transform vs target_transform vs transforms in PyTorch (1)

#python #pytorch #transform #target

*Memos:

My post explains target_transform and transform & target_transform.
My post explains transforms and transform & target_transform & transforms.
My post explains OxfordIIITPet().
My post explains Resize().
My post explains Lambda.

There are the differences between transform, target_transform and transforms as shown below. *It's about origin and transform:
*Memos:

transform is for the function which must have one parameter for transform. *Resize(), RandomRotation(), RandomHorizontalFlip(), etc also can be used.
target_transform is for the function which must have one parameter for target(label).
transforms is for the function which must have two parameters for both transform and target(label). *Resize(), RandomRotation(), RandomHorizontalFlip(), etc also can be used.
Both transform and target_transform can be used at the same time.
transforms cannot be used with transform and/or target_transform at the same time.
According to my experiments, target_transform and transforms are useless.

<origin>

from torchvision.datasets import OxfordIIITPet

origin_data = OxfordIIITPet(
    root="data"
)

origin_data[0]
# (<PIL.Image.Image image mode=RGB size=394x500>, 0)

origin_data[50]
# (<PIL.Image.Image image mode=RGB size=500x333>, 1)

origin_data[100]
# (<PIL.Image.Image image mode=RGB size=333x500>, 2)

<transform>

from torchvision.datasets import OxfordIIITPet
from torchvision.transforms.v2 import Resize

tfresize100_50_data = OxfordIIITPet(
    root="data",
    transform=Resize(size=[100, 50])
)

tfresize100_50_data[0]
# (<PIL.Image.Image image mode=RGB size=50x100>, 0)

tfresize100_50_data[50]
# (<PIL.Image.Image image mode=RGB size=50x100>, 1)

tfresize100_50_data[100]
# (<PIL.Image.Image image mode=RGB size=50x100>, 2)

from torchvision.datasets import OxfordIIITPet

def tf_func(transform):
    return transform

tf_func_data = OxfordIIITPet(
    root="data",
    transform=tf_func
    # transform=lambda transform: transform    
)

tf_func_data[0]
# (<PIL.Image.Image image mode=RGB size=394x500>, 0)

tf_func_data[50]
# (<PIL.Image.Image image mode=RGB size=500x333>, 1)

tf_func_data[100]
# (<PIL.Image.Image image mode=RGB size=333x500>, 2)

from torchvision.datasets import OxfordIIITPet

def tf_func(transform):
    return [0, 1, 2]

tf_func_data = OxfordIIITPet(
    root="data",
    transform=tf_func
    # transform=lambda transform: [0, 1, 2]
)

tf_func_data[0]
# ([0, 1, 2], 0)

tf_func_data[50]
# ([0, 1, 2], 1)

tf_func_data[100]
# ([0, 1, 2], 2)

from torchvision.datasets import OxfordIIITPet

def tf_func():
    return [0, 1, 2]

tf_func_data = OxfordIIITPet(
    root="data",
    transform=tf_func
    # transform=lambda: [0, 1, 2]
)

tf_func_data[0]
# TypeError: tf_func() takes 0 positional arguments but 1 was given

from torchvision.datasets import OxfordIIITPet

def tf_func(transform, target):
    return [0, 1, 2]

tf_func_data = OxfordIIITPet(
    root="data",
    transform=tf_func
    # transform=lambda transform, target: [0, 1, 2]
)

tf_func_data[0]
# TypeError: tf_func() missing 1 required positional argument: 'target'