support int16 grayscale images

[bodokaiser]

This is often the case with medical (MRI) data.

Required changes would be in ToTensor probably something like:

# PIL image mode: 1, L, P, I, F, RGB, YCbCr, RGBA, CMYK
if pic.mode == 'YCbCr':
  nchannel = 3
else:
  nchannel = len(pic.mode)
# handle PIL Image
buf = pic.tobytes()
if len(buf) > pic.width * pic.height * nchannel:
  img = torch.LongTensor(torch.LongStorage.from_buffer(buf))
else:
  img = torch.ByteTensor(torch.ByteStorage.from_buffer(pic.tobytes()))
img = img.view(pic.size[1], pic.size[0], nchannel)

as well as in ToPILImage (just remove normalization to [0, 255] here?).

However I can't assess possible side effects. int16 support may be not very good in pillow (e.g. plt.imshow(Image.fromarray(int16_np_array)) does not work) also there may be other transforms which depend on [0, 255] byte range.

[fmassa]

I'd say as long as the returned tensor is properly converted to float and scaled to [0,1], things should be fine.
But we need to check if standard image transforms (like rotating, cropping, etc) work ok in PIL for int16 type.

Also, LongTensor is actually int64, you might be looking for a ShortTensor instead (which is signed).

[bodokaiser]

Thats the problem. I put a minimal example together where you can examine the problem.

from torchvision.transforms import Compose, ToPILImage, ToTensor
from matplotlib import pyplot as plt

import skimage.io
import numpy as np

img = skimage.io.imread('mr.tif')
print('img', img.shape, img.dtype)

plt.imshow(img)
plt.show()

transform = Compose([
    ToPILImage(),
    ToTensor(),
])

timg = transform(np.expand_dims(img, 2))

plt.imshow(timg[0].numpy())
plt.show()

here you find corresponding the tiff file.
mr.tif.zip

[fmassa]

The example that you mentioned shows that the current code is not adapted to int16 images, or did you try adding the modifications you mentioned?

[bodokaiser]

updated:

# PIL image mode: 1, L, P, I, F, RGB, YCbCr, RGBA, CMYK
if pic.mode == 'YCbCr':
  nchannel = 3
else:
  nchannel = len(pic.mode)
# handle PIL Image
buf = pic.tobytes()
if len(buf) > pic.width * pic.height * nchannel:
  img = torch.ShortTensor(torch.ShortStorage.from_buffer(buf, 'native'))
else:
  img = torch.ByteTensor(torch.ByteStorage.from_buffer(pic.tobytes()))
img = img.view(pic.size[1], pic.size[0], nchannel)

fails with RuntimeError: size '[466 x 394 x 1]' is invalid for input of with 367208 elements at /private/var/folders/y0/d4npmpd50971gpgqxtsvc25m0000gn/T/pip-_fraocf5-build/torch/lib/TH/THStorage.c:59 however changing to:

if len(buf) > pic.width * pic.height * nchannel:
  img = torch.ShortTensor(np.fromstring(buf, dtype=np.int16)[0::2])
else:
  img = torch.ByteTensor(torch.ByteStorage.from_buffer(pic.tobytes()))
img = img.contigouos().view(pic.size[1], pic.size[0], nchannel)

does the job. Furthermore we need to change:

class ToPILImage(object):
    """Converts a torch.*Tensor of range [0, 1] and shape C x H x W
    or numpy ndarray of dtype=uint8, range[0, 255] and shape H x W x C
    to a PIL.Image of range [0, 255]
    """

    def __call__(self, pic):
        npimg = pic
        mode = None
        if not isinstance(npimg, np.ndarray):
            npimg = pic.mul(255).byte().numpy()
            npimg = np.transpose(npimg, (1, 2, 0))

        if npimg.shape[2] == 1:
            npimg = npimg[:, :, 0]
            if npimg.dtype != np.int16:
                mode = "L"

        return Image.fromarray(npimg, mode=mode)

This works but is of course just a quick hack.

[fmassa]

Ok, cool.
But I was wondering, does PIL supports natively image operations in int16 image, such as rotate or crop? If it doesn't, then even if we adapt ToPILImage and ToTensor, we still won't be able to perform these operations. Also, as ToTensor converts the image to float, there would be no way of knowing if the original image was int16 or uint8, meaning that applying ToTensor() followed by ToPILImage would not return the identity.

[bodokaiser]

According to this issue and this PR it does only for grayscale images.

Regarding the behavior of ToTensor() one way to solve this would be that ToTensor() keeps the data type from PIL.Image but can take the target data type as argument.
Alternatively we could also ignore the fact that ToPILImage(ToTensor()) does not return the identity then we would have no API breaks and I also do do think we loose anything through this?

[soumith]

Bodo, it looks like you've been making a lot of progress already.

If you want to fire a few PRs to make torchvision work with int16 out of the box, I would love to have them. If not, I will eventually get to this for sure.

[bodokaiser]

Hey Sounith, yes I could do a PR! Before I would just like to clarify what pre processing to apply as in how pixel ranges are conserved. At the moment ToPILImage scales pixel range for every image to [0, 255] however this would not make sense for uint16/int16. I could then handle uint16/int16 as a special case and scale to [0, 65...] or just remove image scaling at all. What do you think? Also how would we address the concern of losing the identity between ToTensor and ToPILImage?

…

Am 23.03.2017 um 04:24 schrieb Soumith Chintala ***@***.***>: Bodo, it looks like you've been making a lot of progress already. If you want to fire a few PRs to make torchvision work with int16 out of the box, I would love to have them. If not, I will eventually get to this for sure. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[soumith]

0 to 65 sounds fine for int16/uint16. You can remove image scaling if you want too, I dont have experience with this domain, so I'll let you make a call.

In the case of identity preservation, ToPILImage needs to take a kwarg of Int16=True or something for the identity loop to happen. I dont see a better way. Same for ToTensor, taking the target data type as a kwarg seems good.

[bodokaiser]

@soumith @fmassa PR #122 is up for discussion!

[alykhantejani]

@fmassa I think this can now be closed as #122 was merged.