Custom transformer that does transform in batches

I am wondering if there is a way to have custom transformer do transformation in batch instead of per image.

Currently, I am writing custom code in do_transform(self, x, is_y): method.

It is completely possible. You should pass it to the DeviceDataLoader by using the method add_tfm (and then remove_tfm when you’re done). It should take a batch b as an argument (then x = b[0] and y = b[1]).
Oh but by looking at your is_y argument you may be in v0.7 still. What I’m talking about is only possible in v1.