I'm fitting full convolutional network on some image data for semantic segmentation using Keras. However, I'm having some problems overfitting. I don't have that much data and I want to do data augmentation. However, as I want to do pixel-wise classification, I need any augmentations like flips, rotations, and shifts to apply to both feature images and the label images. Ideally I'd like to use the Keras ImageDataGenerator for on-the-fly transformations. However, as far as I can tell, you cannot do equivalent transformations on both the feature and label data.
Does anyone know if this is the case and if not, does anyone have any ideas? Otherwise, I'll use other tools to create a larger dataset and just feed it in all at once.
Thanks!
There are works on extending ImageDataGenerator to be more flexible for exactly these type of cases (see in this issue on Github for examples).
Additionally, as mentioned by Mikael Rousson in the comments, you can easily create your own version of ImageDataGenerator yourself, while leveraging many of its built-in functions to make it easier. Here is an example code I've used for an image denoising problem, where I use random crops + additive noise to generate clean and noisy image pairs on the fly. You could easily modify this to add other types of augmentations. After which, you can use Model.fit_generator to train using these methods.
from keras.preprocessing.image import load_img, img_to_array, list_pictures
def random_crop(image, crop_size):
height, width = image.shape[1:]
dy, dx = crop_size
if width < dx or height < dy:
return None
x = np.random.randint(0, width - dx + 1)
y = np.random.randint(0, height - dy + 1)
return image[:, y:(y+dy), x:(x+dx)]
def image_generator(list_of_files, crop_size, to_grayscale=True, scale=1, shift=0):
while True:
filename = np.random.choice(list_of_files)
try:
img = img_to_array(load_img(filename, to_grayscale))
except:
return
cropped_img = random_crop(img, crop_size)
if cropped_img is None:
continue
yield scale * cropped_img - shift
def corrupted_training_pair(images, sigma):
for img in images:
target = img
if sigma > 0:
source = img + np.random.normal(0, sigma, img.shape)/255.0
else:
source = img
yield (source, target)
def group_by_batch(dataset, batch_size):
while True:
try:
sources, targets = zip(*[next(dataset) for i in xrange(batch_size)])
batch = (np.stack(sources), np.stack(targets))
yield batch
except:
return
def load_dataset(directory, crop_size, sigma, batch_size):
files = list_pictures(directory)
generator = image_generator(files, crop_size, scale=1/255.0, shift=0.5)
generator = corrupted_training_pair(generator, sigma)
generator = group_by_batch(generator, batch_size)
return generator
You can then use the above like so:
train_set = load_dataset('images/train', (patch_height, patch_width), noise_sigma, batch_size)
val_set = load_dataset('images/val', (patch_height, patch_width), noise_sigma, batch_size)
model.fit_generator(train_set, samples_per_epoch=batch_size * 1000, nb_epoch=nb_epoch, validation_data=val_set, nb_val_samples=1000)