How to revert (re-pack) view_as_block for 3D volume arrays?

Question

I have used view_as_block to split my volumes to 64x64x64 volumes by using http://scikit-image.org/docs/dev/api/skimage.util.html#skimage.util.view_as_blocks. After some filters and modifications, I would like to pack them back. Is there any way to pack them in correct order.

previous shape:

print(np.asarray(padded_training_array).shape)

output:

(2240, 576, 1024, 1)

padded_training_array = view_as_blocks(np.squeeze(padded_training_array[:, :, :], block_shape=(64,64,64))

new shape:

(35, 9, 16, 64, 64, 64)

Some modifications.. and desired shape:

(2240, 576, 1024)

Answer 1

You can use numpy.reshape:

>>> import numpy as np
>>> from skimage import util
>>> image = np.random.random((6, 6, 6))
>>> blocks = util.view_as_blocks(image, (2, 2, 2))
>>> blocks.shape
(3, 3, 3, 2, 2, 2)
>>> blocks[(0,) * 6] = 3.0
>>> image2 = np.reshape(blocks, (6, 6, 6))
>>> image2[0, 0, 0]
3.0

But, note that view_as_blocks returns a view. If your modifications are done in-place, then you don't even need to reshape, your original image will be modified already:

>>> image[0, 0, 0]
3.0

If you want to avoid this, use view_as_blocks(...).copy().

Answer 2

To revert from view_as_blocks to the original 3-dimensional array, the axes list for np.transpose is [0, 3, 1, 4, 2, 5].

from skimage.util import view_as_blocks
import numpy as np

a = np.random.randint(5, size=(64, 128, 32))
a_blocks = view_as_blocks(a, (4, 4, 4))
a_reshaped = a_blocks.transpose([0, 3, 1, 4, 2, 5]).reshape((64, 128, 32))
np.array_equal(a_reshaped, a)