I am using a custom image generator to read my data off disk in batches as described here, https://keras.io/examples/vision/oxford_pets_image_segmentation/
The exact generator looks like this:
from tensorflow import keras
import numpy as np
from tensorflow.keras.preprocessing.image import load_img
import tensorflow
class OxfordPets(keras.utils.Sequence):
"""Helper to iterate over the data (as Numpy arrays)."""
def __init__(self, batch_size, img_size, input_img_paths, target_img_paths):
self.batch_size = batch_size
self.img_size = img_size
self.input_img_paths = input_img_paths
self.target_img_paths = target_img_paths
def __len__(self):
return len(self.target_img_paths) // self.batch_size
def __getitem__(self, idx):
"""Returns tuple (input, target) correspond to batch #idx."""
i = idx * self.batch_size
batch_input_img_paths = self.input_img_paths[i : i + self.batch_size]
batch_target_img_paths = self.target_img_paths[i : i + self.batch_size]
x = np.zeros((self.batch_size,) + self.img_size + (3,), dtype="float32")
for j, path in enumerate(batch_input_img_paths):
img = load_img(path, target_size=self.img_size)
x[j] = img
y = np.zeros((self.batch_size,) + self.img_size + (1,), dtype="uint8")
for j, path in enumerate(batch_target_img_paths):
img = load_img(path, target_size=self.img_size, color_mode="grayscale")
y[j] = np.expand_dims(img, 2)
# Ground truth labels are 1, 2, 3. Subtract one to make them 0, 1, 2:
y[j] -= 1
return x, y
This works great and is helpful as my data is too large to read into ram. It takes two paths which are pathways to the input images (batch_input_img_paths) and the masks (batch_target_img_paths).
I would like to modify this generator to use data augmentation. I am trying this:
class OxfordPets(keras.utils.Sequence): """Helper to iterate over the data (as Numpy arrays)."""
def __init__(self, batch_size, img_size, input_img_paths, target_img_paths):
self.batch_size = batch_size
self.img_size = img_size
self.input_img_paths = input_img_paths
self.target_img_paths = target_img_paths
def __len__(self):
return len(self.target_img_paths) // self.batch_size
def __data_augmentation(self, img):
''' function for apply some data augmentation '''
img = tensorflow.keras.preprocessing.image.random_shift(img, 0.2, 0.2)
img = tensorflow.keras.preprocessing.image.random_zoom(img, 0.2)
img = tensorflow.keras.preprocessing.image.random_shear(img, 0.2)
img = tensorflow.keras.preprocessing.image.random_rotation(img, 40)
img = tensorflow.image.random_flip_left_right(img)
img = tensorflow.image.random_flip_up_down(img)
return img
def __getitem__(self, idx):
"""Returns tuple (input, target) correspond to batch #idx."""
i = idx * self.batch_size
batch_input_img_paths = self.input_img_paths[i : i + self.batch_size]
batch_target_img_paths = self.target_img_paths[i : i + self.batch_size]
x = np.zeros((self.batch_size,) + self.img_size + (3,), dtype="float32")
for j, path in enumerate(batch_input_img_paths):
img = load_img(path, target_size=self.img_size)
#apply augmentation
img = self.__data_augmentation(img)
x[j] = img
y = np.zeros((self.batch_size,) + self.img_size + (1,), dtype="uint8")
for j, path in enumerate(batch_target_img_paths):
img = load_img(path, target_size=self.img_size, color_mode="grayscale")
y[j] = np.expand_dims(img, 2)
#apply augmentation
img = self.__data_augmentation(img)
# Ground truth labels are 1, 2, 3. Subtract one to make them 0, 1, 2:
y[j] -= 1
return x, y
but this returns:
AttributeError: 'Image' object has no attribute 'shape'
Another thing I am not positive about is if the exact same augmentation will be applied to the data and to the label. Perhaps a random seed is needed for this? Another thing I would like to do is not only feed in augmented images into the final network but to increase the training size with the augmentation, so that some real images and some augmented images are being fed in, for instance only augment 30% of the input images.
EDIT:
I think I got the augmentation working like this:
class OxfordPets(keras.utils.Sequence):
"""Helper to iterate over the data (as Numpy arrays)."""
def __init__(self, batch_size, img_size, input_img_paths, target_img_paths):
self.batch_size = batch_size
self.img_size = img_size
#augmentation
self.augmentor = tensorflow.keras.preprocessing.image.ImageDataGenerator(
rotation_range=40,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True)
self.input_img_paths = input_img_paths
self.target_img_paths = target_img_paths
def __len__(self):
return len(self.target_img_paths) // self.batch_size
def __getitem__(self, idx):
"""Returns tuple (input, target) correspond to batch #idx."""
i = idx * self.batch_size
batch_input_img_paths = self.input_img_paths[i : i + self.batch_size]
batch_target_img_paths = self.target_img_paths[i : i + self.batch_size]
x = np.zeros((self.batch_size,) + self.img_size + (3,), dtype="float32")
for j, path in enumerate(batch_input_img_paths):
img = load_img(path, target_size=self.img_size)
x[j] = img
y = np.zeros((self.batch_size,) + self.img_size + (1,), dtype="uint8")
for j, path in enumerate(batch_target_img_paths):
img = load_img(path, target_size=self.img_size, color_mode="grayscale")
y[j] = np.expand_dims(img, 2)
# Ground truth labels are 1, 2, 3. Subtract one to make them 0, 1, 2:
y[j] -= 1
return x, y
but this does not solve my problem of actually increasing the dataset size, I believe it is only sending augmented images in now, and no real data.
