Keras 数据增强 (BreakHis)

Question

My problem is to amplify the BreakHis dataset ( https://www.kaggle.com/ambarish/breakhis ) using the following instructions:

We chose the DC subclass as the baselihne, and amplified each of the remaining sublclasses by turning images up and down, left and right, and using counterclowise rotatiotion of 90 ° and 180 °.

Which instructions should I write in Keras to perform the data augmentation? Obviously, I would like to understand how to do those operations in general.

This is the distribution of the original dataset:

This is the distribution of the augmented dataset:

Answer 1

The following function will apply wanted operations with equal probability, if you want to apply more than 1 operation for an array you can alter the code. We provide the number of augmentations to the code as augmentation_number and the image array. Then this function applies transformations on it.

import random
import numpy as np
 
def generate_rotated(array,augmentation_number):
    m=0
    All=[]
    while m!=augmentation_number:
        rand=random.random()    
        if rand<=0.25:
            Array=np.rot90(array)
        elif 0.25<rand and rand<=0.5:
            Array=np.rot90(array,2)
        elif 0.5<rand and rand<=0.75:
            Array=np.flipud(array)
        else:
            Array=np.fliplr(array)
        All.append(Array)
        m+=1
    AugmentedArrays=np.array(All)
    return AugmentedArrays
       

Edit function: -Also if you want to implement a function with %50 probability you can create it like this:

def generate_rotated(array,augmentation_number):
    m=0
    All=[]
    while m!=augmentation_number:
        rand=random.random()    
        if rand<0.5:
            Array=np.rot90(array)
        rand=random.random()    
        if rand<0.5:
            Array=np.rot90(array,2)
        rand=random.random()    
        if rand<0.5:
            Array=np.flipud(array) 
        rand=random.random()    
        if rand<0.5:
            Array=np.fliplr(array)
        All.append(Array)
        m+=1
    AugmentedArrays=np.array(All)
    return AugmentedArrays

Example using:

array=np.array([[12,323,2332],[2323,23452,5195],[95310,450,3]])

generate_rotated(array,5).shape=(5, 3, 3)

generate_rotated(array,5)

I hope this answers your question.