Error with transfroms.Scale when feeding iterable size

[ntomita]

I'm having a problem with transforms.Scale function, which complains
TypeError: unsupported operand type(s) for /: 'tuple' and 'int'
when feeding an iterable size.

It persists even after installed latest packages.
I don't see any problem in a source code, but seems like the if statement gets True for isinstance(self.size, int)

Here's how it fails.
>>> import torch
>>> import collections
>>> from torchvision.transforms import Scale
>>> from PIL import Image

>>> im = Image.open('pos19.jpg')
>>> Scale(size=(512,256))(im)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/ntomita/pytorch-env/lib/python2.7/site-packages/torchvision-0.1.8-py2.7.egg/torchvision/transforms.py", line 139, in __call__
TypeError: unsupported operand type(s) for /: 'tuple' and 'int'

Now checking each portion in a source.
>>> transform = Scale(size=(512, 224))
>>> isinstance(transform.size, int)
False
>>> (isinstance(transform.size, collections.Iterable) and len(transform.size) == 2)
True
Condition terms look fine, but when calling,
>>> transform(im)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/ntomita/pytorch-env/lib/python2.7/site-packages/torchvision-0.1.8-py2.7.egg/torchvision/transforms.py", line 139, in __call__
TypeError: unsupported operand type(s) for /: 'tuple' and 'int'
It fails. If I call resize directly,
>>> im.resize(transform.size, transform.interpolation)
<PIL.Image.Image image mode=L size=512x224 at 0x7F7F8C6D4F50>
works.

[ntomita]

By the way here's the version of torchvision I use
torchvision: 0.1.8-py27_2 soumith
Is this the latest one?

[antspy]

I have the same issue. I also use 0.1.8, which seems to be the latest version.

The problem is easy to solve, apparently they forgot to add the option of size being a tuple (even if it specified this way). In I believe something like this would solve it:

    def __call__(self, img):
        w, h = img.size
        if (w <= h and w == self.size) or (h <= w and h == self.size):
            return img
        if w < h:
            ow = self.size
            oh = int(self.size * h / w)
            return img.resize((ow, oh), self.interpolation)
        else:
            if isinstance(self.size, tuple): #<--- check if tuple, in which case just use it
                 return img.resize(self.size, self.interpolation)
            oh = self.size
            ow = int(self.size * w / h)
            return img.resize((ow, oh), self.interpolation)

[ntomita]

I see, I just avoid using the function and directly call resize() for now.

def _scale(size):
    def resize(im):
        return im.resize(size, Image.BILINEAR)
    return resize

[alykhantejani]

Hi @antspy this has been fixed in this PR. However, a new version of torchvision has not been released. If you install from source you will have this fix.

cc @fmassa I think we can close this