Cache data on CPU memory and then perform non-deterministic GPU transforms

[hw-ju]

Hi!
In this tutorial for using GPU transform, the intermediate data (i.e., after going through a chain of deterministic transforms) is cached on GPU and then non-deterministic GPU transform is applied before feeding data into the model.

Can we cache intermediate data on CPU memory, and then perform non-deterministic GPU transforms? If yes, how to do it?

[KumoLiu]

Hi @hw-ju, yes you can. Just remove EnsureType. I think regular training in this tutorial cache data in CPU instead of GPU.
Thanks!

[KumoLiu]

I'm afraid not.
Perhaps EnsureType can have a flag to indicate whether to cache it.
cc @wyli

[KumoLiu]

I just noticed that you can use PersistentDataset with EnsureType.

[hw-ju]

@KumoLiu
If I use PersistentDataset with EnsureType, then it will cache on some persistent storage and will perform randomized_transform on GPU, corrrect?

[KumoLiu]

Yes, just put EnsureType before randomized_transform. Thanks!

[hw-ju]

@KumoLiu Thank you so much for helping me so patiently :)!!