Is there a faster way to get the Local Binary Pattern of the MNIST dataset?

Question

I need to know if there's a faster way to get the LBP and the resulting histograms of the MNIST dataset. This will be used for handwritten text recognition, through a model that I haven't decided yet..

I've loaded the MNIST dataset and split it to its x, y training sets and x, y test sets based on tensorflow tutorials.

I've then used cv2 to invert the images.

From there I've defined a function using skimage to get the LBP and the corresponding histogram of an input image

I finally used a classic for loop to iterate through the images, get their histograms, store these in a separate list, and return the new list and the unaltered label list of both training and test sets.

Here is the function to load the MNIST dataset:

def loadDataset():
    mnist = tf.keras.datasets.mnist
    (x_train, y_train), (x_test, y_test) = mnist.load_data()

    # should I invert it or not?
    x_train = cv2.bitwise_not(x_train)
    x_test = cv2.bitwise_not(x_test)

    return (x_train, y_train), (x_test, y_test)

Here is the function to get the LBP and the corresponding histogram:

def getLocalBinaryPattern(img, points, radius):
    lbp = feature.local_binary_pattern(img, points, radius, method="uniform")
    hist, _ = np.histogram(lbp.ravel(), 
                bins=np.arange(0, points + 3),
                range=(0, points + 2))

    return lbp, hist

And lastly here's the function to iterate over the images:

def formatDataset(dataset):
    (x_train, y_train), (x_test, y_test) = dataset

    x_train_hst = []
    for i in range(len(x_train)):
        _, hst = getLocalBinaryPattern(x_train[i], 8, 1)
        print("Computing LBP for training set: {}/{}".format(i, len(x_train)))
        x_train_hst.append(hst)

    print("Done computing LBP for training set!")

    x_test_hst=[]
    for i in range(len(x_test)):
        _, hst = getLocalBinaryPattern(x_test[i], 8, 1)
        print("Computing LBP for test set: {}/{}".format(i, len(x_test)))
        x_test_hst.append(hst)

    print("Done computing LBP for test set!")

    print("Done!")

    return (x_train_hst, y_train), (x_test_hst, y_test)

I know it'll be slow, and indeed, it is slow. So I'm kind of looking for more ways to speed it up or if there is already a version of the dataset that has this info I needed.

Answer 1

I don't think there's a straightforward way to speed up the iteration over the images. One might expect that using NumPy's vectorize or apply_along_axis would improve performance, but these solutions are actually slower than a for loop (or a list comprehension).

Demo

Different alternatives for iterating through the images:

def compr(imgs):
    hists = [getLocalBinaryPattern(img, 8, 1)[1] for img in imgs]
    return hists

def vect(imgs):
    lbp81riu2 = lambda img: getLocalBinaryPattern(img, 8, 1)[1]
    vec_lbp81riu2 = np.vectorize(lbp81riu2, signature='(m,n)->(k)')
    hists = vec_lbp81riu2(imgs)
    return hists

def app(imgs):
    lbp81riu2 = lambda img: getLocalBinaryPattern(img.reshape(28, 28), 8, 1)[1]
    pixels = np.reshape(imgs, (len(imgs), -1))
    hists = np.apply_along_axis(lbp81riu2, 1, pixels)
    return hists

Results:

In [112]: (x_train, y_train), (x_test, y_test) = loadDataset()

In [113]: %timeit -r 3 compr(x_train)
1 loop, best of 3: 14.2 s per loop

In [114]: %timeit -r 3 vect(x_train)
1 loop, best of 3: 17.1 s per loop

In [115]: %timeit -r 3 app(x_train)
1 loop, best of 3: 14.3 s per loop

In [116]: np.array_equal(compr(x_train), vect(x_train))
Out[116]: True

In [117]: np.array_equal(compr(x_train), app(x_train))
Out[117]: True