transforms.Lambda fails on class functions

transforms.Lambda appears to fail when a function that is passed to it is a class function:

import torchvision

class MyClass(object):

    def my_transform(self, img):
        return img

def my_transform(img):
    return img

t = MyClass()

print('Pure function works')
torchvision.transforms.Lambda(my_transform)
print('Class function fails')
torchvision.transforms.Lambda(t.my_transform)

Output:

Pure function works
Class function fails
Traceback (most recent call last):
  File "test_pytorch.py", line 17, in <module>
    torchvision.transforms.Lambda(t.my_transform)
  File "/braintree/home/qbilius/libs/miniconda3/lib/python3.6/site-packages/torchvision/transforms/transforms.py", line 279, in __init__
    assert isinstance(lambd, types.LambdaType)
AssertionError

My understanding it that isinstance(lambd, types.LambdaType) is not the best way to check if something is a function or not. Rather, callable or __call__ should be used.

I could submit a quick PR if that seems like a good solution.

[vfdev-5]

Glad that you can adopt a more simple solution 😃

Concerning lambdas, I think the idea is similar to pythonic lambda expressions:

t = Lambda(lambda img: img**2)

More realistic example, can be found even in torchvision with ColorJitter, where lambdas allows to wrap a function with initialization:

t = Lambda(lambda img: F.adjust_contrast(img, contrast_factor))

I agree with you that this can look artificial.

Anyway, maybe maintainers could explain better. cc @fmassa .

PS: if you think that there is an answer to your question you can close the issue and reopen it if there’s still something missing.

[qbilius]

Thanks for your kind help!