helper.py

"""Create images with simple shapes for use in ShapeShop. (28x28 pixels)
   The first batch is for VGG16.
   The second batch is for MNIST.

    Most function contain the following inputs and outputs:

    Args:
        mnist_X_train_sample: a 28x28 image, can have MNIST digit information or be blank.

    Returns:
       image: the created image.
"""

from scipy.misc import imresize
import numpy as np
from scipy.ndimage import imread
from keras import backend as K
from scipy import ndimage


# Misc functions

def preprocess_image(image_path):
    img_width = 28
    img_height = 28
    img = imresize(imread(image_path), (img_width, img_height)).astype('float32')
    img = img.transpose((2, 0, 1))
    img = img.astype('float32')
    # img = img[None,]
    return img


def deprocess_image(x):
    # util function to convert a tensor into a valid image

    # normalize tensor: center on 0., ensure std is 0.1
    x -= x.mean()
    x /= (x.std() + 1e-6)
    x *= 0.1

    # clip to [0, 1]
    x += 0.5
    x = np.clip(x, 0, 1)

    # convert to RGB array
    x *= 255
    x = x.transpose((1, 2, 0))
    x = np.clip(x, 0, 255).astype('uint8')
    return x


def normalize(x):
    # utility function to normalize a tensor by its L2 norm
    return x / (K.sqrt(K.mean(K.square(x))) + 1e-5)


def make_box_img(img_width, img_height):
    img = ndimage.gaussian_filter(np.random.random((1, 3, img_width, img_height))*1.0 + 1, 1)

    img[0, 0, 24:40, 24:40] = 0.0
    img[0, 1, 24:40, 24:40] = 0.0
    img[0, 2, 24:40, 24:40] = 0.0

    return img[0]


def make_noise_img(img_width, img_height):
    img = ndimage.gaussian_filter(np.random.random((1, 3, img_width, img_height))*1.0 + 1, 1)

    return img[0]


def make_box_img_no_blur(img_width, img_height):
    img = np.random.random((1, 3, img_width, img_height))*1.0 + 1

    img[0, 0, 24:40, 24:40] = 0.0
    img[0, 1, 24:40, 24:40] = 0.0
    img[0, 2, 24:40, 24:40] = 0.0

    return img[0]


def make_noise_img_no_blur(img_width, img_height):
    img = np.random.random((1, 3, img_width, img_height))*1.0 + 1

    return img[0]


# Shape functions

def boxify_top_left(mnist_X_train_sample):
    image_with_box = mnist_X_train_sample
    image_with_box[0, 0:6, 0:6] = 1.0
    return image_with_box


def boxify_top_right(mnist_X_train_sample):
    image_with_box = mnist_X_train_sample
    image_with_box[0, 0:6, 22:28] = 1.0
    return image_with_box


def boxify_bottom_left(mnist_X_train_sample):
    image_with_box = mnist_X_train_sample
    image_with_box[0, 22:28, 0:6] = 1.0
    return image_with_box


def boxify_bottom_right(mnist_X_train_sample):
    image_with_box = mnist_X_train_sample
    image_with_box[0, 22:28, 22:28] = 1.0
    return image_with_box


def lineify_top_left(mnist_X_train_sample):
    image_with_line = mnist_X_train_sample
    image_with_line[0, 0:6, 2] = 1.0
    image_with_line[0, 0:6, 3] = 1.0
    return image_with_line


def lineify_bottom_right(mnist_X_train_sample):
    image_with_line = mnist_X_train_sample
    image_with_line[0, 22:28, 28-2] = 1.0
    image_with_line[0, 22:28, 28-3] = 1.0
    return image_with_line


def circleify_top_left(mnist_X_train_sample):
    image_with_circle = mnist_X_train_sample

    a, b = 3, 3
    n = 7
    r = 2
    y, x = np.ogrid[-a:n-a, -b:n-b]
    mask = x*x + y*y <= r*r

