when I have a previous-trained model with parameter fraction=0.1, what happens, if I set fraction=0.5 or fraction=1 when resuming the training (resume=True). Are the images and annotations from the previous run again taken for the resumed training? Is there a way to skip those images/annotations?
The only arguments that can be overridden during resume are:
So your change in fraction will be ignored. You would need to (and should) start a new training sessions anyhow if you want to increase the samples to use.
I asked my question in the wrong way:
I should have written, that I want to train on top of an previously-trained-model (yolo task=detect model=‘best.pt’ …), but your answer would also answer this type of question.
In a short summary of your answer: Neither a resume of training or a training on top of an existing model is considering the fraction of images/annotations of the previous training for the current training.