image_preprocessing.py

"""Fairly basic set of tools for real-time data augmentation on image data.
Can easily be extended to include new transformations,
new preprocessing methods, etc...
"""
from __future__ import absolute_import
from __future__ import print_function

import math
import os
import random
import re
import warnings

import cv2
import numpy as np
import scipy.ndimage as ndi
from keras import backend as K
from keras.applications.imagenet_utils import preprocess_input
from keras.preprocessing.image import DirectoryIterator, Iterator
from scipy import linalg
from six.moves import range
from torchvision import transforms

from bcolz_array_iterator import BcolzArrayIterator
from utils.image_utils import resize_and_pad

try:
    from PIL import Image as pil_image
except ImportError:
    pil_image = None


def randomCropFlips(size=224):
    transform = transforms.Compose([
        transforms.Lambda(lambda x: randomHorizontalFlip(x, u=0.5)),
        transforms.Lambda(lambda x: randomCrop(x, size)),
        transforms.Lambda(lambda x: preprocess_input(x, mode='tf')),
    ])
    return transform


def centerCrop(size=224):
    transform = transforms.Compose([
        transforms.Lambda(lambda x: cropCenter(x, height=size, width=size)),
        transforms.Lambda(lambda x: preprocess_input(x, mode='tf')),
    ])
    return transform


# http://enthusiaststudent.blogspot.jp/2015/01/horizontal-and-vertical-flip-using.html
# http://qiita.com/supersaiakujin/items/3a2ac4f2b05de584cb11
def randomVerticalFlip(img, u=0.5):
    if random.random() < u:
        img = cv2.flip(img, 0)  # np.flipud(img)  #cv2.flip(img,0) ##up-down
    return img


def randomHorizontalFlip(img, u=0.5):
    shape = img.shape
    if random.random() < u:
        img = cv2.flip(img, 1)  # np.fliplr(img)  #cv2.flip(img,1) ##left-right
    return img


def randomFlip(img, u=0.5):
    if random.random() < u:
        img = cv2.flip(img, random.randint(-1, 1))
    return img


def randomTranspose(img, u=0.5):
    if random.random() < u:
        img = img.transpose(1, 0, 2)  # cv2.transpose(img)
    return img


def cropCenter(img, height, width):
    h, w, c = img.shape
    dx = (h - height) // 2
    dy = (w - width) // 2

    y1 = dy
    y2 = y1 + height
    x1 = dx
    x2 = x1 + width
    img = img[y1:y2, x1:x2, :]

    return img


def randomCrop(img, crop_size=224):
    h, w, c = img.shape
    dy = random.randint(0, h - crop_size)
    dx = random.randint(0, w - crop_size)
    img = img[dy:dy + crop_size, dx:dx + crop_size]
    return img


# http://stackoverflow.com/questions/16265673/rotate-image-by-90-180-or-270-degrees
def randomRotate90(img, u=0.25):
    if random.random() < u:
        angle = random.randint(1, 3) * 90
        if angle == 90:
            img = img.transpose(1, 0, 2)  # cv2.transpose(img)
            img = cv2.flip(img, 1)
            # return img.transpose((1,0, 2))[:,::-1,:]
        elif angle == 180:
            img = cv2.flip(img, -1)
            # return img[::-1,::-1,:]
        elif angle == 270:
            img = img.transpose(1, 0, 2)  # cv2.transpose(img)
            img = cv2.flip(img, 0)
            # return  img.transpose((1,0, 2))[::-1,:,:]
    return img


def randomRotate(img, u=0.25, limit=90):
    if random.random() < u:
        angle = random.uniform(-limit, limit)  # degree

        height, width = img.shape[0:2]
        mat = cv2.getRotationMatrix2D((width / 2, height / 2), angle, 1.0)
        img = cv2.warpAffine(img, mat, (height, width), flags=cv2.INTER_LINEAR,
                             borderMode=cv2.BORDER_REFLECT_101)
        # img = cv2.warpAffine(img, mat, (height,width),flags=cv2.INTER_LINEAR,borderMode=cv2.BORDER_CONSTANT)

    return img


def randomShift(img, u=0.25, limit=4):
    if random.random() < u:
        dx = round(random.uniform(-limit, limit))  # pixel
        dy = round(random.uniform(-limit, limit))  # pixel

        height, width, channel = img.shape
        img1 = cv2.copyMakeBorder(img, limit + 1, limit + 1, limit + 1,
                                  limit + 1, borderType=cv2.BORDER_REFLECT_101)
        y1 = limit + 1 + dy
        y2 = y1 + height
        x1 = limit + 1 + dx
        x2 = x1 + width
        img = img1[y1:y2, x1:x2, :]

    return img


def randomShiftScale(img, u=0.25, limit=4):
    if random.random() < u:
        height, width, channel = img.shape
        assert (width == height)
        size0 = width
        size1 = width + 2 * limit
        img1 = cv2.copyMakeBorder(img, limit, limit, limit, limit,
                                  borderType=cv2.BORDER_REFLECT_101)
        size = round(random.uniform(size0, size1))

        dx = round(random.uniform(0, size1 - size))  # pixel
        dy = round(random.uniform(0, size1 - size))

