Array passing without copying

In [ ]:

import shapeworks as sw

In [ ]:

import numpy as np

shapeworks Image from numpy array¶

In [ ]:

dims = (1,3,2) # NOTE: numpy dims are specified in z, y, x order
farr = np.ndarray(dims, dtype=np.float32)

ival = 10; jval = 50; kval = 1.75
for i in range(0, farr.shape[2]):
    for j in range(0, farr.shape[1]):
        for k in range(0, farr.shape[0]):
            farr[k][j][i] = ival*(i/farr.shape[2]) + jval*(j/farr.shape[1]) + kval/farr.shape[0]

In [ ]:

farr.mean()

In [ ]:

farr.dtype

In [ ]:

farr.flags['OWNDATA']

In [ ]:

farrimg = sw.Image(farr)
farrimg  # NOTE: sw.Image dims are specified in x, y, z order

In [ ]:

farrimg.mean()

While the numpy can still look at the memory, it no longer has ownership:¶

In [ ]:

farr.flags['OWNDATA']

In [ ]:

farrimg += 100

In [ ]:

farrimg.mean()

In [ ]:

farr.mean()

...so the safest thing to do now is let the array go out of scope:¶

• having used a temporary during Image construction: img = sw.Image(np.array(arr))
• variable replacement after Image construction: arr = np.zeros(1)
• explicit deletion after Image construction: del arr

In [ ]:

del farr

Only dtype.float32 arrays can be used to initialize an image:¶

In [ ]:

dims = (12,3,21)
darr = np.ndarray(dims, dtype=np.float64)

ival = 10; jval = 50; kval = 1.75
for k in range(0, dims[0]):
    for j in range(0, dims[1]):
        for i in range(0, dims[2]):
            darr[k][j][i] = ival*(i/darr.shape[2]) + jval*(j/darr.shape[1]) + kval/darr.shape[0]

In [ ]:

darr.dtype

In [ ]:

darr.flags['OWNDATA']

In [ ]:

# note: this try/catch is only used so the notebook runs to completion; not typically necessary
try:
    darrimg = sw.Image(darr) # Throws an exception because dtype must be same as Image's pixel type
except Exception as ex:
    print(ex)

In [ ]:

darrimg = sw.Image(np.array(darr, dtype=np.float32)) # Makes a copy of the array when passsed
darrimg

No unnecessary copies and no memory leaks!¶

The numpy array still owns its data since a copy was passed to create the Image.¶

Ownership of the copy's memory was taken by the image before it went out of scope.¶

In [ ]:

darr.flags['OWNDATA']

In [ ]:

darrimg.mean()

In [ ]:

darr.mean()

In [ ]:

darrimg += 50

In [ ]:

darrimg.mean()

In [ ]:

darr.mean()

In [ ]:

darr *= 10

In [ ]:

darrimg.mean()

In [ ]:

darr.mean()

Now for the other way around.¶

numpy array from shapeworks Image¶

In [ ]:

ellipsoid_path = "../../../Testing/data/1x2x2.nrrd"
femur_path = "../../../Testing/data/femur.nrrd"

In [ ]:

img = sw.Image(ellipsoid_path)
img

In [ ]:

arr = img.toArray()
arr.dtype

In [ ]:

arr.mean()

In [ ]:

img.mean()

In [ ]:

arr.shape  # remember, numpy dims are zyx and Image dims are xyz

In [ ]:

img.dims()

The numpy array references the memory of the current Image and can change it:¶

In [ ]:

arr += 100

In [ ]:

img.mean()

In [ ]:

arr.mean()

...but it still doesn't have ownership.¶

Since many Image operations reallocate internally, it's still safest to let it go out of scope as shown above.¶

In [ ]:

arr.flags['OWNDATA']

In [ ]:

del arr

If a copy is needed, pass copy=True to toArray()¶

In [ ]:

arr = img.toArray(copy=True)
arr.flags['OWNDATA']

This can be useful when the array is created from a temporary Image:¶

In [ ]:

arr = sw.Image(ellipsoid_path).toArray(copy=True)
arr.mean()

In [ ]:

def use_arr(arr):
    return arr.mean()

In [ ]:

use_arr(sw.Image(ellipsoid_path).toArray(copy=True))

viewing the image using pyvista¶

One common reason to get the Image's array is for viewing using pyvista using the for_viewing argument:¶

In [ ]:

import pyvista as pv

In [ ]:

pv.set_jupyter_backend(backend="ipyvtklink")

In [ ]:

#help(pv.Plotter)

In [ ]:

plotter = pv.Plotter(shape    = (1, 1),
                     notebook = True,
                     border   = True)
plotter.add_axes()
plotter.add_bounding_box()
#plotter.show_bounds()  # for some reason extremely slow on osx
#plotter.show_grid()    # for some reason extremely slow on osx

In [ ]:

# NOTE: pyvisya-wrapped vtk images require 'F' ordering to prevent copying
arr = img.toArray(for_viewing = True) # 'F' is `for_viewing`
arr.flags

In [ ]:

arr.flags

In [ ]:

# sw2vtkImage takes care of this for us
vtkimg = sw.sw2vtkImage(img, verbose=True)

In [ ]:

vol = plotter.add_volume(vtkimg, shade=True, show_scalar_bar=True)

In [ ]:

plotter.show()

Finally, we can assign a numpy array to an existing Image¶

This retains the Image's origin, scale, and coordsys.¶

In [ ]:

plotter = pv.Plotter(shape    = (1, 1),
                     notebook = True,
                     border   = True)
plotter.add_axes()

In [ ]:

img1 = sw.Image(femur_path)

In [ ]:

img1.setSpacing((1.5, 0.75, 1))  # set spacing to show that it's preserved on both copy and assign

In [ ]:

img2 = sw.Image(img1)  # make a copy to be processed by a scipy Python filter (spacing preserved)

Let's use a scipy operation on the image:¶

In [ ]:

from scipy import ndimage

In [ ]:

ck = ndimage.gaussian_filter(img2.toArray(), 12.0)

The return from this filter is the right size and type, but it's a copy:¶

In [ ]:

ck.shape

In [ ]:

ck.dtype

In [ ]:

ck.flags['OWNDATA']

Let's assign it back to Image so we can retain Image's origin, scale, and coordsys:¶

In [ ]:

img2.assign(ck)

In [ ]:

# notice numpy array ownership has been transferred to Image
ck.flags['OWNDATA']

Now we can look at it again in the plotter:¶

In [ ]:

plotter.add_volume(sw.sw2vtkImage(img2), shade=True, show_scalar_bar=True)

plotter.add_volume(sw.sw2vtkImage(img1), shade=True, show_scalar_bar=True)

In [ ]:

plotter.show()