不同等级的张量批量tf.matmul

Question

Given a Tensor of shape (A, B, C, D, E) and a tensor of shape (A, B, E) , I would like to do a batch multiplication with automatic broadcasting of the second tensor, such that:

    In [1]: X = tf.placeholder(dtype=tf.float32, shape=[A, B, C, D, E])
    In [2]: Y = tf.placeholder(dtype=tf.float32, shape=[A, B, E])
    In [3]: # DO SOMETHING LIKE tf.matmul(X, Y)
    Out[3]: <tf.Tensor 'MatMul_1:0' shape=(A, B, C, D) dtype=float64>

Is it even possible?

Thanks in advance

Answer 1

Broadcasting is mostly supported for those operations which makes use of element-wise computation. tf.matmul is not element-wise operation but tf.multiply is element-wise based.

Also, Tensorflow may or may not do automatic broadcasting for higher order tensors even in those operations where broadcasting is supported. In those cases, you will manually have to do it.

I have written another answer which does automatic broadcasting for higher order tensors.