        y1 = dy
        y2 = y1 + size
        x1 = dx
        x2 = x1 + size

        if size == size0:
            img = img1[y1:y2, x1:x2, :]
        else:
            img = cv2.resize(img1[y1:y2, x1:x2, :], (size0, size0),
                             interpolation=cv2.INTER_LINEAR)

    return img


def randomScale(img, u=0.25, scale_factor=0.150):
    if random.random() < u:
        height, width, channel = img.shape
        new_min_width = width * (1 - scale_factor)
        new_width = round(random.uniform(new_min_width, width))
        dx = round(random.uniform(0, width - new_width))
        dy = round(random.uniform(0, width - new_width))

        y1 = dy
        y2 = y1 + new_width
        x1 = dx
        x2 = x1 + new_width
        img = cv2.resize(img[y1:y2, x1:x2, :], (width, width),
                         interpolation=cv2.INTER_LINEAR)
    return img


def makeRandomFlips():
    transform = transforms.Compose([
        transforms.Lambda(lambda x: randomFlip(x, u=0.5)),
        transforms.Lambda(lambda x: randomTranspose(x, u=0.5)),
        transforms.Lambda(lambda x: randomRotate90(x, u=0.5)),
    ])
    return transform


def randomShiftScaleRotate(img, u=0.5, shift_limit=4, scale_limit=4,
                           rotate_limit=45):
    if random.random() < u:
        height, width, channel = img.shape
        assert (width == height)
        size0 = width
        size1 = width + 2 * scale_limit

        angle = random.uniform(-rotate_limit, rotate_limit)  # degree
        size = round(random.uniform(size0, size1))
        dx = round(random.uniform(0, size1 - size))  # pixel
        dy = round(random.uniform(0, size1 - size))

        cc = math.cos(angle / 180 * math.pi) * (size / size0)
        ss = math.sin(angle / 180 * math.pi) * (size / size0)
        rotate_matrix = np.array([[cc, -ss], [ss, cc]])

        box0 = np.array([[0, 0], [size0, 0], [size0, size0], [0, size0], ])
        box1 = box0 - np.array([width / 2, height / 2])
        box1 = np.dot(box1, rotate_matrix.T) + np.array(
            [width / 2 + dx, height / 2 + dy])

        box0 = box0.astype(np.float32)
        box1 = box1.astype(np.float32)
        mat = cv2.getPerspectiveTransform(box0, box1)
        img = cv2.warpPerspective(img, mat, (height, width),
                                  flags=cv2.INTER_LINEAR,
                                  borderMode=cv2.BORDER_REFLECT_101)

    return img


def random_rotation(x, rg, row_axis=1, col_axis=2, channel_axis=0,
                    fill_mode='nearest', cval=0.):
    """Performs a random rotation of a Numpy image tensor.

    # Arguments
        x: Input tensor. Must be 3D.
        rg: Rotation range, in degrees.
        row_axis: Index of axis for rows in the input tensor.
        col_axis: Index of axis for columns in the input tensor.
        channel_axis: Index of axis for channels in the input tensor.
        fill_mode: Points outside the boundaries of the input
            are filled according to the given mode
            (one of `{'constant', 'nearest', 'reflect', 'wrap'}`).
        cval: Value used for points outside the boundaries
            of the input if `mode='constant'`.

    # Returns
        Rotated Numpy image tensor.
    """
    theta = np.pi / 180 * np.random.uniform(-rg, rg)
    rotation_matrix = np.array([[np.cos(theta), -np.sin(theta), 0],
                                [np.sin(theta), np.cos(theta), 0],
                                [0, 0, 1]])

    h, w = x.shape[row_axis], x.shape[col_axis]
    transform_matrix = transform_matrix_offset_center(rotation_matrix, h, w)
    x = apply_transform(x, transform_matrix, channel_axis, fill_mode, cval)
    return x


def random_shift(x, wrg, hrg, row_axis=1, col_axis=2, channel_axis=0,
                 fill_mode='nearest', cval=0.):
    """Performs a random spatial shift of a Numpy image tensor.

    # Arguments
        x: Input tensor. Must be 3D.
        wrg: Width shift range, as a float fraction of the width.
        hrg: Height shift range, as a float fraction of the height.
        row_axis: Index of axis for rows in the input tensor.
        col_axis: Index of axis for columns in the input tensor.
        channel_axis: Index of axis for channels in the input tensor.
        fill_mode: Points outside the boundaries of the input
            are filled according to the given mode
            (one of `{'constant', 'nearest', 'reflect', 'wrap'}`).
        cval: Value used for points outside the boundaries
            of the input if `mode='constant'`.

    # Returns
        Shifted Numpy image tensor.
    """
    h, w = x.shape[row_axis], x.shape[col_axis]
    tx = np.random.uniform(-hrg, hrg) * h
    ty = np.random.uniform(-wrg, wrg) * w
    translation_matrix = np.array([[1, 0, tx],
                                   [0, 1, ty],
                                   [0, 0, 1]])

    transform_matrix = translation_matrix  # no need to do offset
    x = apply_transform(x, transform_matrix, channel_axis, fill_mode, cval)
    return x


def random_shear(x, intensity, row_axis=1, col_axis=2, channel_axis=0,
                 fill_mode='nearest', cval=0.):
    """Performs a random spatial shear of a Numpy image tensor.

