Applying transforms to a batch of images

Hi.

I am aware of this guide that shows how to apply transforms to multiple images, masks, etc. But I am not sure how I could extend this to use inside a more generic pipeline with tf.data.

Consider the existing guide that shows how to use albumentations with TensorFlow. It applies the augmentation transforms to single examples and for a larger dataset, it can quickly become very expensive. Rather, applying the pipeline on batches might be more efficient as stated here.

How can we achieve this i.e. how can we modify the augmentation pipeline, in this case, to allow a batch of images to be processed?

[sayakpaul]

I see. But having it supporting batches might beneficial for the tf.data pipeline since tf.numpy_function is already very expensive.

[GuillaumeLeclerc]

GPU augmentations are very useful in nodes having many GPUs (8-16). In environments with many jobs running it is not acceptable to use 20+ cores for a single GPU job as machines run out of available cores before they run out of GPUs. Having (most of the time) very expensive GPUs idle is a massive waste of resources. So even if it was slightly slower it is still a crucial addition in my opinion.