Fold out a numpy array along a new dimension using values as index

Definitely more elegant: use np.ndenumerate().

for (i,j), val in np.ndenumerate(img):
    img_norm[val-1,i,j] = val

Looks like this should be faster than yours because O(N^2) rather than O(N^3). Let's try it out on an array with size and content as you describe:

def getnorm_ndenumerate(img):
    img_norm = np.zeros([np.max(img), *img.shape])
    for (i,j), val in np.ndenumerate(img):
        img_norm[val-1,i,j] = val  
    return img_norm

b = np.int64(25*np.random.rand(224, 224)) 

%timeit getnorm_ndenumerate(b)

gives

47.8 ms ± 1.38 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

It is indeed faster than yours. But the elegance comes at a price, because it is slower than acdr's method.

I made a mistake, in the output array, all the non-zeros should be 1. Sorry for my silly mistake.

thanks for all your help. I tested the three methods above, including code from Jean-François Corbett, acdr + Jean-François Corbett and mine. It turns out the method from acdr + Jean-François Corbett is the fastest.

Here is my testing code

def test_time():
    def func1(img, max_num):
        w, h = img.shape
        img_norm = np.zeros([w, h, max_num], np.float32)
        for (i, j), val in np.ndenumerate(img):
            # img_norm[i, j, val - 1] = val
            img_norm[i, j, val - 1] = 0 if val == 0 else 1
        return img_norm

    def func2(img, max_num):
        w, h = img.shape
        img_norm = np.zeros([w, h, max_num], np.float32)
        for idx in range(1, max_num + 1):
            # img_norm[:, :, idx - 1] = idx*(img == idx)
            img_norm[:, :, idx - 1] = (img == idx)
        return img_norm

    def func3(img, max_num):
        w, h = img.shape
        img_norm = np.zeros([w, h, max_num], np.float32)
        for idx in range(max_num):
            # img_norm[:, :, idx] = (idx+1) * (img[:, :, 0] == (idx + np.ones(shape=img[:, :, 0].shape)))
            img_norm[:, :, idx] = (img == (idx + np.ones(shape=img.shape)))
        return img_norm

    import cv2
    img_tmp = cv2.imread('dat.png', cv2.IMREAD_UNCHANGED)
    img_tmp = np.asarray(img_tmp, np.int)

    # img_tmp = np.array([
    #     [0, 0, 1, 0, 0],
    #     [2, 0, 3, 0, 1],
    #     [0, 2, 3, 1, 0],
    #     [0, 0, 1, 0, 0],
    #     [1, 0, 2, 0, 1],
    # ])

    img_bkp = np.array(img_tmp, copy=True)
    print(img_bkp.shape)
    import time
    cnt = 100
    maxnum = 24
    start_time = time.time()
    for i in range(cnt):
        _ = func1(img_tmp, maxnum)
    print('1 total time =', time.time() - start_time)

    start_time = time.time()
    for i in range(cnt):
        _ = func2(img_tmp, maxnum)
    print('2 total time =', time.time() - start_time)

    start_time = time.time()
    for i in range(cnt):
        _ = func3(img_tmp, maxnum)
    print('3 total time =', time.time() - start_time)

    print((img_tmp == img_bkp).all())
    img1 = func1(img_tmp, maxnum)
    img2 = func2(img_tmp, maxnum)
    img3 = func3(img_tmp, maxnum)
    print(img1.shape, img2.shape, img3.shape)
    print((img1 == img2).all())
    print((img2 == img3).all())
    print((img1 == img3).all())
    # print(type(img1[0, 0, 0]), type(img2[0, 0, 0]), type(img3[0, 0, 0]))
    # print('img1\n', img1[:, :, 2])
    # print('img3\n', img3[:, :, 2])

    (224, 224)
    1 total time = 4.738261938095093
    2 total time = 0.7725710868835449
    3 total time = 1.5980615615844727
    True
    (224, 224, 24) (224, 224, 24) (224, 224, 24)
    True
    True
    True

Thanks for all your kind help!