difference between transform & target_transform?

[isalirezag]

Can you please explain the difference between transform & target_transform?
I understand that we usually use transform for doing changes to our images, but target_transform is still unclear for me? what is exactly used for? can we use that for images? the definition of it says: a function that takes in the target and transforms it, but it is not clear to me what it exactly means

[Cadene]

Lets say you want to apply some data augmentations over a dataset of images with their segmentation masks. If you rotate the image to the left with transform, you would have to rotate the mask to the left as well. That's one of the point of target_transform.

[isalirezag]

cannot we just rotate both of them via transform?

[alykhantejani]

@isalirezag currently the transform of the datasets only takes the input as an argument, hence the need for both a transform and a target_transform. However, there is a discussion in #230 to extend transforms to handle both input and target together.

cc @fmassa I think we can close this.

[eduardo4jesus]

Lets say you want to apply some data augmentations over a dataset of images with their segmentation masks. If you rotate the image to the left with transform, you would have to rotate the mask to the left as well. That's one of the point of target_transform.

It took me a while to understand this explanation. So, I just want to highlight a term that @Cadene mentioned that should be in kept in mind in order to understand it right away.

For classification problems, the target are labels. Thus, target_transforms wouldn't make sense.
However, for segmentation problems, in which we have segmentation masks as targets, it totally make sense.

[fmassa]

@eduardo4jesus I wouldn't say that target_transforms doesn't make sense for classification problems.
One could for example want to merge labels into common categories during training (say, via some sort of taxonomy), and the target_transform could be used for that.

[eduardo4jesus]

@fmassa. (Wow) That is true! Thank you so much. I hadn't thought about that usage!

[SmallPigPeppa]

Thank you a lot!!! very helpful answer

[andrew-miao]

@fmassa Cool! Thank you for your informational answer!

[chenfengshijie]

Very useful answer!

[brando90]

what is an example of what the target transform should be for a classification problem?

[EddieAy]

@fmassa Thank you for your informational answer

[hanli435]

very helpful.thank you a lot!!