However, whenever I try to create a dataset with the same name, it gives me an error. (See image below.) But then, if I change the name slightly, it works.
I already created a model, but it doesn’t work well, so I want to add to it without having to reprocess all the frames I used for the original model. Any suggestions?
Almost! Additional training depends almost entirely on the LABELS in the models, not the arbitrary names of datasets or models.
You have two options for additional training:
(1) Continue to train more steps on the model with the original dataset.
(2) Add a new dataset to the existing model
You can always do (1), and you do that by selecting the model you want to continue training on (from the Models tab), click the “More Training” button, provide how many additional steps you want to train for, and GO! Since we’re adding extra training to the model, we always need to save the NEW model (it’s going to be the old model PLUS additional training) as a different name.
For #2, you can add an additional dataset to a model IF AND ONLY IF the new dataset has all of the exact same LABELS as the original dataset used to train the model. The ftc-ml software will look at your datasets and show you datasets that are compatible that HAVE NOT yet been included in the model. If there are any datasets that are compatible, it will create a list of these for you to select on the “More Training” dialog, like so:
Just select the dataset you want to add, and train away. Just like the other scenario, this will create a NEW model (OLD model PLUS an additional dataset) so we need to give a new description name for the model.
Remember, Datasets can ONLY be added if they have the EXACT SAME labels. If the original dataset used to train the Model has “Ducks” and “Balls” as labels, any NEW datasets must have EXACTLY AND ONLY “Ducks” and “Balls” as labels. Not simply just “Ducks”, not simply “Balls”, and definitely NO NEW LABELS (you cannot add “Cubes” to the list of labels). Fun fact, this is why the original labels in the Model Zoo models are unavailable (they’re wiped) when we train a new model on top of an existing Model Zoo model.
Okay, thanks so much! It actually makes a lot of sense that it depends on the labels not the names. And I don’t think I realized that there were two different ways to improve the model, so thanks for adding that as well!