    circle = np.ones((n, n))
    circle = circle*mask

    image_with_circle[0, 0:7, 0:7] = circle
    return image_with_circle


def circleify_bottom_right(mnist_X_train_sample):
    image_with_circle = mnist_X_train_sample

    a, b = 3, 3
    n = 7
    r = 2
    y, x = np.ogrid[-a:n-a, -b:n-b]
    mask = x*x + y*y <= r*r

    circle = np.ones((n, n))
    circle = circle*mask

    image_with_circle[0, 28-7:28, 28-7:28] = circle
    return image_with_circle


def triangulify_top_left(mnist_X_train_sample):
    image_with_tri = mnist_X_train_sample

    n = 7
    tri = np.zeros((n, n))
    tri[3, 0:7] = 1.0
    tri[2, 1:6] = 1.0
    tri[1, 2:5] = 1.0
    tri[0, 3] = 1.0
    image_with_tri[0, 0:7, 0:7] = tri

    return image_with_tri


def triangulify_bottom_right(mnist_X_train_sample):
    image_with_tri = mnist_X_train_sample

    n = 7
    tri = np.zeros((n, n))
    tri[3, 0:7] = 1.0
    tri[2, 1:6] = 1.0
    tri[1, 2:5] = 1.0
    tri[0, 3] = 1.0
    image_with_tri[0, 28-7:28, 28-7:28] = tri

    return image_with_tri


def boxify_center(mnist_X_train_sample):
    image_with_box = mnist_X_train_sample
    image_with_box[0, 11-4:17+4, 11-4:17+4] = 1.0
    return image_with_box


def lineify_center(mnist_X_train_sample):
    image_with_line = mnist_X_train_sample
    image_with_line[0, 11-4:17+4, 2+11] = 1.0
    image_with_line[0, 11-4:17+4, 3+11] = 1.0
    return image_with_line


def circleify_center(mnist_X_train_sample):
    image_with_circle = mnist_X_train_sample

    a, b = 7, 7
    n = 15
    r = 8
    y, x = np.ogrid[-a:n-a, -b:n-b]
    mask = x*x + y*y <= r*r

    circle = np.ones((n, n))
    circle = circle*mask

    image_with_circle[0, 7:22, 7:22] = circle
    return image_with_circle


def triangulify_center(mnist_X_train_sample):
    image_with_tri = mnist_X_train_sample

    a = 12
    b = a/2
    n = 7+a
    tri = np.zeros((n, n))
    tri[n-1, 0:n] = 1.0
    tri[n-2, 1:n-1] = 1.0
    tri[n-3, 2:n-2] = 1.0
    tri[n-4, 3:n-3] = 1.0
    tri[n-5, 4:n-4] = 1.0
    tri[n-6, 5:n-5] = 1.0
    tri[n-7, 6:n-6] = 1.0
    tri[n-8, 7:n-7] = 1.0
    tri[n-9, 8:n-8] = 1.0
    tri[n-10, 9:n-9] = 1.0

    image_with_tri[0, 11-b-5:18+b-5, 11-b:18+b] = tri
    return image_with_tri


def noiseify():
    image_with_noise = np.random.random([1, 28, 28])

    return image_with_noise


def boxify_center_hollow(mnist_X_train_sample):
    image_with_box = mnist_X_train_sample
    border = 2
    image_with_box[0, 11-4:17+4, 11-4:17+4] = 1.0
    image_with_box[0, 11-4+border:17+4-border, 11-4+border:17+4-border] = 0.0
    return image_with_box


def lineify_center_horizontal(mnist_X_train_sample):
    image_with_line = mnist_X_train_sample
    image_with_line[0, 2+11, 11-4:17+4] = 1.0
    image_with_line[0, 3+11, 11-4:17+4] = 1.0
    return image_with_line


def circleify_center_hollow(mnist_X_train_sample):
    image_with_circle = mnist_X_train_sample
    image_with_circle_big = np.copy(mnist_X_train_sample)

    circleify_center(image_with_circle_big)

