Questions about ragged batching of higher dimensional tensors

After reading the ragged batching documentation a couple of things still remain unclear to me.

The example provided is really quite helpful, but I would like to know how I can adapt it to my situation in which I have higher dimensional tensors. Let us say that, like the example in my situation I also send 3 requests but instead of sending shapes [1, 3], [1, 4], [1, 5] I have shapes [1, 3, 32, 32, 1], [1, 4, 32, 32, 1], [1, 5, 32, 32, 1].

1. Is ragged batching of higher dimensional tensors like this supported?
2. How will it then batch the examples? Will it concatenate the examples in the first (batch) dimension: [12, 32, 32, 1]? Or will it concatenate and flatten all values into a big 1-dimensional tensor: [12*32*32*1] and should I modify my model to reshape the ragged input tensor back to shape of [12, 32, 32, 1]?

[GuanLuo]

I agree that your suggestion can solve this particular higher dimensional tensor scenario with existing ragged batch expressiveness, but as you mentioned this is not the generic solution as it can’t describe tensors with multiple variable dimensions properly. And since there is already the plan for adding the new kind of batch input, I would incline to keep the existing behavior and let the new batch input to address the multi-dimensional ragged batching.

[stengoes]

Ok thanks for sharing your insights and plans for the future support of ragged batching of higher dimensional tensors.

I will close this issue now.