mean_std.py

"""
Module to compute the mean and std of the whole training dataset once cropped
"""
import glob
import multiprocessing

from PIL import Image
import numpy as np
import pickle as pkl


def append_img(args):
    """
    Function to crop, resize and convert an image to a numpy array
    :param args: object containing the settings of the preprocess
    :return: the image transformed
    """
    print(args[0])
    meta = args[2]
    min_left = min(meta['left'])
    max_left = max(meta['left']) + max(meta['width'])
    min_top = min(meta['top'])
    max_top = max(meta['top']) + max(meta['height'])
    image = Image.open(args[1])
    image = image.crop((
        (1 - 0.3) * min_left,
        (1 - 0.3) * min_top,
        (1 + 0.3) * max_left,
        (1 + 0.3) * max_top))
    image = image.resize((64,64))
    return np.array(image)


if __name__ == '__main__':
    metadata = pkl.load(open("./data/SVHN/train_metadata.pkl", "rb"))
    imgs_paths = [path for path in glob.glob("./data/SVHN/train/*.png")]
    indexes = [int(path[path.rfind("\\") + 1:path.rfind(".")])
               for path in imgs_paths]
    metas = [metadata[index - 1]["metadata"] for index in indexes]
    args = zip(indexes, imgs_paths, metas)

    p = multiprocessing.Pool()
    imgs = [p.map(append_img, args)]
    imgs = np.array(imgs).squeeze()
    print(imgs.mean(axis=(0,1,2)) / 255.0)
    print(imgs.std(axis=(0,1,2)) / 255.0)