In one of the videos (can’t remember if it was 2019 Deep Learning or 2018 Machine Learning), Jeremy showed a tool that one of the study groups had created to remove mislabeled data (it showed you a bunch and you clicked on the ones that were labeled wrong).
I’m working with one of my own datasets now and noticed that many of the “errors” in plot_top_losses are actually mislabeled in my dataset. I’ve been searching the web and this forum for a link to that plugin but can’t find it.
Update: after fiddling around with this for way too long, I wanted to report back to say that the step the docs take about creating a new learner using .no_split() and loading it with your saved weights is actually very important. I was getting non-sensical loss values and showing seemingly random images prior to doing that (I think because it was trying to use DatasetType.Fix internally).