8. Verify the effect of the model.

After the model training is complete, you can calculate the model's indicators on the verification set, first modify the model location in the configuration file.

Modify ./algorithms/ivgSegmentation/config/sample/config.py.

9. Model reasoning.

Model reasoning can specify a certain picture, and infer the segmented area of ​​the picture, and visualize it. First, you need to specify the picture path that needs to be reasoned.

Modify ./algorithms/ivgSegmentation/config/sample/config.py

Execute the following command to infer the effect of the model.

The image path of the inference output is under ./output/deeplabv3_resnet50_standard-sample_512x1024.

10. Export the SDK.

The algorithm development kit supports exporting the model into a model SDK to facilitate downstream tasks such as model deployment.

1. First install the mmocr suite.

2. After installing the mmocr suite, the ./algorithms/mmocr/config/textrecog folder includes config.py (algorithm shell configuration), and the configuration file needs to be modified according to the required algorithm and data set path. The robust_scanner algorithm is taken as an example below.

Modify ./algorithms/mmocr/algorithm/configs/textrecog/robustscanner_r31_academic.py,

3. Install the robust_scanner pre-training model.

4. Train the OCR model.

When using mmcv for the first time, you need to compile mmcv-full. This process is slow, and you can directly use the official precompiled dependency package.

Precompiled package URL: https://download.openmmlab.com/mmcv/dist/cu101/torch1.6.0/index.html

Change the epoch (number of iterations) in ./algorithms/mmocr/config/textrecog/config.py to 2, as shown in the following figure:

The trained model will be saved in the specified location, which is output/${algorithm} by default.

5. Verify the effect of the model.

After the model training is complete, you can calculate the model's indicators on the verification set, first modify the model location in the configuration file.

Modify ./algorithms/mmocr/config/textrecog/config.py

6. Model reasoning.

Model reasoning can specify a certain picture, and infer the segmented area of ​​the picture, and visualize it. First, you need to specify the image path to be reasoned, and modify the algorithms/mmocr/config/textrecog/config.py file, as follows.

Modify ./algorithms/mmocr/algorithm/configs/textrecog/robust_scanner/config.py

The image path of the inference output is under output/robustscanner_r31_academic/vis

7. Export the SDK.

This demonstration only deploys the OCR service, including local deployment and online deployment. After the deployment is online, the deployment service is called to infer the local image and obtain the predicted reading of the water meter. To deploy online services, you need to specify an OBS bucket to save the files required for deployment.

1. Configure the OBS bucket in the algorithms/mmocr/config/textrecog/config.py file, namely obs_bucket=.

2. Execute the following command:

Click here to view the successfully deployed online services

After learning through this example and completing the process of creating an algorithm suite, if you no longer use it, it is recommended that you clear the relevant resources to avoid waste of resources and unnecessary costs.

Click to follow and learn about Huawei Cloud's fresh technologies for the first time~