    # Arguments
        x: Input tensor. Must be 3D.
        intensity: Transformation intensity.
        row_axis: Index of axis for rows in the input tensor.
        col_axis: Index of axis for columns in the input tensor.
        channel_axis: Index of axis for channels in the input tensor.
        fill_mode: Points outside the boundaries of the input
            are filled according to the given mode
            (one of `{'constant', 'nearest', 'reflect', 'wrap'}`).
        cval: Value used for points outside the boundaries
            of the input if `mode='constant'`.

    # Returns
        Sheared Numpy image tensor.
    """
    shear = np.random.uniform(-intensity, intensity)
    shear_matrix = np.array([[1, -np.sin(shear), 0],
                             [0, np.cos(shear), 0],
                             [0, 0, 1]])

    h, w = x.shape[row_axis], x.shape[col_axis]
    transform_matrix = transform_matrix_offset_center(shear_matrix, h, w)
    x = apply_transform(x, transform_matrix, channel_axis, fill_mode, cval)
    return x


def random_zoom(x, zoom_range, row_axis=1, col_axis=2, channel_axis=0,
                fill_mode='nearest', cval=0.):
    """Performs a random spatial zoom of a Numpy image tensor.

    # Arguments
        x: Input tensor. Must be 3D.
        zoom_range: Tuple of floats; zoom range for width and height.
        row_axis: Index of axis for rows in the input tensor.
        col_axis: Index of axis for columns in the input tensor.
        channel_axis: Index of axis for channels in the input tensor.
        fill_mode: Points outside the boundaries of the input
            are filled according to the given mode
            (one of `{'constant', 'nearest', 'reflect', 'wrap'}`).
        cval: Value used for points outside the boundaries
            of the input if `mode='constant'`.

    # Returns
        Zoomed Numpy image tensor.

    # Raises
        ValueError: if `zoom_range` isn't a tuple.
    """
    if len(zoom_range) != 2:
        raise ValueError('zoom_range should be a tuple or list of two floats. '
                         'Received arg: ', zoom_range)

    if zoom_range[0] == 1 and zoom_range[1] == 1:
        zx, zy = 1, 1
    else:
        zx, zy = np.random.uniform(zoom_range[0], zoom_range[1], 2)
    zoom_matrix = np.array([[zx, 0, 0],
                            [0, zy, 0],
                            [0, 0, 1]])

    h, w = x.shape[row_axis], x.shape[col_axis]
    transform_matrix = transform_matrix_offset_center(zoom_matrix, h, w)
    x = apply_transform(x, transform_matrix, channel_axis, fill_mode, cval)
    return x


def random_channel_shift(x, intensity, channel_axis=0):
    x = np.rollaxis(x, channel_axis, 0)
    min_x, max_x = np.min(x), np.max(x)
    channel_images = [
        np.clip(x_channel + np.random.uniform(-intensity, intensity), min_x,
                max_x)
        for x_channel in x]
    x = np.stack(channel_images, axis=0)
    x = np.rollaxis(x, 0, channel_axis + 1)
    return x


def transform_matrix_offset_center(matrix, x, y):
    o_x = float(x) / 2 + 0.5
    o_y = float(y) / 2 + 0.5
    offset_matrix = np.array([[1, 0, o_x], [0, 1, o_y], [0, 0, 1]])
    reset_matrix = np.array([[1, 0, -o_x], [0, 1, -o_y], [0, 0, 1]])
    transform_matrix = np.dot(np.dot(offset_matrix, matrix), reset_matrix)
    return transform_matrix


def apply_transform(x,
                    transform_matrix,
                    channel_axis=0,
                    fill_mode='nearest',
                    cval=0.):
    """Apply the image transformation specified by a matrix.

    # Arguments
        x: 2D numpy array, single image.
        transform_matrix: Numpy array specifying the geometric transformation.
        channel_axis: Index of axis for channels in the input tensor.
        fill_mode: Points outside the boundaries of the input
            are filled according to the given mode
            (one of `{'constant', 'nearest', 'reflect', 'wrap'}`).
        cval: Value used for points outside the boundaries
            of the input if `mode='constant'`.

    # Returns
        The transformed version of the input.
    """
    x = np.rollaxis(x, channel_axis, 0)
    final_affine_matrix = transform_matrix[:2, :2]
    final_offset = transform_matrix[:2, 2]
    channel_images = [ndi.interpolation.affine_transform(
        x_channel,
        final_affine_matrix,
        final_offset,
        order=0,
        mode=fill_mode,
        cval=cval) for x_channel in x]
    x = np.stack(channel_images, axis=0)
    x = np.rollaxis(x, 0, channel_axis + 1)
    return x


def flip_axis(x, axis):
    x = np.asarray(x).swapaxes(axis, 0)
    x = x[::-1, ...]
    x = x.swapaxes(0, axis)
    return x


def array_to_img(x, data_format=None, scale=True):
    """Converts a 3D Numpy array to a PIL Image instance.

