Model Creator

Image Classification

Image classification uses the transfer learning algorithm to perform minute-level learning training on hundreds of images in specific fields (such as vehicles and animals) based on the base classification model with good generalization capabilities, and can automatically generate a model for image classification. The generated model can automatically identify the category to which the image belongs. HMS Toolkit also allows you to generate API files for the model at the click of a button and to utilize a demo project for quickly verifying and calling the image classification model in your app.

Resource Preparations

Before using the image classification capability, prepare image resources for training as required.

Model Training

1. In Coding Assistant, go to AI > Model Creator.
2. Select Image as the transfer learning capability.
3. Set required parameters, then click Confirm.
   NOTE

   After you click Confirm, HMS Toolkit checks whether the training environment is ready. If the environment is not ready, HMS Toolkit automatically downloads and configures the environment.

   • Operation type: Select New Model.
   • Model Deployment Location: Select Deployment Cloud. It is supported on all phones running Android 4.1 or later. Huawei phones with built-in NPUs can use hardware to accelerate inference.

   

4. Drag or add the image classification folders to the Please select train image folder area, then set Output model file path and train parameters. Retain the default values of the parameters in Train parameter. If you have extensive experience in deep learning development, you can modify the parameter settings to improve the accuracy of the image recognition model.
   • Iteration count: The default value is 100.
   • Learning rate: The default value is 0.01.

   

5. Click Create Model to start training and generate an image classification model.
6. After the model is generated, view the model learning result (training precision and verification precision), corresponding learning parameters, and training data.

   

7. After the model training is complete, verify the model by adding the image folders in Please select test image folder under Add test image. The tool automatically uses the trained model to perform the test and displays the test result.

   

8. When the model is ready, click either Generate API or Generate Demo.
   • Generate API: HMS Toolkit automatically adds the dependency package, model file, and API sample source file of the trained model to the corresponding project folders, and adds dependencies to the build.gradle file of the module. You can then call the corresponding API in your app to integrate the image classification capability.

     

   • Generate Demo: HMS Toolkit automatically generates a demo project, which automatically integrates the trained model. You can directly build and run the demo project to generate an APK file, and run the file on the simulator or real device to check the image classification performance.

Text Classification

Text classification uses the transfer learning algorithm to perform minute-level learning training on text data in specific fields (such as sports, entertainment, and education) based on the base classification model with good generalization capabilities, and can automatically generate a model for text classification. The generated model can automatically identify the category to which the text belongs. HMS Toolkit also allows you to generate API files for the model at the click of a button and to utilize a demo project for quickly verifying and calling the text classification model in your app.

Resource Preparations

Model Training

1. In Coding Assistant, go to AI > Model Creator.
2. Select Text as the transfer learning capability.
3. Set Operation type to New Model and Model Deployment Location to Load From Obs Cloud, then click Confirm.
   NOTE

   After you click Confirm, HMS Toolkit checks whether the training environment is ready. If the environment is not ready, HMS Toolkit automatically downloads and configures the environment.

   

4. Drag or add the classified text resources (JSON text or TXT category folders) to the Please select train text dataset area. If the text resources are not encoded in UTF-8 format, the following dialog box is displayed. Click OK to convert the resources. (Currently, only the resources in Unicode or GBK format can be converted). After the conversion is complete, select a file name and file type, and click Confirm.

   

   Set Language type, Output model file path, and Train parameter. Retain the default values of the parameters in Train parameter. If you have extensive experience in deep learning development, you can modify the parameter settings to improve the accuracy of the text recognition model.

   • High precision mode: If this option is selected, a high-precision training model is generated.
   • Iteration count: The default value is 1000.
   • Learning rate: The default value is 0.01.

   

5. Click Create Model to start training and generate a text classification model.
6. After the model is generated, view the model learning result (training precision and verification precision), corresponding learning parameters, and training data.

   

7. After the model training is complete, you can verify the model by adding the text resources (JSON text or TXT category folders) in Please select test text dataset under Add test text. The tool automatically uses the trained model to perform the test and displays the test result.
8. When the model is ready, click either Generate API or Generate Demo.
   • Generate API: HMS Toolkit automatically adds the dependency package, model file, and API sample source file of the trained model to the corresponding project folders, and adds dependencies to the build.gradle file of the module. You can then call the corresponding API in your app to integrate the text classification capability.

     

   • Generate Demo: HMS Toolkit automatically generates a demo project, which automatically integrates the trained model. You can directly build and run the demo project to generate an APK file, and run the file on the simulator or real device to check the text classification performance.

Weakly Supervised Object Detection

Weakly supervised object detection uses the transfer learning algorithm to perform minute-level learning and training on easy-to-identify images of specific categories based on the base classification model with good generalization capabilities, analyze image features, extract objects from the images from the recognition model, and mark them. HMS Toolkit also allows you to generate API files for the model at the click of a button and to utilize a demo project for quickly verifying and calling the image classification model in your app.

