Float PILImage not converted as writeable

[lukasHoel]

🐛 Bug

When we have a float PIL-Image (e.g. mode='F'), e.g. for the purpose of applying transforms to it, and finally convert it with ToTensor then it will print the warning

/opt/conda/conda-bld/pytorch_1587428094786/work/torch/csrc/utils/tensor_numpy.cpp:141: UserWarning: The given NumPy array is not writeable, and PyTorch does not support non-writeable tensors. This means you can write to the underlying (supposedly non-writeable) NumPy array using the tensor. You may want to copy the array to protect its data or make it writeable before converting it to a tensor. This type of warning will be suppressed for the rest of this program.

The writeability of the numpy array got lost somewhere when converting a numpy array to PILImage with Image.fromarray(numpy_array, mode='F') and then after some transforms to a tensor with ToTensor.
This does not happen with PIL Images other than float (e.g. mode='RGB').

This warning is especially annoying since it gets printed every epoch.

To Reproduce

Steps to reproduce the behavior:

    from torchvision.transforms import ToTensor
    import numpy as np
    from PIL import Image
    a = np.array([[1.0,0.5], [1.0,0.5]])
    print(a.flags.writeable)
    pilimg = Image.fromarray(a, mode='F')
    tensor = ToTensor()(pilimg)
    print(tensor.numpy().flags)

    b = np.asarray(pilimg)
    c = np.array(pilimg)
    print(b.flags)
    print(c.flags)

This code will print above warning.

Also note the following:

• the numpy array b is NOT writeable, the numpy array c is writeable.
• This suggests that the error is located in the conversion from numpy to PIL.
• But since I do not have the easy possibility to convert PIL to numpy and then to tensor within a transforms Compose and since numpy array c is writeable, I open this issue here.

As workaround I do the following:

class ToNumpy(object):
    def __call__(self, sample):
        return np.array(sample)

def fix_compose_transform(transform):
        if isinstance(transform.transforms[-1], torchvision.transforms.ToTensor):
            transform = torchvision.transforms.Compose([
                *transform.transforms[:-1],
                ToNumpy(),
                torchvision.transforms.ToTensor()
            ])
        return transform

Expected behavior

Warning is not printed and ToTensor method can deal with the misbehaviour of PIL image.

Environment

Collecting environment information...
PyTorch version: 1.5.0
Is debug build: No
CUDA used to build PyTorch: 10.1

OS: Ubuntu 18.04.4 LTS
GCC version: (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0
CMake version: version 3.10.2

Python version: 3.8
Is CUDA available: Yes
CUDA runtime version: 10.1.243
GPU models and configuration: GPU 0: GeForce GTX 970
Nvidia driver version: 418.87.01
cuDNN version: Could not collect

Versions of relevant libraries:
[pip3] numpy==1.18.3
[conda] blas                      1.0                         mkl  
[conda] cudatoolkit               10.1.243             h6bb024c_0  
[conda] mkl                       2020.0                      166  
[conda] mkl-service               2.3.0            py38he904b0f_0  
[conda] mkl_fft                   1.0.15           py38ha843d7b_0  
[conda] mkl_random                1.1.0            py38h962f231_0  
[conda] numpy                     1.18.1           py38h4f9e942_0  
[conda] numpy-base                1.18.1           py38hde5b4d6_1  
[conda] numpydoc                  0.9.2                      py_0  
[conda] pytorch                   1.5.0           py3.8_cuda10.1.243_cudnn7.6.3_0    pytorch
[conda] pytorch3d                 0.1.1                    pypi_0    pypi
[conda] torchvision               0.6.0                py38_cu101    pytorch

[zhangguanheng66]

In your snippet code, Image is not defined.

[lukasHoel]

Sorry, forgot that. Now added, it's the PIL image...

[zhangguanheng66]

Yeap. I got the same warning as you. @fmassa Do you thing if we need a fix for this?

[pmeier]

We started to split to_tensor in pil_to_tensor (#2092) and convert_image_dtype (#2078). Maybe this already covered by what we already have. If not, I think it would be good fit there.

[lukasHoel]

@pmeier had a quick look at both suggestions:

• pil_to_tensor has the same problem, as it uses np.asarray instead of np.array to convert from PIL to numpy. I mentioned this in the corresponding line of code, see Add pil_to_tensor to functionals #2092. If it is ok to swap the numpy method, this would solve the problem in a similar way as in my workaround above, but also integrate it nicely into transforms.Compose.

• convert_image_dtype is not practical in this use case as users would like to keep the data as-is and only get them as tensor without converting types.

[pmeier]

It was meant as suggestion where we should fix this. If I understood @fmassa correctly convert_image_dtype(pil_to_tensor(pil_image)) will replace to_tensor(pil_image) in the future.

[fmassa]

@pmeier is correct, and convert_image_dtype(pil_to_tensor(pil_image)) will be a substitute for to_tensor(pil_image) for most of the cases.

About the writeable issue, I think this lies within PyTorch itself.
Making a copy in torchvision would add more overhead, which might not be necessary in most of the cases.

[bwesen]

FYI this happens with "normal" loading as well like opening a png with PIL Image, as luminance image for example, and then doing tvt.functional.pil_to_tensor warns this warning. Doesn't have to be a float image.

pic = Image.open(fn).convert(mode='L')
tensor = tvt.functional.pil_to_tensor(img).type(torch.ByteTensor)

=> the warning, ending with:
(Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:141.)
img = torch.as_tensor(np.asarray(pic))