Document Understanding Auto-Training Is Failing

News Knowledge Base
, , ,
1

Document Understanding Auto-Training is Failing.

Issue Description: Document Understanding Auto-Training is Failing

Root Cause: Most issues with auto-retraining are caused by misconfigurations.


Diagnosing/Resolving:

  1. First, review the docs to make sure the correct steps are being taken. See: The Auto-Fine-tuning Loop (Public Preview) ( Ensure to go to the version specific documents)
  2. Login To AI Center
  3. Check the package version. Document Understand Package Versions of 22.5.X have a known issue with auto retraining. If using one of these packages, either upgrade to a newer release (if one is out) or downgrade to 22.4.X.
    1. AI Center -> ML Packages and click the package in question. The page should have version information.
  4. Validate the Dataset.
    1. Go to the dataset that is being used for the pipeline.
      1. AI Center->DataSets and select the dataset.
    2. In the Dataset folder there should be a folder called export. Select this folder.
    3. In the folder there should be folders titled "auto-export-XXX"
    4. Additionally there should be a file called latest.txt.
    5. Here is an example of what we are looking for:
      1. image.png
    6. Once it has been validated that this files exist, please take a screenshot to share with support if a ticket is opened.
    7. If the dataset does not contain the auto-export folders that means that auto-retraining has not be enabled. If the latest.txt file is not present, it could be it was deleted by a user. If that is the case it should be recreated on the next export. For enabling scheduled exports see: The Auto-Fine-tuning Loop (Public Preview)
  5. Validate the Pipeline settings.
    1. AICenter -> Pipelines
    2. When creating a pipeline, its important that the following rules are followed:
      1. For the Input dataset, the export folder should be selected.
      2. For the evaluation dataset, a specific export is select (This should only change when your model changes. So from run to run it will usually be the same.)
      3. The auto-retraining variable is set to True.
      4. Here is an example
        1. In this example the Input dataset is set to EvalTest/export (i.e. /, and export folder will always be named 'export'.
        2. Evaluation dataset is set to the initial evaluation dataset that was made when labeling data, EvalTest/export/Test_22-08-32T214352
        3. Here is a screenshot.image.png
  6. With the proper steps for creating a pipeline in mind, go to the failed pipeline and make sure the above rules were followed.
    1. AI Center -> Pipelines
    2. Click the Failed pipeline.
    3. Validate the above settings. For the environment variable, make sure to click on the down arrow for parameters. Make sure to take a screenshot to share with support.
    4. Here is an example:
      1. image.png
    5. If the settings are not correct, then the pipeline needs to be recreated with the correct settings.
  7. Finally check the pipeline logs.
    1. On the failed pipeline, download the logs.
      1. Click the download button on the bottom of the pipeline page.
      2. image.png
    2. In the logs search for: 'Auto-retraining enabled, checking latest.txt'
    3. If the above is not present it means that auto-training is enabled correctly. If it is not, then it means either:
      1. Auto training was not enabled correctly
      2. Or it could be a bug.
    4. If the line is present, that mostly likely means that another issue is causing the failure. We recommend looking at the error in that case and checking other KB articles, etc.
  8. If this article does not resolve the issue, open a ticket with UiPath and include the following:
    1. Screenshot of the Dataset captured in the step for validating the dataset.
    2. Screenshot of the pipeline settings in the step for validating the pipeline settings.
    3. The Pipeline logs.
    4. Include the ML package version that is being used.
    5. AI Center version. Or if on cloud include your support ID: Managing Organization Settings and the tenant name.