Idea for Keras Preprocessing API

[PhilipMay]

As far as I know (in the new RFC Keras Preprocessing API) the different augmentation layers (flip, roation, ...) are always used after each other. But what about the following idea:

If you could configure a certain percentage for each layer that gives the chance to activate this layer. This way not all layers are activated the same time. The advantage is that you could configure more agressive augmentation values per layer but do not "overaugment" the images.

My assumption is that this would bring a wider range of different images and improves augmentation with better cnn performance (accuracy or whatever).

What do you think?

[RaphaelMeudec]

I agree with this, I've always loved how imgaug pipelines are built. A problem I have with the ImageDataGenerator as of today is that you can't set the randomization of the augmentations, nor define a maximum number of augmentations to perform. It can lead to too much data augmentation and force to select smaller ranges for each augmentation.

[PhilipMay]

What do others think about this? @fchollet @karmel , ...

[Dref360]

Unfortunatly, the last Keras community meeting was cancelled due to TF Summit, I'll raise this issue during the next one.
Sorry for the delay.

[Dref360]

With the new API, this could be done using a K.switch right?

[bhack]

For who Is interested also in autoaugmentation there is a thread in keras-team/autokeras#828 (comment)

[RaphaelMeudec]

@Dref360 Do you know when will be the next Keras community meeting? By the way, I'm not sure Keras is the best place to hold a flexible data augmentation pipeline. It would make more sense to hold it in tf.data or as part of the tensorflow/addons (cc @seanpmorgan) to begin with?

[LukeWood]

As far as I know (in the new RFC Keras Preprocessing API) the different augmentation layers (flip, roation, ...) are always used after each other. But what about the following idea:

If you could configure a certain percentage for each layer that gives the chance to activate this layer. This way not all layers are activated the same time. The advantage is that you could configure more agressive augmentation values per layer but do not "overaugment" the images.

My assumption is that this would bring a wider range of different images and improves augmentation with better cnn performance (accuracy or whatever).

What do you think?

Hey Phillip! I know it has been a long time, but check out what we have done in KerasCV! KerasCV API Docs