#
Model retraining
In order to improve the effectiveness of the trained model, we can subject it to a retraining process. To start, click the Re-train
button on the selected model tile.
Retraining process consists of three stages:
- The first to edit dataset;
- The second to merge object classes;
- The third relates to model configuration.
#
Editing datasets
In the first stage, we choose the datasets that we want to use for the retraining process.
By default, the datasets that were used in the previous (original) training session are selected. You can delete them if you want to use only the new datasets.
#
Merging object classes
By clicking the Next
button, we go to the category merge view. It is slightly different from the original merge view:
- Categories from the previous training are automatically assigned to buckets and marked in gray
- If you selected new categories, they will be assigned to the automatically generated "Unassigned" bucket
- If the name of a new category matches the name of an existing bucket, it will be automatically assigned to it
To be able to start retraining, we need to move all dataset categories from the "Unassigned" bucket to the appropriate buckets. Only then the "Unassigned" bucket will be automatically deleted and the Re-train
button will be unlocked. Each bucket must always contain at least one dataset category.
#
Model configuration
Clicking Next
takes you to the datasets view.
In the last stage, define the name of the retrained model and the number of planned epochs (in the case of classification) or iterations (in the case of object detection).
The model name
must meet the same validation requirements as for regular training (See training).