In our project, a invoice recipe validation process, we are periodically adding new images to the models to improve classification as new samples become available.
So, every night, when we have collected a reasonable amount of new samples of classification errors (we manually audit some cases), we add the audited images to the model tags, retrain the model, and publish the new iteration as the default for predictions.
Everything is working very well so far, but we have some 20 iterations by now, with the most recent published.
My question is about the older iterations.
It is safe to delete the unpublished iterations? It will affect somehow the training of new iterations?
Keeping them have some impact on cost?
Based on the official doc, it is safe to delete iterations that are obsolete.
And based on Azure Custom vision pricing sheet, you will not be charged if you keep them and not do any actions on them