150_151_custom_data_augmentation.py has seeding problem

[moienr]

since your code, changes the random seed to do the same transformation on both image and mask, it will cause the number = random.randint(0, len(images)) to have only two different outputs, which results in only two images being used for data augmentation.

to solve this, I changed the transformation function, to get image and mask as input arguments and removed the seed, so when for example an angle is generated through random.randint() function uses that angle on both mask and image, instead of calling the function twice with a seed.

the code below works just fine, without using a seed. thank you for all your amazing tutorials.

#https://youtu.be/k4TqxHteJ7s
#https://youtu.be/mwN2GGA4mqo
"""
@author: Sreenivas Bhattiprolu
"""

import numpy as np
from matplotlib import pyplot as plt
from skimage.transform import AffineTransform, warp
from skimage import io, img_as_ubyte
import random
import os
from scipy.ndimage import rotate

images_to_generate=1000
seed_for_random = 42

#Define functions for each operation
#Define seed for random to keep the transformation same for image and mask

# Make sure the order of the spline interpolation is 0, default is 3. 
#With interpolation, the pixel values get messed up.
def rotation(image, mask):

    angle= random.randint(-180,180)
    r_img = rotate(image, angle, mode='reflect', reshape=False, order=0)
    r_msk = rotate(mask, angle, mode='reflect', reshape=False, order=0)
    return r_img,r_msk

def h_flip(image, mask):
    hflipped_img= np.fliplr(image)
    hflipped_msk= np.fliplr(mask)
    return  hflipped_img,hflipped_msk

def v_flip(image, mask):
    vflipped_img= np.flipud(image)
    vflipped_msk= np.flipud(mask)
    return vflipped_img,vflipped_msk

def v_transl(image, mask):
    n_pixels = random.randint(-64,64)
    vtranslated_img = np.roll(image, n_pixels, axis=0)
    vtranslated_msk = np.roll(mask, n_pixels, axis=0)
    return vtranslated_img,vtranslated_msk

def h_transl(image, mask):
    n_pixels = random.randint(-64,64)
    htranslated_img = np.roll(image, n_pixels, axis=1)
    htranslated_msk = np.roll(mask, n_pixels, axis=1)
    return htranslated_img,htranslated_msk

transformations = {'rotate': rotation,
                      'horizontal flip': h_flip, 
                      'vertical flip': v_flip,
                   'vertical shift': v_transl,
                   'horizontal shift': h_transl
                 }                #use dictionary to store names of functions 

images_path="E:\_palm\dataset_v2\patched_has_tree\w_palm\images\\" #path to original images
masks_path = "E:\_palm\dataset_v2\patched_has_tree\w_palm\masks\\"
img_augmented_path="E:\_palm\dataset_v2\patched_has_tree\w_palm\\augmented\images\\" # path to store aumented images
msk_augmented_path="E:\_palm\dataset_v2\patched_has_tree\w_palm\\augmented\masks\\" # path to store aumented images
images=[] # to store paths of images from folder
masks=[]

for im in os.listdir(images_path):  # read image name from folder and append its path into "images" array     
    images.append(os.path.join(images_path,im))

for msk in os.listdir(masks_path):  # read image name from folder and append its path into "images" array     
    masks.append(os.path.join(masks_path,msk))

i=1   # variable to iterate till images_to_generate

print(len(images))

while i<=images_to_generate: 
    number = random.randint(0, len(images))  #PIck a number to select an image & mask
    print(number)
    image = images[number]
    mask = masks[number]
    #print(image, mask)
    #image=random.choice(images) #Randomly select an image name
    original_image = io.imread(image)
    original_mask = io.imread(mask)
    transformed_image = None
    transformed_mask = None
#     print(i)
    n = 0       #variable to iterate till number of transformation to apply
    transformation_count = random.randint(1, len(transformations)) #choose random number of transformation to apply on the image

    while n <= transformation_count:
        key = random.choice(list(transformations)) #randomly choosing method to call
          #Generate seed to supply transformation functions. 
        transformed_image,transformed_mask = transformations[key](original_image, original_mask)

        n = n + 1

    new_image_path= "%s/augmented_image_%s.png" %(img_augmented_path, i)
    new_mask_path = "%s/augmented_mask_%s.png" %(msk_augmented_path, i)   #Do not save as JPG
    io.imsave(new_image_path, transformed_image)
    io.imsave(new_mask_path, transformed_mask)
    i =i+1