CODEWITHSUNDEEP
pythontensorflowkeras
mad
mad

Reputation: 2789

How does data augmentation work in Keras?

I have the following code to do data augmentation using my data in a list as input:

from keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img import PIL

def augment(file_images_path, dir_save):

    datagen = 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, fill_mode='nearest')

    with open(file_images_path) as f:

       images_names = f.readlines()
       images_names = [x.strip() for x in images_names]
       for line in images_names:
           img=PIL.Image.open(line)             
           img=img.resize((28,28))                        
           x = img_to_array(img)                                                    
           x = x.reshape((1,) + x.shape)        
           # the .flow() command below generates batches of randomly transformed 
           #images and saves the results to the `dir_save` directory            
           i = 0            
           for batch in datagen.flow(x, batch_size=1, save_to_dir=dir_save, save_prefix='augmented', save_format='tif'):
                i += 1
                if i > 2:
                    break  # otherwise the generator would loop indefinitely

I am quite rookie in data augmentation in Keras and I want to know how many image operations does Keras perform on my images per iteration. For example, if I run this code on a list containing 14 images it will generate 126 augmented images. If I run it on a list containing 125 images, it will generate 370 augmented images. My question is: why?

Upvotes: 3

Views: 959

Answers (1)

Jonas Adler
Jonas Adler

Reputation: 10759

If you use data augmentation in Keras then each time you generate some data, the data will be slightly modified.

Now some data augmentation steps have a finite number of options (e.g. you can either flip the image or you don't), so using these would perhaps double the amount of images you have.

Others have a (practically) infinite number of options. For example, when you specify rotation_range=40, this means that each time you generate an image, this image will be rotated with a randomly chosen angle between -40 and 40 degrees.

Hence, with the data augmentation you have used, you actually have infinitely many different images that can be generated. However, these are going to be highly correlated, and is obviously not as good as actually having infinitely many images.

Upvotes: 4

Related Questions