bug in rotation part of tio.transforms.RandomAffine

[javidsss]

Is there an existing issue for this?

• I have searched the existing issues

Bug summary

I experimented with various "degrees" values in tio.transforms.RandomAffine, and found that I randomly get empty images. Overall, due to the nature of rotation transformation, I expected some inconsistencies between different runs, but not getting completely zero images.

Code for reproduction

for i in range(20):
    transform_rotate = tio.transforms.RandomAffine(degrees=(0,0,0,0,5,5), p=1) # I experimented with degrees=(0,0,0,0,5,10), etc.
    torchio_image_rotate = transform_rotate(torchio_image)
    plt.figure()
    plt.imshow(torchio_image_rotate.data[0, :, :, 1], cmap='gray') #torchio_image_rotate shape: torch.Size([1, 224, 224, 48])

Actual outcome

Out of 20 images, maybe around 13-14 of them look very similar to each other and are the correct rotated images by around 5 degrees, however, about 6-7 of the output images are just zeros

Error messages

No error, the outcome is not correct

Expected outcome

I expected about 20 images relatively similar to each other. They should be about 5 degrees rotated with respect to their last axis

System info

No response

[romainVala]

Hi
sound weird, I suspect it comes from your images
but try with
transform_rotate = tio.transforms.RandomAffine(degrees=(0,0,0,0,5,5), p=1, scale=0)
by default scaling factor is 0.1

[fepegar]

@javidsss, can you please share some data, a minimal working example and screenshots?

[fepegar]

By the way, you are showing the top slice, which will probably be black if you are looking, for example, at an MRI in RAS orientation:

torchio_image_rotate.data[0, :, :, 1]

For example:

In [1]: import torchio as tio

In [2]: t1 = tio.datasets.FPG().t1

In [3]: import matplotlib.pyplot as plt

In [4]: plt.imshow(t1.data[0, :, :, 1], cmap='gray')
Installed osx event loop hook.
Out[4]: <matplotlib.image.AxesImage object at 0x16c7cc950>

In [5]: plt.show()

You can use the plot method instead:

In [6]: t1.plot()

I'm going to assume that it's simply a visualization issue and not a bug. Feel free to reopen if you disagree.

[javidsss]

Thank you all for your help. The trick with adding scales=0 solved the issue.