create_DAGAN_database

create_DAGAN_database.py

#!/usr/bin/env python3
# coding: utf8
import numpy as np
from pathlib import Path
import matplotlib.image as mpimg

"""
Suppose a file tree i.e:
imgs/
    |class1/
        |images of type class1.png
    |class2/
        |...
    |...

the images are normalized into numpy arrays in the form: (n_classes, n_samples, h, w, c)
"""

def create_db(images_dir_path):
    if not Path(images_dir_path).is_dir():
        print('Error')
        exit()

    dataset = []
    for d in Path(images_dir_path).glob('*'):
        # skipping simple files (readme, licences, ...)
        if not d.is_dir():
            continue
        
        classData = []
        for f in Path(d).glob('*.png'):
            img = mpimg.imread(str(f)) 
            img = img.astype(np.float)
            img /= 255.0
            img = np.reshape(img, newshape=(img.shape[0], img.shape[1], 3))
            classData.append(img)
        
        dataset.append(classData)

    dataset.sort(key=len)

    return np.array(dataset)


train_dir = '/path/to/imgs'
data = create_db(train_dir)
print("dataset shape:", data.shape)
np.save('my_database.npy', data)

hi gurujam ,
Thanks for the script
I am also preparing my dataset into 5-dim array. Please correct me if I am missing : you basically appending each image into 'classDataSet' and having 3-dim (h,w,c) . Then appending again into 'dataset' which is being 4-dim (num_example, h,w,c). isn't it supposed to be 5-dim?

Hi,
Talking about shapes implies working with array, and not lists.
Calculating the shape of an array result in the number of his elements plus their dimension (see doc
So, np.array(classDataSet).shape would give us a 4-dim array (num_samples, h, w, c).

Hi,
Reading this (a year later, seeing that people still scratch their brains with some DAGAN experiments), I feel like to continue my previous comment :
...
Following the same logic, data.shape will give you (num_classes, num samples, h, w, c)
[I also updated the script to make the things easier for you :)]
This suppose that all your images are png of the same shape.
Use np.reshape(img, newshape=(img.shape[0], img.shape[1], 1)) if your images are all black and white.