how to process data generated by ImageDataGenerator?

Question

I am trying to use Keras ImageDataGenerator to train my model with a large stereo Dataset.

For each scene, I have two rgb images, I have to split them and concatenate them to have 6 one-channel images as an input of my model (i.e. this shape (6,224,224,1)). For small datasets, it is easy because I can upload the two sub-datasets in the memory and process the concatenation of images as ndarrays. But with ImageDataGenerator, it is not the same, as I have to make sure that it takes the same batches from the two sub-datasets, and be able to process the concatenation before passing the input to my model.

Inspired from this post, I tried this code:

input_imgen = ImageDataGenerator()

def generate_generator_multiple(generator,dir1, dir2, batch_size):
    genX1 = generator.flow_from_directory(directory=dir1,
                                          color_mode="rgb",
                                          batch_size=batch_size,
                                          class_mode="categorical",
                                          shuffle=False)



    genX2 = generator.flow_from_directory(directory=dir2,
                                          color_mode="rgb",
                                          batch_size=batch_size,
                                          class_mode="categorical",
                                          shuffle=False)
    while True:
            X1i = genX1.next()
            X2i = genX2.next()
            yield [X1i[0], X2i[0]], X2i[1]

In my case, how could I process the two series X1i[0] and X2i[0] with batches of images of size( 224,224,3) to have X1Sum of size(6,224,224), and instead of: yield [X1i[0], X2i[0]], X2i[1] I will have: yield X1Sum, X2i[1]

Can't you do `np.concatenate(X1i, X2i)` and reshape ? I am guessing `X1i` and `X2i` have shapes `(224, 224, 3)` each. Or am i wrong in my assumption? — Anakin, May 07 '19 at 12:04
X1i and X2i are series of images of size (224, 224, 3), I am not sure that I can do np.concatenate(list(X1i), list(X2i)). — E.gh, May 07 '19 at 12:43

score 1 · Accepted Answer · answered May 07 '19 at 13:09

It looks to me that you can simply convert the list to arrays and then use np.concatenate to merge the channels and np.swapaxes to put the channels in front. Example code with output:

import numpy as np

list_a = []
list_b = []

for i in range(5):
    list_a.append(np.random.rand(4, 4, 3))
    list_b.append(np.random.rand(4, 4, 3))

a = np.array(list_a)
b = np.array(list_b)
print(a.shape)
print(b.shape)
c = np.concatenate((a, b), axis=-1)
print(c.shape)
c = np.swapaxes(c, 0, -1)
print(c.shape)

Output is:

(5, 4, 4, 3)
(5, 4, 4, 3)
(5, 4, 4, 6)
(6, 4, 4, 5)

I guess this is what you want to do.

how to process data generated by ImageDataGenerator?

1 Answers1