    a, b = 7, 7
    n = 15
    r = 5
    y, x = np.ogrid[-a:n-a, -b:n-b]
    mask = x*x + y*y <= r*r

    circle = np.ones((n, n))
    circle = circle*mask

    image_with_circle[0, 7:22, 7:22] = circle
    final_image = image_with_circle_big[0]-image_with_circle[0]

    return final_image[None, ]


def triangulify_center_hollow(mnist_X_train_sample):
    image_with_tri = mnist_X_train_sample
    image_with_tri_big = np.copy(mnist_X_train_sample)

    triangulify_center(image_with_tri_big)

    a = 6
    b = a/2
    n = 7+a
    tri = np.zeros((n, n))
    tri[n-1, 0:n] = 1.0
    tri[n-2, 1:n-1] = 1.0
    tri[n-3, 2:n-2] = 1.0
    tri[n-4, 3:n-3] = 1.0
    tri[n-5, 4:n-4] = 1.0
    tri[n-6, 5:n-5] = 1.0
    tri[n-7, 6:n-6] = 1.0
    tri[n-8, 7:n-7] = 1.0
    tri[n-9, 8:n-8] = 1.0
    tri[n-10, 9:n-9] = 1.0

    image_with_tri[0, 11-b-3:18+b-3, 11-b:18+b] = tri
    final_image = image_with_tri_big[0] - image_with_tri[0]

    return final_image[None, ]


def noiseify_blur():
    image_with_noise = ndimage.gaussian_filter(np.random.random((1, 28, 28))*1.0, 1)

    return image_with_noise


# Random shapes

def boxify(mnist_X_train_sample):
    box_diameter = np.random.randint(28)
    i = np.random.randint(28-box_diameter)
    j = np.random.randint(28-box_diameter)
    image_with_box = mnist_X_train_sample
    image_with_box[0, i:i+box_diameter, j:j+box_diameter] = 1.0
    return image_with_box


def lineify(mnist_X_train_sample):
    line_diameter = np.random.randint(28)
    i = np.random.randint(28-line_diameter)
    j = np.random.randint(28-line_diameter)
    image_with_line = mnist_X_train_sample
    image_with_line[0, i:i+line_diameter, j] = 1.0
    image_with_line[0, i:i+line_diameter, j+1] = 1.0
    return image_with_line


def circleify(mnist_X_train_sample):
    image_with_circle = mnist_X_train_sample

    n = 28
    r = np.random.randint(1, high=n/2)

    a = np.random.randint(r, high=n-r)
    b = np.random.randint(r, high=n-r)

    y, x = np.ogrid[-a:n-a, -b:n-b]
    mask = x*x + y*y <= r*r

    circle = np.ones((n, n))
    circle = circle*mask

    image_with_circle[0, 0:n, 0:n] = circle
    return image_with_circle


# Half shapes

def boxify_top_left_half(mnist_X_train_sample):
    image_with_box = mnist_X_train_sample
    image_with_box[0, 3:6, 3:6] = 1.0
    return image_with_box


# Testing

def house(mnist_X_train_sample):
    image_with_house = mnist_X_train_sample
    box = boxify_center(np.zeros([1, 28, 28]))
    shifted_box = np.roll(box[0], 5, axis=0)

    tri = triangulify_center(np.zeros([1, 28, 28]))
    shifted_tri = np.roll(tri[0], -7, axis=0)

    image_with_house = shifted_box + shifted_tri
    return image_with_house[None, ]


# Saving

# def save_image(data, cm, fn, dpi):

#     sizes = np.shape(data)
#     height = float(sizes[0])
#     width = float(sizes[1])

#     fig = plt.figure()
#     fig.set_size_inches(width/height, 1, forward=False)
#     ax = plt.Axes(fig, [0., 0., 1., 1.])
#     ax.set_axis_off()
#     fig.add_axes(ax)

#     ax.matshow(data, cmap=cm)
#     plt.savefig(fn, dpi=dpi)
#     plt.close()