how do you ensure you aren’t just blindly deleting images that your model genuinely failed to predict vs. the intention of deleting actually mislabeled/bad pictures?
does fastai has support for image preprocessing to make classifier work better with the images which are not good in quality… there are augmentations but there are many like contours,canny ,contrasting etc which could be required …In real world we may not get many good picture specially satellite images
That is exactly why this is not done automatically. You must delete images that do not correspond to the category, not delete images that are misclassified. In this example, if we have a Google Image of a music band called ‘Grizzly’ we should delete it. If we have an image of a small grizzly that the model predicted as a teddy bear, we should not delete it.
Jeremy is talking about production inference, which I also find fast on CPU, but how come there are these companies coming out with new hardware for inferencing (start ups if I recall)?