Override the original transforms.Compose

vision
1

I overrode the original transforms.Compose (vision/transforms.py at master · pytorch/vision · GitHub for preprocessing multiple targets, I found the type of img1 seemed turned to be class ‘torch.Tensor’ rather than normal class ‘PIL.Image.Image’ like img2. This seemed weird. Does anyone know the reason for this? Thanks!

Here’ my code,

class Compose(object):
    def __init__(self, transforms):
        self.transforms = transforms
    def __call__(self, img1, img2, mask1,mask2):

        print(type(img1))
        print(type(img2))
        print(type(mask1))
        print(type(mask2))
        print("img1.size",img1.size)
        print("mask1.size",mask1.size)
        print("mask2.size",mask2.size)

        assert img1.size() == mask1.size
        assert img1.size() == mask2.size

        for t in self.transforms:
            img1, img2,mask1,mask2 = t(img1,img2,mask1,mask2)
        return img1, img2,mask1,mask2

Then I got ,

type(img1) <class 'torch.Tensor'>
type(img2) <class 'PIL.Image.Image'>
type(mask1) <class 'PIL.Image.Image'>
type(mask2) <class 'PIL.Image.Image'>

img1.size() <built-in method size of Tensor object at 0x2b18bf80eb80>
mask1.size() (876, 376)
mask2.size() (876, 376)
... 

Traceback (most recent call last):
  File "train.py", line 210, in <module>
    main()
  File "train.py", line 110, in main
    train(net, optimizer)
  File "train.py", line 117, in train
img1.size() <built-in method size of Tensor object at 0x2b18bf7f3d00>
mask1.size() (550, 488)
mask2.size() (550, 488)
    for i, data in enumerate(train_loader):
  File "/home/public/software/anaconda3/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 363, in __next__
    data = self._next_data()
  File "/home/public/software/anaconda3/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 989, in _next_data
    return self._process_data(data)
  File "/home/public/software/anaconda3/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 1014, in _process_data
    data.reraise()
  File "/home/public/software/anaconda3/lib/python3.8/site-packages/torch/_utils.py", line 395, in reraise
    raise self.exc_type(msg)
AssertionError: Caught AssertionError in DataLoader worker process 0.
Original Traceback (most recent call last):
  File "/home/public/software/anaconda3/lib/python3.8/site-packages/torch/utils/data/_utils/worker.py", line 185, in _worker_loop
    data = fetcher.fetch(index)
  File "/home/public/software/anaconda3/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 44, in fetch
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/home/public/software/anaconda3/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 44, in <listcomp>
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/home/public/MyModel/datasets.py", line 75, in __getitem__
    img1, img2, mask1, mask2 = self.joint_transform(img1, img2, mask1, mask2)
  File "/home/public/MyModel/joint_transforms.py", line 16, in __call__
    assert img1.size() == mask1.size
AssertionError
2

Hi,

To answer your question: (I found the type of img1 seemed turned to be class ‘torch.Tensor’ rather than normal class ‘PIL.Image.Image’ like img2)

Since you don’t call much operations in the call part of your class then I would suggest checking the types of your train loader.

You can run next(iter(train_loader)) and then recieve all the data to which you can find the type. Since your call function does nothing to im1 to convert it to a Tensor then I suggest checking your getitem function in your data loader or your collate_fn if you pass one to the data loader as that could also possiblt convert im1 to a Tensor before it is being called by Compose().

Sarthak

3

Many thanks. I found it’s due to the transform order’s matter in getitem at last.