Broadcasting for Ragged Tensors

TensorFlow General Discussion
,
1

I am trying to change a function I wrote to make it work for ragged tensors in addition to normal (dense) tensors. One issue I ran into is that it seems that broadcasting is not supported for ragged tensors. As a minimal example (using Python 3.7, TF 2.8.2), consider the following:

tf.linalg.matvec( tf.constant([ [[0,0,1],[0,1,0],[1,0,0]] ]), tf.ragged.constant([ [1,1,0], [0,0,1] ]) )
# InvalidArgumentError: Batch dimensions of `a` and `b` do not have the same size.
tf.linalg.matvec( tf.constant([ [[0,0,1],[0,1,0],[1,0,0]] ]), tf.constant([ [1,1,0], [0,0,1] ]) )
# <tf.Tensor: shape=(1, 2, 3), dtype=int32, numpy=array([[[0, 1, 1],[1, 0, 0]]], dtype=int32)>

My first question is: why does this not work in the ragged case? Is there a fundamental reason I am not seeing or has it just not been implemented?

In my case, I have a ragged tensor X of shape [B, None, None, C] representing a batch of B images of different sizes and same number of channels C = 3. I’m trying to apply a per-pixel linear transformation represented by a tensor A of shape [B, 1, 1, C, C] as follows:

Y = tf.linalg.matvec(A, X)
# InvalidArgumentError: Batch dimensions of `a` and `b` do not have the same size.

So my second question is: what would be a good way to do this instead?

Thanks for any advice!

3

Hi @mdehling,

Above error is due to tf.linalg.matvec function requires both inputs has to be consistent in shape and same type , where ragged tensors don’t have fixed shape in inner dimension.

I suggest you to convert the ragged to dense tensor first, pad this converted dense tensor(X) to match the shape of A, then you can perform this transformation like matmul operation and convert back it to ragged tensor as you want.Here’s a sample code for ragged tensor operation with broadcasting for your reference.

Hope this helps.Thank You.