I’m new here (sorry, I if this is the wrong category).
In fastai v1 I could do the following:
from fastai.basics import * from fastai.vision.data import * from fastai.vision import models from fastai.vision import * from fastai.metrics import error_rate,accuracy from fastai.widgets import * from fastai.vision.augmentations import * dataset = ImageDataBunch.from_folder(data_dir, "train","val" ds_tfms=transforms, size=224) classifier = cnn_learner(dataset, models.squeezenet1_1, metrics=(error_rate,accuracy)) learn.fit_one_cycle(4) ds,idxs= DatasetFormatter().from_toplosses(learn) ImageCleaner(ds,idxs,data_dir)
So basically, I want to be able to clean my images and remove potentially unwanted images from my dataset. I want the images with the highest loss to be shown first.
I can’t seem to find neither DatasetFormatter, nor ImageDataBunch in fastai v2. How can I achieve this behavior in fastai v2? I searched for hours but did not find anything…
Bonus: Can I use my PyTorch-model and my PyTorch-dataset (or DataLoaders) without having to load them in again and train a completely fresh model?