    # Arguments
        x: Input Numpy array.
        data_format: Image data format.
        scale: Whether to rescale image values
            to be within [0, 255].

    # Returns
        A PIL Image instance.

    # Raises
        ImportError: if PIL is not available.
        ValueError: if invalid `x` or `data_format` is passed.
    """
    if pil_image is None:
        raise ImportError('Could not import PIL.Image. '
                          'The use of `array_to_img` requires PIL.')
    x = np.asarray(x, dtype=K.floatx())
    if x.ndim != 3:
        raise ValueError('Expected image array to have rank 3 (single image). '
                         'Got array with shape:', x.shape)

    if data_format is None:
        data_format = K.image_data_format()
    if data_format not in {'channels_first', 'channels_last'}:
        raise ValueError('Invalid data_format:', data_format)

    # Original Numpy array x has format (height, width, channel)
    # or (channel, height, width)
    # but target PIL image has format (width, height, channel)
    if data_format == 'channels_first':
        x = x.transpose(1, 2, 0)
    if scale:
        x = x + max(-np.min(x), 0)
        x_max = np.max(x)
        if x_max != 0:
            x /= x_max
        x *= 255
    if x.shape[2] == 3:
        # RGB
        return pil_image.fromarray(x.astype('uint8'), 'RGB')
    elif x.shape[2] == 1:
        # grayscale
        return pil_image.fromarray(x[:, :, 0].astype('uint8'), 'L')
    else:
        raise ValueError('Unsupported channel number: ', x.shape[2])


def img_to_array(img, data_format=None):
    """Converts a PIL Image instance to a Numpy array.

    # Arguments
        img: PIL Image instance.
        data_format: Image data format.

    # Returns
        A 3D Numpy array.

    # Raises
        ValueError: if invalid `img` or `data_format` is passed.
    """
    if data_format is None:
        data_format = K.image_data_format()
    if data_format not in {'channels_first', 'channels_last'}:
        raise ValueError('Unknown data_format: ', data_format)
    # Numpy array x has format (height, width, channel)
    # or (channel, height, width)
    # but original PIL image has format (width, height, channel)
    x = np.asarray(img, dtype=K.floatx())
    if len(x.shape) == 3:
        if data_format == 'channels_first':
            x = x.transpose(2, 0, 1)
    elif len(x.shape) == 2:
        if data_format == 'channels_first':
            x = x.reshape((1, x.shape[0], x.shape[1]))
        else:
            x = x.reshape((x.shape[0], x.shape[1], 1))
    else:
        raise ValueError('Unsupported image shape: ', x.shape)
    return x


def load_img(path, grayscale=False, target_size=None):
    """Loads an image into PIL format.

    # Arguments
        path: Path to image file
        grayscale: Boolean, whether to load the image as grayscale.
        target_size: Either `None` (default to original size)
            or tuple of ints `(img_height, img_width)`.

    # Returns
        A PIL Image instance.

    # Raises
        ImportError: if PIL is not available.
    """
    if pil_image is None:
        raise ImportError('Could not import PIL.Image. '
                          'The use of `array_to_img` requires PIL.')
    img = pil_image.open(path)
    if grayscale:
        if img.mode != 'L':
            img = img.convert('L')
    else:
        if img.mode != 'RGB':
            img = img.convert('RGB')
    if target_size:
        hw_tuple = (target_size[1], target_size[0])
        if img.size != hw_tuple:
            img = img.resize(hw_tuple)
    return img


def list_pictures(directory, ext='jpg|jpeg|bmp|png'):
    return [os.path.join(root, f)
            for root, _, files in os.walk(directory) for f in files
            if re.match(r'([\w]+\.(?:' + ext + '))', f)]


class ImageDataGenerator(object):
    """Generate minibatches of image data with real-time data augmentation.

    # Arguments
        featurewise_center: set input mean to 0 over the dataset.
        samplewise_center: set each sample mean to 0.
        featurewise_std_normalization: divide inputs by std of the dataset.
        samplewise_std_normalization: divide each input by its std.
        zca_whitening: apply ZCA whitening.
        zca_epsilon: epsilon for ZCA whitening. Default is 1e-6.
        rotation_range: degrees (0 to 180).
        width_shift_range: fraction of total width.
        height_shift_range: fraction of total height.
        shear_range: shear intensity (shear angle in radians).
        zoom_range: amount of zoom. if scalar z, zoom will be randomly picked
            in the range [1-z, 1+z]. A sequence of two can be passed instead
            to select this range.
        channel_shift_range: shift range for each channel.
        fill_mode: points outside the boundaries are filled according to the
            given mode ('constant', 'nearest', 'reflect' or 'wrap'). Default
            is 'nearest'.
        cval: value used for points outside the boundaries when fill_mode is
            'constant'. Default is 0.
        horizontal_flip: whether to randomly flip images horizontally.
        vertical_flip: whether to randomly flip images vertically.
        rescale: rescaling factor. If None or 0, no rescaling is applied,
            otherwise we multiply the data by the value provided. This is
            applied after the `preprocessing_function` (if any provided)
            but before any other transformation.
        preprocessing_function: function that will be implied on each input.
            The function will run before any other modification on it.
            The function should take one argument:
            one image (Numpy tensor with rank 3),
            and should output a Numpy tensor with the same shape.
        data_format: 'channels_first' or 'channels_last'. In 'channels_first' mode, the channels dimension
            (the depth) is at index 1, in 'channels_last' mode it is at index 3.
            It defaults to the `image_data_format` value found in your
            Keras config file at `~/.keras/keras.json`.
            If you never set it, then it will be "channels_last".
    """

