Affine Transform: why is translate a list[int] when the code suggests it could be floating point?

[varung]

vision/torchvision/transforms/functional.py

Line 956 in f16322b

img: Tensor, angle: float, translate: List[int], scale: float, shear: List[float],

cc @vfdev-5

[varung]

vision/torchvision/transforms/functional.py

Line 1055 in f16322b

translate_f = [1.0 * t for t in translate]

This line converts the translation into floating point...it seems like we should be able to pass in floating point list

[voldemortX]

@varung I think I can give you some preliminary insight on this matter before someone knows more about this comes in later.

Translation is just defined as integer for this transform:

vision/torchvision/transforms/functional.py

Line 967 in f16322b

translate (sequence of integers): horizontal and vertical translations (post-rotation translation)

But the matrix getter takes float inputs:

vision/torchvision/transforms/functional.py

Line 833 in f16322b

center: List[float], angle: float, translate: List[float], scale: float, shear: List[float]

For tensor transforms, we need to consider torchscript support as well, which means the data type needs to be consistent, hence the conversion to float.
You can see here the PIL transform does not use a conversion.

[vfdev-5]

@voldemortX thanks for explaining tech things!

@varung in addition to what @voldemortX said about type consistency, yes, we may also think about having translate API as float to support subpixel translations...