I would like to finetune my model for the planet-kaggle competition data by training with the misclassified labels but I am not quite sure, how to figure this out.
The docu of ‘plot_multi_top_losses’ says:
Note that you can passsave_misclassified=True
(by default it'sFalse
). In such case, the method will return a list containing the misclassified images which you can use to debug your model and/or tune its hyperparameters.
So I would have to use this list to filter my dataloader?! How would I do that?