BIAP2 - Slicecopy


Matthew Brett








Alternative implementation as of Nibabel 2.0 with image proxy slicing : see


Please see for motivation.

Sometimes we have a biiig images and we don’t want to load the whole array into memory. In this case it is useful to be able to load as a proxy:

img = load('my_huge_image.nii')

and then take out individual slices, as in something very approximately like:

slice0 = img.get_slice(0)


Should slice0 be a copy or a view?

As from the previous discussion - BIAP1 - Towards immutable images - an image may be a proxy or an array.

If the image is an array, the most natural thing to return is a view. That is, modifying slice0 will modify the underlying array in img.

If the image is a proxy, it would be self-defeating to return a view, because that would involve reading the whole image into memory, exactly what we are trying to avoid. So, for a proxied image, we’d nearly always want to return a copy.

What slices should the slicing allow?

The img.get_slice(0) syntax needs us to know what slice 0 is. In a nifti image of 3 dimensions, the first is fastest changing on disk. To be useful 0 will probably refer to the slowest changing on disk. Otherwise we’ll have to load nearly the whole image anyway. So, for a nifti, 0 should be the first slice in the last dimension.

For Minc on the other hand, you can and I (MB) think always do get C ordered arrays back, so that the slowest changing dimension in the image array is the first. Actually, I don’t know how to read a Minc file slice by slice, but the general point is that, to know which slice is worth reading, you need to know the relationship of the image array dimensions to fastest / slowest on disk.

We could always solve this by assuming that we always want to do this for Analyze / Nifti1 files (Fortran ordered). It’s a little ugly of course.

Note that taking the slowest changing slice in a nifti image would be the equivalent of taking a slice from the last dimension:

arr = img.get_data()
slice0 = arr[...,0]

In general, we can get contiguous data off disk for the same data as contiguous data in memory (perhaps obviously). So, all of these are contiguous in the Fortran ordering case:


That is, in general, : up until the first specified dimension, then contiguous slices, followed by integer slices. So, all of these can be read directly off disk as slices. Obviously the rules are the reverse for c-ordered arrays.

Option 1: fancy slice object

It’s option 1 because it’s the first one I thought of:

slice0 = img.slicecopy[...,0]

Here we solve the copy or view problem with ‘always copy’. We solve the ‘what slicing to allow’ by letting the object decide how to do the slicing. We could obviously just do the full load (deproxy the image) and return a copy of the sliced array, as in:

class SomeImage:
    class Slicer:
        def __init__(self, parent):
            self.parent = parent
        def __getitem__(self, slicedef):
            data = parent._data
            if is_proxy(data) and iscontinguous(slicedef, order='F'):
                return read_off_disk_somehow(slicedef, data)
            data = parent.get_data(unproxy=True)
            return data.__getitem__(slicedef)
    def __init__(self, stuff):
        self.slicecopy = Slicer(self)

The problem with this is that:

slice0 = img.slicecopy[...,1]

might unproxy the image. At the moment, it’s rather hidden whether the image is proxied or not on the basis that it’s an optimization that should be transparent.

Option 2: not-fancy method call

slice0 = img.get_slice(0, copy=True)

‘slice or view’ solved with explicit keyword. ‘which slice’ solved by assuming you always mean one slice in the last dimension. Or we could also allow:

slices = img.get_slices(slice(0,3), copy=True)

This is ugly, but fairly clear. This simple ‘I mean the last dimension’ might be annoying because it assumes the last dimension is the slowest changing, and it does not get to optimize the more complex contiguous cases above. So we could even allow full slicing with stuff like:

slice = img.get_slices((slice(None), slice(None), slice(3)), copy=True)

Again - this looks a lot more ugly than the slicecopy syntax above.

Now, when would you choose copy=True? I think, when the image is a proxy. Otherwise you’d want a view. Probably. So what you mean, probably, is something like this:

slices = img.get_slices(slicedef, copy_if='is_proxy')

But, we’ve established that for some slices, you’re going to have to load the whole image anyway. So in fact probably what you want is to:

  1. Take a view if this image is not a proxy

  2. Take a copy if we can read this directly off disk

  3. Unproxy the image if we have to read the whole thing off disk anyway to get the slices we want, on the basis that we have to read the whole thing into memory anyway, we might as well do that and save ourselves lots of disk thrashing getting the individual slices.

Of course that’s what option 1 boils down to. So I think I prefer version 1.