    def __init__(self,
                 featurewise_center=False,
                 samplewise_center=False,
                 featurewise_std_normalization=False,
                 samplewise_std_normalization=False,
                 zca_whitening=False,
                 zca_epsilon=1e-6,
                 rotation_range=0.,
                 width_shift_range=0.,
                 height_shift_range=0.,
                 shear_range=0.,
                 zoom_range=0.,
                 channel_shift_range=0.,
                 fill_mode='nearest',
                 cval=0.,
                 horizontal_flip=False,
                 vertical_flip=False,
                 rescale=None,
                 preprocessing_function=None,
                 data_format=None):
        if data_format is None:
            data_format = K.image_data_format()
        self.featurewise_center = featurewise_center
        self.samplewise_center = samplewise_center
        self.featurewise_std_normalization = featurewise_std_normalization
        self.samplewise_std_normalization = samplewise_std_normalization
        self.zca_whitening = zca_whitening
        self.zca_epsilon = zca_epsilon
        self.rotation_range = rotation_range
        self.width_shift_range = width_shift_range
        self.height_shift_range = height_shift_range
        self.shear_range = shear_range
        self.zoom_range = zoom_range
        self.channel_shift_range = channel_shift_range
        self.fill_mode = fill_mode
        self.cval = cval
        self.horizontal_flip = horizontal_flip
        self.vertical_flip = vertical_flip
        self.rescale = rescale
        self.preprocessing_function = preprocessing_function

        if data_format not in {'channels_last', 'channels_first'}:
            raise ValueError(
                '`data_format` should be `"channels_last"` (channel after row and '
                'column) or `"channels_first"` (channel before row and column). '
                'Received arg: ', data_format)
        self.data_format = data_format
        if data_format == 'channels_first':
            self.channel_axis = 1
            self.row_axis = 2
            self.col_axis = 3
        if data_format == 'channels_last':
            self.channel_axis = 3
            self.row_axis = 1
            self.col_axis = 2

        self.mean = None
        self.std = None
        self.principal_components = None

        if np.isscalar(zoom_range):
            self.zoom_range = [1 - zoom_range, 1 + zoom_range]
        elif len(zoom_range) == 2:
            self.zoom_range = [zoom_range[0], zoom_range[1]]
        else:
            raise ValueError('`zoom_range` should be a float or '
                             'a tuple or list of two floats. '
                             'Received arg: ', zoom_range)

    def flow(self, x, y=None, batch_size=32, shuffle=True, seed=None,
             save_to_dir=None, save_prefix='', save_format='png'):
        return NumpyArrayIterator(
            x, y, self,
            batch_size=batch_size,
            shuffle=shuffle,
            seed=seed,
            data_format=self.data_format,
            save_to_dir=save_to_dir,
            save_prefix=save_prefix,
            save_format=save_format)

    def flow_from_directory(self, directory,
                            target_size=(256, 256), color_mode='rgb',
                            classes=None, class_mode='categorical',
                            batch_size=32, shuffle=True, seed=None,
                            save_to_dir=None,
                            save_prefix='',
                            save_format='png',
                            follow_links=False,
                            interpolation='nearest'):
        return DirectoryIterator(
            directory, self,
            target_size=target_size, color_mode=color_mode,
            classes=classes, class_mode=class_mode,
            data_format=self.data_format,
            batch_size=batch_size, shuffle=shuffle, seed=seed,
            save_to_dir=save_to_dir,
            save_prefix=save_prefix,
            save_format=save_format,
            follow_links=follow_links,
            interpolation=interpolation)

    def flow_bcolz(self, x, y=None, batch_size=32, shuffle=True, seed=None,
                   save_to_dir=None, save_prefix='', save_format='jpeg'):
        return BcolzArrayIterator(
            x, y, self,
            batch_size=batch_size,
            shuffle=shuffle,
            seed=seed,
            # data_format=self.data_format,
            # save_to_dir=save_to_dir,
            # save_prefix=save_prefix,
            # save_format=save_format,
            # switch_data_format=self.switch_data_format
        )

    def flow_from_filelist(self, filenames, y, batch_size=32,
                           image_size=256, cropped_image_size=224,
                           color_mode='rgb',
                           shuffle=True,
                           seed=None, save_to_dir=None, save_prefix='',
                           save_format='jpeg'):
        return ImageListIterator(filenames, y, self, image_size=image_size,
                                 cropped_image_size=cropped_image_size,
                                 color_mode=color_mode,
                                 batch_size=batch_size, shuffle=shuffle,
                                 seed=seed, data_format=self.data_format,
                                 save_to_dir=save_to_dir,
                                 save_prefix=save_prefix,
                                 save_format=save_format)

    def standardize(self, x):
        """Apply the normalization configuration to a batch of inputs.

