Skip to main content
In the Self-Assessment section of the Results tab, you can monitor the training progress and adjust the training duration. The statistics displayed in this section will help you understand if the training is going on successfully or if changes should be made.

​
Training Progress Graph

The graph on the left shows the percentage of correctly extracted field values across all epochs. AD_Semistructured_Accuracy_Metrics This graph is updated after each epoch and reflects the actual accuracy of field extraction. The accuracy is calculated on a validation subset, taking into consideration recognition errors and some additional metrics.
Note: During the training process, tests are carried out on the validation set, which consists of 20% of the documents. After training, tests are carried out on all documents. Therefore, the results may differ. The more documents there are in total, the less the results should differ.

​
Interpreting the Graph

This graph can help you decide whether you should carry on with the training or stop training the activity:
  • If the accuracy is sufficiently high and is no longer increasing, then the training can be stopped.
  • If the accuracy is not sufficiently high but is still increasing, then the training should go on.
  • If the accuracy is not sufficiently high and is no longer increasing, then you should analyze the quality of extraction of individual fields (there may be fields in the documents that are too rare or that have been marked up incorrectly).
Note: Training does not affect the quality of recognition and incorrectly detected fields can still be trained.

​
How to Decide When to Stop Training Your Activity

By default, a Deep Learning activity is trained over 20 epochs. However, in certain cases you may want to reduce the training duration or stop the training manually. You may want to stop training if:
  • The accuracy percentage is sufficiently high and there have not been any significant changes in accuracy for several epochs in a row.
  • The accuracy percentage is low and there have not been any significant changes in accuracy for several epochs in a row (consider using a larger document set or make sure that all fields are well represented and marked up correctly).
  • You want to test the activity manually on all the documents (resume the training if you are not satisfied with the field extraction results).

​
How to Adjust the Training Duration

You can reduce the number of epochs or the time spent on training a document set any time during the training (such as document types with a small number of variants will take less time to train). To adjust the training duration, do the following:
  1. Go to the Self-Assessment section of the Results tab.
  2. Click the settings icon to the right of the number of epochs above the graph.
  3. Enter a new number of epochs and/or time in “hours:minutes” format and click Apply. The new limit must be greater than the number of elapsed epochs. The maximum number of epochs is 20. The training will stop after one of the limits is reached.
Note: You can stop the training at any time by clicking Stop Training. The training results for any unfinished epoch will be discarded.