Data transforms for specific format (e.g. torch)?

[Modexus]

Is it necessary for custom transforms to format to tensors (pytorch) in the transform itself?
.set_transform(transform) removes the previous useful set.format("torch").
From the example in the docs it seems like one needs to format to tensor themselves in the transform.

Could I use map without caching instead?
Is it possible to somehow reuse TensorFormatter that already provides the formatting to torch?
Another option would be to wrap the dataset in a torch dataset class?

What would be the suggested way of doing transforms?