Data Augmentation Question


Seeing the example C2_W2_Lab_1_cats_v_dogs_augmentation, I have some questions but the most important one now is:

After using the ImageDataGenerator method, how many generated images are added during training and validation? Is the amount of original images incremented by the amount of generated images? Or do the same generated images replace the same original images during training? Despite having scheduled the next class, I still have this doubt.

Could you clarify this for me?

Does this help?

Hi, In that perspective, I understand that the generated images will replace some original images randomly, right?

Hi @Raphael_Garcia_Morei.
The generated images will not replace the original ones you will keep original ones and generated images will be added.
The number of generated images depends on the parameters you set when using ImageDataGenerator . The generator applies various transformations to the original images, creating new versions of those images. These transformations can include rotations, flips, shifts, brightness adjustments, etc.


1 Like

If no augmentation is applied, then, generated image will be the same as the original image. As the number of augmentation parameters increase, odds are slim that a generated image will exactly match the original image.

Please see for the details on how NumpyArrayIterator works as an example.

1 Like