If I am already collating my mini-batch, what should I use instead of a batch_tfms?

batch_tfms want you to return a list of examples ready to be mini-batched … but what if you’ve already made your minibatch?

For example, I have a batch transform that returns this …

All I really need is to make sure all the tensors end up on the right device.

Obviously, this throws an exception because again, a batch_tfm expects to return a list of mini-batch ready examples … not a mini-batch as I have above.

So my question is … how can I use the DataBlock API and handle mini-batch creation myself?