keras flow_from_directory returns 0 images

Question

When I try to use the following snippet of code to try to predict on a batch of images, I get a message saying that no image were found:

test_datagen = ImageDataGenerator()

test_generator = test_datagen.flow_from_directory(
        directory='test/',
        target_size=(300, 300),
        color_mode="rgb",
        shuffle = False,
        class_mode='binary',
        batch_size=1)

filenames = test_generator.filenames

My directory structure is as follows. I first have a main directory named dogs_vs_cats in which I have two sub directories train and test containing the respective images and also the notebook which contains this code.

Answer 1

Keras generator alway looks for subfolders (representing the classes). Images insight the subfolders are associated with a class.

So when you work on C:\images\ and you have two classes, say C1, C2, you need to create subfolders C:\images\C1\ and C:\images\C2\. The directory insight the generator function should point to C:\images\.

See this post for the case of image prediction: https://stackoverflow.com/a/55991598/9524424

Answer 2

There is another option where you do not have to copy the test file into another test file:

datagen = ImageDataGenerator()

test_data = datagen.flow_from_directory('.', classes=['test'])

This solved my problem. For more info see this.

Answer 3

There could be two situations when you run ImageDataGenerator on test set of images.

Case #1: Test folder has subfolders representing the classes. Peter has answered this part.

Case #2: There are no labelled test images. In this case, you will have a single test folder which contains all the images that you want to classify.

Kyle Banks has written a tip to handle this in his blog: https://kylewbanks.com/blog/loading-unlabeled-images-with-imagedatagenerator-flowfromdirectory-keras