        # Arguments
            x: batch of inputs to be normalized.

        # Returns
            The inputs, normalized.
        """
        if self.preprocessing_function:
            x = self.preprocessing_function(x)
        if self.rescale:
            x *= self.rescale
        if self.samplewise_center:
            x -= np.mean(x, keepdims=True)
        if self.samplewise_std_normalization:
            x /= np.std(x, keepdims=True) + 1e-7

        if self.featurewise_center:
            if self.mean is not None:
                x -= self.mean
            else:
                warnings.warn('This ImageDataGenerator specifies '
                              '`featurewise_center`, but it hasn\'t '
                              'been fit on any training data. Fit it '
                              'first by calling `.fit(numpy_data)`.')
        if self.featurewise_std_normalization:
            if self.std is not None:
                x /= (self.std + 1e-7)
            else:
                warnings.warn('This ImageDataGenerator specifies '
                              '`featurewise_std_normalization`, but it hasn\'t '
                              'been fit on any training data. Fit it '
                              'first by calling `.fit(numpy_data)`.')
        if self.zca_whitening:
            if self.principal_components is not None:
                flatx = np.reshape(x, (-1, np.prod(x.shape[-3:])))
                whitex = np.dot(flatx, self.principal_components)
                x = np.reshape(whitex, x.shape)
            else:
                warnings.warn('This ImageDataGenerator specifies '
                              '`zca_whitening`, but it hasn\'t '
                              'been fit on any training data. Fit it '
                              'first by calling `.fit(numpy_data)`.')
        return x

    def random_transform(self, x, seed=None):
        """Randomly augment a single image tensor.

        # Arguments
            x: 3D tensor, single image.
            seed: random seed.

        # Returns
            A randomly transformed version of the input (same shape).
        """
        # x is a single image, so it doesn't have image number at index 0
        img_row_axis = self.row_axis - 1
        img_col_axis = self.col_axis - 1
        img_channel_axis = self.channel_axis - 1

        if seed is not None:
            np.random.seed(seed)

        # use composition of homographies
        # to generate final transform that needs to be applied
        if self.rotation_range:
            theta = np.pi / 180 * np.random.uniform(-self.rotation_range,
                                                    self.rotation_range)
        else:
            theta = 0

        if self.height_shift_range:
            tx = np.random.uniform(-self.height_shift_range,
                                   self.height_shift_range) * x.shape[
                     img_row_axis]
        else:
            tx = 0

        if self.width_shift_range:
            ty = np.random.uniform(-self.width_shift_range,
                                   self.width_shift_range) * x.shape[
                     img_col_axis]
        else:
            ty = 0

        if self.shear_range:
            shear = np.random.uniform(-self.shear_range, self.shear_range)
        else:
            shear = 0

        if self.zoom_range[0] == 1 and self.zoom_range[1] == 1:
            zx, zy = 1, 1
        else:
            zx, zy = np.random.uniform(self.zoom_range[0], self.zoom_range[1],
                                       2)

        transform_matrix = None
        if theta != 0:
            rotation_matrix = np.array([[np.cos(theta), -np.sin(theta), 0],
                                        [np.sin(theta), np.cos(theta), 0],
                                        [0, 0, 1]])
            transform_matrix = rotation_matrix

        if tx != 0 or ty != 0:
            shift_matrix = np.array([[1, 0, tx],
                                     [0, 1, ty],
                                     [0, 0, 1]])
            transform_matrix = shift_matrix if transform_matrix is None else np.dot(
                transform_matrix, shift_matrix)

        if shear != 0:
            shear_matrix = np.array([[1, -np.sin(shear), 0],
                                     [0, np.cos(shear), 0],
                                     [0, 0, 1]])
            transform_matrix = shear_matrix if transform_matrix is None else np.dot(
                transform_matrix, shear_matrix)

        if zx != 1 or zy != 1:
            zoom_matrix = np.array([[zx, 0, 0],
                                    [0, zy, 0],
                                    [0, 0, 1]])
            transform_matrix = zoom_matrix if transform_matrix is None else np.dot(
                transform_matrix, zoom_matrix)

        if transform_matrix is not None:
            h, w = x.shape[img_row_axis], x.shape[img_col_axis]
            transform_matrix = transform_matrix_offset_center(transform_matrix,
                                                              h, w)
            x = apply_transform(x, transform_matrix, img_channel_axis,
                                fill_mode=self.fill_mode, cval=self.cval)

        if self.channel_shift_range != 0:
            x = random_channel_shift(x,
                                     self.channel_shift_range,
                                     img_channel_axis)
        if self.horizontal_flip:
            if np.random.random() < 0.5:
                x = flip_axis(x, img_col_axis)

        if self.vertical_flip:
            if np.random.random() < 0.5:
                x = flip_axis(x, img_row_axis)

        return x

    def fit(self, x,
            augment=False,
            rounds=1,
            seed=None):
        """Fits internal statistics to some sample data.

        Required for featurewise_center, featurewise_std_normalization
        and zca_whitening.

