Data augmentation for multiple images

vision
1

I’m beginner for Pytorch. I want to increase the number of datasets (data augmentation).
I have two images (cat_1 and cat_2) in folder and I want to use albumentation to increase the number of images as follows:

import cv2
import torch
import albumentations as A
import numpy as np
import matplotlib.pyplot as plt
import os
from os import listdir
from os.path import join, isfile
from torchvision.utils import save_image

# Read files from Directory
Dir_img = '/content/drive/MyDrive/Cat_Dog'
files = os.listdir(Dir_img)
print(files) # ['cat_1.jpg', 'cat_2.jpg']

After that, I use OpenCV to read images from Dir_img and keep array of each image into list as follows:

dataset = [] # <class 'list'> In list consist of array of each image 
             # len(dataset) = 2

for img in files:
  img_ = cv2.imread(os.path.join(Dir_img,img), cv2.IMREAD_UNCHANGED) # default BGR
  Org_ = cv2.cvtColor(img_, cv2.COLOR_BGR2RGB) # Convert BGR -> RGB
  dataset.append(Org_)

  # cat_1: dataset[0].shape = (666, 1000, 3)
  # cat_2: dataset[1].shape = (825, 1100, 3)

Next, I defined an augmentation pipeline (multiple files) as follows:

TRANSFORM = A.Compose([
    A.Blur(p = 0.5, blur_limit = (3, 7)),  
    A.RandomBrightnessContrast(p = 0.5, brightness_limit = 1, contrast_limit = 1)
])

After that, I will pass all images to the augmentation pipeline and receive augmented images.

for IMG in dataset: # IMG is <class 'numpy.ndarray'>
  MY_TRANSFORMED = TRANSFORM(image=IMG)  #<class 'dict'>

  #Convert 'dict' to 'numpy.array'
  MY_TRANSFORMED_IMAGE = MY_TRANSFORMED['image']  # <class 'numpy.ndarray'>
  
  MY_TRANSFORMED_IMAGE_1 = TRANSFORM(image=IMG)['image']
  MY_TRANSFORMED_IMAGE_2 = TRANSFORM(image=IMG)['image']
  MY_TRANSFORMED_IMAGE_3 = TRANSFORM(image=IMG)['image']

From above code, I received a total of 3 images of different cat_2. But I think the resulting image should be of a cat_1.

I have questions as follows:

  1. Every time, I run TRANSFORM. I receive a different image every time. Is there a way to get me the same results every time?
  2. Why haven’t I received an image of a cat_1?
  3. If I want 1 original images (cat_1 and cat_2) to do 10 different data augmentations, I can use a for loop ?

I’ve tried to find document from various sources, but I’m not getting the answer I’m wondering.

2

  1. If you are using random transformations you would need to seed the corresponding library.

  2. Are you checking the last output after the for loop was executed? If so, note that the second iteration will transform cat_2 while the first iteration should be transforming cat_1. Add a debug print statement to check which image is currently being transformed inside the loop.

  3. Yes.