Vision: Transforms for C++ API

[ShahriarSS]

🚀 Feature

Hi.
I want to implement transforms using OpenCV for the C++ API after finishing the C++ models.

[fmassa]

While using OpenCV is the natural choice, I'm considering modifying the transforms in torchvision to use PyTorch ops. It would thus be natural to also use PyTorch ops in C++.

Thoughts?

[soumith]

i'm more looking forward to using pytorch ops, like grid_sample, F.interpolate, etc.

[ShahriarSS]

Well I guess PyTorch ops would be a better choice. I'll implement the new transforms when they are done. I can help with the new implementation as well if you'd like.

[moskomule]

Do you plan to use PyTorch ops for all operations, including random crop and random color operations as well?

[fmassa]

@moskomule for all ops that are currently supported, including random crop and random color operations, yes.

[FabianSchuetze]

Is there any work already done on including the transform ops in the cpp api?

I see that the python implementations transform ops often use the PIL Image object as arguments for function calls. How would images be passed in the cpp api?