        # Arguments
            x: Numpy array, the data to fit on. Should have rank 4.
                In case of grayscale data,
                the channels axis should have value 1, and in case
                of RGB data, it should have value 3.
            augment: Whether to fit on randomly augmented samples
            rounds: If `augment`,
                how many augmentation passes to do over the data
            seed: random seed.

        # Raises
            ValueError: in case of invalid input `x`.
        """
        x = np.asarray(x, dtype=K.floatx())
        if x.ndim != 4:
            raise ValueError('Input to `.fit()` should have rank 4. '
                             'Got array with shape: ' + str(x.shape))
        if x.shape[self.channel_axis] not in {1, 3, 4}:
            warnings.warn(
                'Expected input to be images (as Numpy array) '
                'following the data format convention "' + self.data_format + '" '
                                                                              '(channels on axis ' + str(
                    self.channel_axis) + '), i.e. expected '
                                         'either 1, 3 or 4 channels on axis ' + str(
                    self.channel_axis) + '. '
                                         'However, it was passed an array with shape ' + str(
                    x.shape) +
                ' (' + str(x.shape[self.channel_axis]) + ' channels).')

        if seed is not None:
            np.random.seed(seed)

        x = np.copy(x)
        if augment:
            ax = np.zeros(tuple([rounds * x.shape[0]] + list(x.shape)[1:]),
                          dtype=K.floatx())
            for r in range(rounds):
                for i in range(x.shape[0]):
                    ax[i + r * x.shape[0]] = self.random_transform(x[i])
            x = ax

        if self.featurewise_center:
            self.mean = np.mean(x, axis=(0, self.row_axis, self.col_axis))
            broadcast_shape = [1, 1, 1]
            broadcast_shape[self.channel_axis - 1] = x.shape[self.channel_axis]
            self.mean = np.reshape(self.mean, broadcast_shape)
            x -= self.mean

        if self.featurewise_std_normalization:
            self.std = np.std(x, axis=(0, self.row_axis, self.col_axis))
            broadcast_shape = [1, 1, 1]
            broadcast_shape[self.channel_axis - 1] = x.shape[self.channel_axis]
            self.std = np.reshape(self.std, broadcast_shape)
            x /= (self.std + K.epsilon())

        if self.zca_whitening:
            flat_x = np.reshape(x, (
            x.shape[0], x.shape[1] * x.shape[2] * x.shape[3]))
            sigma = np.dot(flat_x.T, flat_x) / flat_x.shape[0]
            u, s, _ = linalg.svd(sigma)
            self.principal_components = np.dot(
                np.dot(u, np.diag(1. / np.sqrt(s + self.zca_epsilon))), u.T)


class NumpyArrayIterator(Iterator):
    """Iterator yielding data from a Numpy array.

    # Arguments
        x: Numpy array of input data.
        y: Numpy array of targets data.
        image_data_generator: Instance of `ImageDataGenerator`
            to use for random transformations and normalization.
        batch_size: Integer, size of a batch.
        shuffle: Boolean, whether to shuffle the data between epochs.
        seed: Random seed for data shuffling.
        data_format: String, one of `channels_first`, `channels_last`.
        save_to_dir: Optional directory where to save the pictures
            being yielded, in a viewable format. This is useful
            for visualizing the random transformations being
            applied, for debugging purposes.
        save_prefix: String prefix to use for saving sample
            images (if `save_to_dir` is set).
        save_format: Format to use for saving sample images
            (if `save_to_dir` is set).
    """

    def __init__(self, x, y, image_data_generator,
                 batch_size=32, shuffle=False, seed=None,
                 data_format=None,
                 save_to_dir=None, save_prefix='', save_format='jpeg',
                 switch_data_format=False):
        if y is not None and len(x) != len(y):
            raise ValueError('X (images tensor) and y (labels) '
                             'should have the same length. '
                             'Found: X.shape = %s, y.shape = %s' %
                             (np.asarray(x).shape, np.asarray(y).shape))

        if data_format is None:
            data_format = K.image_data_format()
        # self.x = np.asarray(x, dtype=K.floatx())
        self.x = x

        if self.x.ndim != 4:
            raise ValueError('Input data in `NumpyArrayIterator` '
                             'should have rank 4. You passed an array '
                             'with shape', self.x.shape)
        channels_axis = 3 if data_format == 'channels_last' else 1
        if self.x.shape[channels_axis] not in {1, 3, 4}:
            raise ValueError('NumpyArrayIterator is set to use the '
                             'data format convention "' + data_format + '" '
                                                                        '(channels on axis ' + str(
                channels_axis) + '), i.e. expected '
                                 'either 1, 3 or 4 channels on axis ' + str(
                channels_axis) + '. '
                                 'However, it was passed an array with shape ' + str(
                self.x.shape) +
                             ' (' + str(
                self.x.shape[channels_axis]) + ' channels).')
        if y is not None:
            self.y = np.asarray(y)
        else:
            self.y = None
        self.image_data_generator = image_data_generator
        self.data_format = data_format
        self.save_to_dir = save_to_dir
        self.save_prefix = save_prefix
        self.save_format = save_format
        self.switch_data_format = switch_data_format
        super(NumpyArrayIterator, self).__init__(x.shape[0], batch_size,
                                                 shuffle, seed)

    def next(self):
        """For python 2.x.