Resource Preparations

Model Training

1. In Coding Assistant, go to AI > Model Creator.
2. Select Object as the transfer learning capability.
3. Set required parameters, then click Confirm.
   NOTE

   After you click Confirm, HMS Toolkit checks whether the training environment is ready. If the environment is not ready, HMS Toolkit automatically downloads and configures the environment.

   • Operation type: Select New Model.
   • Model type: Select a model type as required.
     • Multi Object Detection: Indicates weakly supervised multi-object detection.
     • Weak Supervised Single Object Detection: Indicates weakly supervised single-object detection.

   

4. Drag or add the image classification folders to Please select train image folder.
   • When Model type is set to Multi Object Detection, you need to set Augmented dataset path, Output model file path, Train parameter, and Augmented parameter. Retain the default values of the parameters in Train parameter and Augmented parameter. If you have extensive experience in deep learning development, you can modify the parameter settings to improve the accuracy of the image recognition model.
     • Iteration count: The default value is 5000.
     • Learning rate: The default value is 0.001.
     • Batch size: The default value is 16.
     • Saturation: image saturation. The value can be an integer or a number with two decimal places at most. The default value is 1.5.
     • Hue: image hue. The value can be an integer or a number with two decimal places at most. The default value is 0.1.
     • Exposure: image impressions. The value can be an integer or a number with two decimal places at most. The default value is 1.5.
     • Flip: image flipping. The value can only be 0 or 1 (default value).
     • Blur: image blur. The value must be an integer. The default value is 1.
     • Gaussian noise: The value must be an integer. The default value is 1.
     • Mosaic: The value must be an integer. The default value is 1.
     • Jitter: random adjustment range of the aspect ratio. The value can be an integer or a number with two decimal places at most ranging from 0 to 1. The default value is 0.1.
     • Augment Times: number of data augmentation times. The data set doubles each time the data is augmented. The value must be an integer. The default value is 10.

       

   • When Model type is set to Weak Supervised Single Object Detection, you need to set Output model file path and Train parameter. Retain the default values of the parameters in Train parameter. If you have extensive experience in deep learning development, you can modify the parameter settings to improve the accuracy of the image recognition model.
     • Iteration count: The default value is 1000.
     • Learning rate: The default value is 0.001.

     

5. Click Create Model to start training and generate an image classification model.
6. After the model is generated, view the model learning result (training precision and verification precision), corresponding learning parameters, and training data.

   

7. After the model training is complete, verify the model. You only need to add the image classification folder (weakly supervised single-object detection) or a single image (weak supervision multi-object detection) to Please select test image folder in Add test image. The tool automatically uses the trained model to perform the test and displays the test result.

   

8. When the model is ready, click either Generate API or Generate Demo.
   • Generate API: HMS Toolkit automatically adds the dependency package, model file, and API sample source file of the trained model to the corresponding project folders, and adds dependencies to the build.gradle file of the module. You can then call the corresponding API in your app to integrate the weakly supervised object detection capability.

     

   • Generate Demo: HMS Toolkit automatically generates a demo project, which automatically integrates the trained model. You can directly build and run the demo project to generate an APK file, and run the file on the simulator or real device to check the weakly supervised object detection performance.

Sound Classification

Sound classification uses the transfer learning algorithm to perform minute-level learning training on sound data in specific fields based on the base classification model with good generalization capabilities, and can automatically generate a model for sound classification. The generated model can automatically identify the category to which the sound belongs. HMS Toolkit also allows you to generate a demo project for the model at the click of a button and to utilize the project for quickly verifying and calling the sound classification model in your app.

Resource Preparations

Model Training

1. In Coding Assistant, go to AI > Model Creator.
2. Select Sound as the transfer learning capability.
3. Set Operation type to New Model and Model type to Deployment App, then click Confirm.
   NOTE

   After you click Confirm, HMS Toolkit checks whether the training environment is ready. If the environment is not ready, HMS Toolkit automatically downloads and configures the environment.

   

4. Drag or add the sound classification folders to the Please select train sound dataset area, and set train parameters.
   • Iteration count: The default value is 100.
   • Learning rate: The default value is 0.01.

     

5. Click Create Model to start training and generate a sound classification model.
6. After the model is generated, view the model learning result (training precision and verification precision), corresponding learning parameters, and training data.

   

7. After the model training is complete, verify the model by adding the sound resources to be tested to Please select test sound dataset under Add test sound. The tool automatically uses the trained model to perform the test and displays the test result.
8. When the model is ready, click either Generate API or Generate Demo.
   • Generate API: HMS Toolkit automatically adds the dependency package, model file, and API sample source file of the trained model to the corresponding project folders and adds dependencies to the build.gradle file of the module. You can then call the corresponding API in your app to integrate the sound classification capability.
   • Generate Demo: HMS Toolkit automatically generates a demo project, which automatically integrates the trained model. You can directly build and run the demo project to generate an APK file, and run the file on the simulator or real device to check the sound classification performance.