是否有更快的方法使用numpy中的矢量化操作从大型二维数组恢复图像

Question

I have large 2-D array (typically 0.5 to 2GB) of dimension of nx 1008. This array contains the several images and the values in the array are actually the pixel value. Basically what is done to recover these images is as follow

1. Start iterating over the array.
2. Take the first 260 rows ie you will have 260*1008=262080 values.
3. For the 261st row only take the first 64 values(the rest values in that row are junk). Thus now we have 262144 pixel values.
4. Dump all these values in a 1-D array say dump and do np.reshape(dump, (512,512))) to obtain the image. Notice that 512x512=262144
5. Repeat the same thing starting from 262nd row again.

This is my solution

counter=0
dump=np.array([], dtype=np.uint16)
#pixelDat is the array shaped n x 1008 containing the pixel values
for j in xrange(len(pixelDat)):
    #Check if it is the last row for a particular image
    if(j == (260*(counter+1)+ counter)):
        counter += 1
        dump=np.append(dump, pixelDat[j][:64])
        #Reshape dump to form the image and write it to a fits file
        hdu = fits.PrimaryHDU(np.reshape(dump, (512,512)))
        hdu.writeto('img'+str("{0:0>4}".format(counter))+'.fits', clobber=True)
        #Clear dump to enable formation of next image
        dump=np.array([], dtype=np.uint16)
    else:
        dump=np.append(dump, pixelDat[j])

I have been wondering if there is a way to speed up this whole process. The first thing that came to my mind is using vectorized numpy operations. However I am not very sure how to apply it in this case.

PS: Do not worry about the fits and hdu part. Its just creating a .fits file for my image.

Answer 1

Here is an attempt using flattening and np.split . It avoids copying data.

def chop_up(pixelDat):
    sh = pixelDat.shape
    try:
        # since the array is large we do not want a copy
        # the next line will succeed only if we can reshape in-place
        pixelDat.shape = -1
    except:
        return False # user must resort to other method
    N = len(pixelDat)
    split = (np.arange(0, N, 261*1008)[:, None] + (0, 512*512)).ravel()[1:]
    if split[-1] > N:
       split = split[:-2]
    result = [x.reshape(512,512) for x in np.split(pixelDat, split) if len(x) == 512*512]
    pixelDat.shape = sh
    return result