        # Returns
            The next batch.
        """
        # Keeps under lock only the mechanism which advances
        # the indexing of each batch.
        with self.lock:
            index_array, current_index, current_batch_size = next(
                self.index_generator)
        # The transformation of images is not under thread lock
        # so it can be done in parallel
        batch_x = np.zeros(tuple([current_batch_size] + list(self.x.shape)[1:]),
                           dtype=K.floatx())
        for i, j in enumerate(index_array):
            x = self.x[j]
            x = self.image_data_generator.random_transform(x.astype(K.floatx()))
            x = self.image_data_generator.standardize(x)
            batch_x[i] = x
        if self.save_to_dir:
            for i in range(current_batch_size):
                img = array_to_img(batch_x[i], self.data_format, scale=False)
                fname = '{prefix}_{index}_{hash}.{format}'.format(
                    prefix=self.save_prefix,
                    index=current_index + i,
                    hash=np.random.randint(1e4),
                    format=self.save_format)
                img.save(os.path.join(self.save_to_dir, fname))
        if self.y is None:
            return batch_x
        batch_y = self.y[index_array]
        if self.switch_data_format:
            batch_x = np.swapaxes(batch_x, 1, 3)
        return batch_x, batch_y


class ImageListIterator(Iterator):
    """Iterator capable of reading images given by a filelist from disk.

    # Arguments
        filenames: list containing full paths to image files.
        labels: labels
        image_data_generator: Instance of `ImageDataGenerator`
            to use for random transformations and normalization.
        image_size: image file will be resized to image_size x image_size,
                    with padding, keeping the aspect ratio
        cropped_image_size: if the image will be cropped, the final image
                    size will be cropped_image_size x cropped_image_size
        color_mode: One of `"rgb"`, `"grayscale"`. Color mode to read images.
        batch_size: Integer, size of a batch.
        shuffle: Boolean, whether to shuffle the data between epochs.
        seed: Random seed for data shuffling.
        data_format: String, one of `channels_first`, `channels_last`.
        save_to_dir: Optional directory where to save the pictures
            being yielded, in a viewable format. This is useful
            for visualizing the random transformations being
            applied, for debugging purposes.
        save_prefix: String prefix to use for saving sample
            images (if `save_to_dir` is set).
        save_format: Format to use for saving sample images
            (if `save_to_dir` is set).
    """

    def __init__(self, filenames, labels, image_data_generator,
                 image_size=256, cropped_image_size=256, color_mode='rgb',
                 batch_size=32, shuffle=True, seed=None,
                 data_format=None,
                 save_to_dir=None, save_prefix='', save_format='jpeg'):
        if data_format is None:
            data_format = K.image_data_format()
        self.filenames = filenames
        self.labels = labels
        self.image_data_generator = image_data_generator
        self.target_size = (image_size, image_size)
        self.final_size = (cropped_image_size, cropped_image_size)
        if color_mode not in {'rgb', 'grayscale'}:
            raise ValueError('Invalid color mode:', color_mode,
                             '; expected "rgb" or "grayscale".')
        self.color_mode = color_mode
        self.data_format = data_format
        if self.color_mode == 'rgb':
            if self.data_format == 'channels_last':
                self.image_shape = self.final_size + (3,)
            else:
                self.image_shape = (3,) + self.final_size
        else:
            if self.data_format == 'channels_last':
                self.image_shape = self.final_size + (1,)
            else:
                self.image_shape = (1,) + self.final_size
        self.save_to_dir = save_to_dir
        self.save_prefix = save_prefix
        self.save_format = save_format
        self.samples = len(self.filenames)
        assert self.samples == len(self.labels)
        super(ImageListIterator, self).__init__(self.samples, batch_size,
                                                shuffle, seed)

    def _get_batches_of_transformed_samples(self, index_array):
        batch_x = np.zeros((len(index_array),) + self.image_shape,
                           dtype=K.floatx())
        # build batch of image data
        for i, j in enumerate(index_array):
            fname = self.filenames[j]
            img = resize_and_pad(fname, desired_size=self.target_size[0])
            x = img_to_array(img, data_format=self.data_format)
            # x = self.image_data_generator.random_transform(x)
            x = self.image_data_generator.standardize(x)
            batch_x[i] = x
        # optionally save augmented images to disk for debugging purposes
        if self.save_to_dir:
            for i, j in enumerate(index_array):
                img = array_to_img(batch_x[i], self.data_format, scale=True)
                fname = '{prefix}_{index}_{hash}.{format}'.format(
                    prefix=self.save_prefix,
                    index=j,
                    hash=np.random.randint(1e7),
                    format=self.save_format)
                img.save(os.path.join(self.save_to_dir, fname))

        batch_y = self.labels[index_array]
        return batch_x, batch_y

    def next(self):
        """For python 2.x.

        # Returns
            The next batch.
        """
        with self.lock:
            index_array = next(self.index_generator)
        # The transformation of images is not under thread lock
        # so it can be done in parallel
        return self._get_batches_of_transformed_samples(index_array)