Rotation transform not working - Semantic Segmentation

I’m trying to apply rotation data augmentation to my images’ dataset. The matter is that I’m using satellite images which has 32-bit values so I have created my own Items and ItemLists in order to deal with those images.

Everything works right, I can train on pretained neural networks and so on. The problem comes when I add rotation transform to train DataLoader. It throws me: " ‘tuple’ object has no attribute ‘pixel’ "

· I haven’t overwrite apply_tfms methods
· In order to add the tfms I do: src.train_dl.add_tfm(get_transforms()[0][3])

Any idea?

The data augmentation transforms should be applied to the dataset, you are trying to apply them to the dataloader.

1 Like

The only places in which I can add transformations are:
· doing .transformations() after calling .databunch()
· after generating the databunch() throught its dataloader.

Both ways give me errors.

No, that’s not true. Although it’s hard to point you in the right direction without seeing any code.