Document Understanding 用户指南

迁移传统项目

1. 从传统项目或基于 AI Center 的项目中导出数据集。
2. 将数据集导入新式项目。

当前限制​

• Currently, importing datasets larger than 5000pages is not supported. Only the initial 5000 pages will be successfully imported, with any additional pages failing to do so. For example, if your dataset consists of 4999 pages and you try to import a document of 4 pages, the process will not succeed.
• 批次名称和相应的批次结果当前不可用。如果您的数据已组织为批次，则系统不会显示此信息，但会保存此信息。
• 不支持从 AI Center 导出。 仅支持从 Document Manager 导出。

从传统项目导出数据集​

1. 导航到要迁移的传统项目并将其打开。
2. Go to the document type you want to export and select Open document type.

3. From the Filter documents drop-down list, select Training and validation set.

4. 选择“导出” 。
5. Leave Current search results selected and fill in a name for your export job.
6. Select Download.

从基于 AI Center 的项目中导出数据集​

1. Open AI Center and navigate to the Data Labeling page.

2. Select the Data Labeling Session you want to migrate.

   

3. Once Document Manager is open, from the Filter documents drop-down list, select Training and validation set.

4. 选择“导出” 。
5. Leave Current search results selected and fill in a name for your export job.
6. Select Download.

导入数据集​

1. 导航到要向其中导入数据的项目，并将其打开。
2. Select Add document type and create a new custom document type.

3. On the new custom document type, select Upload and choose the zip file of the classic project you exported. Wait for the upload to finish.
   备注：

   不支持从 AI Center 导出。 仅支持从 Document Manager 导出。

模型训练​

• Initialization: Machine learning uses optimization methods that need initial guesses to trigger the optimization algorithms. Different initial guesses during each training could lead to various outcomes due to the unpredictable nature of these algorithms.
• Random state: Some algorithms use randomness in their operations. For instance, when training a neural network, procedures like stochastic gradient descent and mini-batch gradient descent introduce randomness. Therefore, even with identical initial model parameters and datasets, the performance of models may vary in different runs.
• Regularization: Certain algorithms include a penalty term that encourages the model to maintain smaller weights. Due to the randomness involved, the model could operate with a different weight set each time.

在文档类型管理器中更改基本模型​

1. Select the three-dot menu from your custom document type and choose Document type manager.

   

2. Navigate to the Settings tab.

3. Select the desired model from the Base model drop-down list.

   

4. After making your selection, select Save. To exit, select Back.

导出类型​

导入架构​

1. Create a custom document type in the Build section.
2. 导入包含架构的 zip 文件。
   备注：

   • 架构导入仅限于没有预先存在架构的自定义文档类型。
   • 如果您将架构导入已包含架构的文档类型，则导入将失败。

Migrate the automation workflow​

Replace the ML Extractor Activity​

1. In your Studio project, open the Data Extraction Scope activity.
2. Remove the existing ML Extractor Activity.
3. Add a Document Understanding Project Extractor inside the Data Extraction Scope.
4. Select Get or refresh extractor capabilities to open the configuration wizard.
5. Under Design time credentials, enter your App Id, App Secret, and Tenant Url.
6. Select Get Projects to load the list of available modern projects.
7. For Project, select your desired modern project from the dropdown list.
8. For Version, select a deployed version of the project. Alternatively, select a Tag linked to a specific version. Version and Tag are mutually exclusive.
9. Select Get Capabilities.
10. Make sure Update Activity Arguments is checked.

Migrate classification​