How albumentations work with keras Sequence

Question

I have read this tutorial for using albumentations with keras sequence. The code is as follows: `

from tensorflow.python.keras.utils.data_utils import Sequence

class CIFAR10Sequence(Sequence):
    def __init__(self, x_set, y_set, batch_size, augmentations):
        self.x, self.y = x_set, y_set
        self.batch_size = batch_size
        self.augment = augmentations

    def __len__(self):
        return int(np.ceil(len(self.x) / float(self.batch_size)))

    def __getitem__(self, idx):
        batch_x = self.x[idx * self.batch_size:(idx + 1) * self.batch_size]
        batch_y = self.y[idx * self.batch_size:(idx + 1) * self.batch_size]
        
        return np.stack([
            self.augment(image=x)["image"] for x in batch_x
        ], axis=0), np.array(batch_y)

`

The thing is I don't understand how it is augmenting ( ie providing more samples ) the data. The way I see it, it is just transforming the samples in the dataset, and not generating newer ones.

Answer 1

Following the tutorial you provided you may see that the author defines AUGMENTATIONS_TRAIN and AUGMENTATIONS_TEST objects which perform the actual augmentation.

Then these objects are passed to the sequence generator above:

train_gen = CIFAR10Sequence(x_train, y_train, hparams.train_batch_size, augmentations=AUGMENTATIONS_TRAIN)

so that calling self.augment actually augments every image in the batch:

self.augment(image=x)["image"] for x in batch_x

And yes, augmentation doesn't mean creating new objects but applying random transformation to existing ones to create 'artifical' objects which are somewhat different from the originals.