I currently have a Vertex AI Model trained on 75,000 images and a few hundred hours of training.
I am looking to train the model on new images while keeping its old memory, however the last time I trained it was more than 6 months ago, (not allowing me to incrementally train it).
If I select “Train New Version” (Trains model as a version of an existing model) in the model details of pipeline training; will it retain my old models memory and accuracy? I don’t want to lose all of these hours of training on all of these images all because I lapsed 6 months.
When you select “Train New Version” , it creates a completely new version of your existing model. The new version is independent of the original model. This method is typically used when you want to create a new version, which usually involves training on the entire dataset from scratch, incorporating both old and new data.
If you train on the same dataset, the new version can retain some level of accuracy from the previous version, but it may also adapt to any changes in data distribution or patterns.
Your previous model versions will remain accessible in the Vertex AI console, allowing you to revert to any of those versions if needed.
For more information regarding Model Version, you can read through this documentation.
Thank you; I am still a little confused though. Let’s say I have a model that I trained for 50 hours called Model A. When I select “Train New Version” of Model A and train it for 5 more hours, does that mean that it has 55 hours of training? My issue is; on the next screen I cannot enable incremental training from Model A. How does this affect my model training on comparison?