Specifying the probability of applying an augmentation layer

Short Description The idea is to specify the probability rate at which a generic augmentation layer will be applied to the inputs. This can be achieved in two ways:

1. Having a RanomApply layer which accepts the Augmentation Layer and the probability rate. The augmentation layer will be applied in accordance to the probability provided.

Pseudocode:

custom_random_flip = RandomApply(
    layer=RandomFlip(),
    rate=0.8
)

This would signify that we want our custom_random_flip layer to be applied with a probability of 0.8.

1. Accepting a rate argument for all the augmentation layers prefixed Random. The suggestion is based on the assumption that these layers have a without Random counterpart, which does not need any probability rate. The Random<AUGMENATION_LAYER> would accept the rate argument which would define with what probability should this layer be applied to the inputs.

Existing Implementations

• PyTorch transform RandomApply
• PyToch transform RandomHorizontalFlip with the p arguments.

[sayakpaul]

custom_random_flip = RandomApply(
    layer=RandomFlip(),
    rate=0.8
)

An extension of this could be to allow the users to specify a list of augmentation transforms and the ability to set their probabilities. If individual probabilities are not specified default will be picked.

Just a minor nit for other participants:

RandomFlip in keras.layers applies flipping with a 50-50 chance already.

[sebastian-sz]

I think you mean that RandomApply works similar to MaybeApply, but RandomApply works on the entire batch, while MaybeApply works on individual inputs within the batch?

I think RandomApply would be interesting addition 😃