Feature Request: Detect when writing to open mmap

[fepegar]

I have written a small function in IPython to delete the origin information from an image:
In [1]: def removeOrigin(imPath):
...: nii=nib.load(imPath)
...: affine = nii.affine
...: affine[:3, 3] = 0
...: data = nii.get_data()
...: new = nib.Nifti1Pair(data, affine)
...: nib.save(new, imPath)

My images are in Nifti1Pair format and very small (500 kB - 1 MB).

When I call my function, IPython crashes, the Terminal says Bus error and something very bad happens: my image data is gone!! The .hdr/.imgs file contain now 0 bytes, so I just lost my data.

I've tried outside the function and it happens the same when I try to overwrite the original image, i.e.:
nib.save(new, imPath) -> Bus error
nib.save(new, imPath.replace('.hdr', '_2.hdr')) -> Ok

[effigies]

Hi, sorry about the delay.

My guess is that the image is an mmap, so that when you begin writing, you''re truncating the file, and thus losing the data you want to write.

You could try data = nii.get_data().copy() to make sure the data is loaded into memory.

[fepegar]

Better late than never.

I don't have the technical background to fully understand your solution, but I'm sure it works. However, shouldn't nibabel be able to prevent this from hapenning?

[effigies]

Perhaps. Certainly worth a discussion. I do believe (though I haven't tested) that #466 or something similar would resolve the problem by unlinking the input file before writing to that name.

It should also be possible to verify that the target filename is the same as the source mmap, but there might be weirdness there I'm not immediately seeing.

As an aside, I would say that, for reproducibility and documentation reasons, I'd be very cautious about overwriting files. In this case, performing your action twice is the same as performing it once, so at least there's no danger to forgetting you've already done it. But in general, I'd strongly recommend setting up your workflows such that there is a clear flow of (input file, operation, output file).

[fepegar]

Sure, one has to be cautious when overwriting. But if everything I want is, say, change the origin of my 2GB nifti, maybe I wouldn't want two copies of my file...

[effigies]

Fair enough. Well, in the short term, I'd say use nib.load(fname, mmap=False), which won't require a manual copy command.

[fepegar]

Duly noted, thanks!

[VincentXWD]

Thank you @effigies . .get_data().copy() helps me a lot!

[matthew-brett]

I should also say, that in general you should prefer img.get_fdata(), which, I believe, always does a copy.

[fepegar]

I confirm this problem is still happening: TorchIO-project/torchio#77