TypeError: pic should be PIL Image or ndarray. Got <class 'numpy.ndarray'>

1

when i ues "torchvision.transforms.Totensor()"to conver a numpy.array to tensor.But numpy.array is not ndarray?

3 Likes
2

Could you print the shape and type or your input?

1 Like
3

maybe your input dimension is not a accept shape, can not deal with like a image

1 Like
4

@luodahei Hi, have you fixed the problem?
I have encountered the same problem, but no idea to fix it.
Could you please help me?

Best regards,
purespring

5

torchvision.transforms.ToTensor() does accept a ndarray. What doesn’t is the torchvision.transforms.Resize() or CenterCrop().
Resize() takes a PIL image, so you need to convert the ndarray like this:

from PIL import Image
PIL_image = Image.fromarray(ndarray_image)

Hope this does the trick for you!

4 Likes
6

I have the same error, and the reason in my case is the array is None, i.e, obtained from np.array(x) where x = None

7

I have the same issue, being that my data is of size torch.Size([32, 3, 64, 64]), and I’m using the transformations:

transform = transforms.Compose(
    [transforms.RandomHorizontalFlip(p=0.2),
    transforms.ColorJitter(brightness=.5, hue=.1),
    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])

I am calling as this:

transform(data)

According to the docs, this composition should accept tensor. My PyTorch version is 1.7.1

Why is there even such difference between accepting PIL or tensor types? This makes everything so much harder

8

This code works for me, so you might need to update PyTorch and torchvision:

transform = transforms.Compose(
    [transforms.RandomHorizontalFlip(p=0.2),
    transforms.ColorJitter(brightness=.5, hue=.1),
    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])

x = torch.randn(32, 3, 64, 64)
out = transform(x)
print(out.shape)
> torch.Size([32, 3, 64, 64])

The ability to apply torchvision.transforms on tensors is relatively new and makes it easier now.

1 Like
9

Hi I’m also getting this error with the most updated pytorch and torchvision. My transforms are

transform = transforms.Compose([torchvision.transforms.RandomCrop(size=(16, 16)), 
                                torchvision.transforms.RandomHorizontalFlip(p=0.5), 
                                torchvision.transforms.ToTensor(),
                                torchvision.transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
                               ])

If I comment the ToTensor() transform out I instead get an “Unexpected type <class ‘numpy.ndarray’>” error.

The error shows up when I’m calling this line in my training function.

for i, (x, y) in enumerate(loader):

It works though when I just have the ToTensor and Normalise transform, but not the first 2

10

The transformations would work on PIL.Images as well as PyTorch tensors (in newer torchvision releases), while it seems that you are trying to use a numpy array based on the error message.
Pass on of the supported types to transform and it should work.

1 Like