Creating a new training dataset when there are previous datasets

I have previously gone through the whole labeling and training process with my data (and extracted outlier frames, merged the datasets, and reran the training). However, I’ve recently found mistakes in those labeled frames, so I’ve gone through and fixed them all.

How should I go about training the model without incorporating the old models?

Right now, if I try to run create_training_dataset, it says that the folders for iteration-1 already exist. But I don’t want to redo what was in iteration-1 because those frames weren’t accurately labeled.

Do I use merge_datasets (even though I haven’t added any new frames) and the run create_training_dataset? Or do I change some of the parameters in create_training_dataset?

Nevermind, I just used merge_datasets and ran the training as iteration-2 with default weights

1 Like