preprocessing_util.py

# -*- coding: utf-8 -*-
import numpy as np

def autoencoder_vae_add_noise(batch_xs,batch_ys): 
    #add noising step
    batch_xs_target = batch_xs
    batch_xs = batch_xs * np.random.randint(2, size=batch_xs.shape)
    batch_xs += np.random.randint(2, size=batch_xs.shape)
    
    batch_xs = batch_xs
    batch_ys = batch_xs_target

    return batch_xs,batch_ys

# Augment training data
def expend_training_data(images, labels):

    expanded_images = []
    expanded_labels = []

    j = 0 # counter
    for x, y in zip(images, labels):
        j = j+1
        if j%100==0:
            print ('expanding data : %03d / %03d' % (j,numpy.size(images,0)))

        # register original data
        expanded_images.append(x)
        expanded_labels.append(y)

        # get a value for the background
        # zero is the expected value, but median() is used to estimate background's value
        bg_value = np.median(x) # this is regarded as background's value
        image = np.reshape(x, (-1, 28))

        for i in range(4):
            # rotate the image with random degree
            angle = np.random.randint(-15,15,1)
            new_img = ndimage.rotate(image,angle,reshape=False, cval=bg_value)

            # shift the image with random distance
            shift = np.random.randint(-2, 2, 2)
            new_img_ = ndimage.shift(new_img,shift, cval=bg_value)

            # register new training data
            expanded_images.append(np.reshape(new_img_, 784))
            expanded_labels.append(y)

    # images and labels are concatenated for random-shuffle at each epoch
    # notice that pair of image and label should not be broken
    expanded_train_total_data = np.concatenate((expanded_images, expanded_labels), axis=1)
    np.random.shuffle(expanded_train_total_data)

    return expanded_train_total_data