Cannot reproduce literature results

[ivdorelian]

The best I can get this to is 94.2% accuracy on the test set. I've done this with this implementation of the described augmentation method in the original paper:

def crop_mirror(image):

    pad = 4
    im_shape = image.shape
    new_image = np.zeros((im_shape[0] + 2*pad, im_shape[1] + 2*pad, 3))
    for c in range(im_shape[-1]):
        new_image[:, :, c] = np.pad(image[:, :, c], 4, mode='reflect')
    rand_row_start = np.random.randint(0, 8)
    rand_col_start = np.random.randint(0, 8)

    row_end = rand_row_start + im_shape[0]
    col_end = rand_col_start + im_shape[1]
    return new_image[rand_row_start:row_end, rand_col_start:col_end, :]

datagen = ImageDataGenerator(
    horizontal_flip=True,  # randomly flip images
    preprocessing_function=crop_mirror)

And std normalization:

(X_train, y_train), (X_test, y_test) = tf.keras.datasets.cifar10.load_data()
X_train = (X_train - X_train.mean(axis=0)) / (X_train.std(axis=0))
X_test = (X_test - X_test.mean(axis=0)) / (X_test.std(axis=0))

And I've also had to use 3 augmentation rounds to get there.

Any other idea about what might be the problem with reproducing the results in the paper using TF?

[asmith26]

Hi @ivdorelian,

Thanks for your information and code regarding cropping/augmentation. The best I have achieved is test accuracy 0.93%, but I haven't tried experimenting with the full paper implementations as I noted in the README

If you do manage to improve/reproduce the results in the paper, contributions are very welcome :)
Many thanks!