how to multiply by batches in keras backend or tf

Question

我有两个形状为（batch size，15,500），（batch size，500,98）的张量，我想将它们作为矩阵乘以每个批次大小的矩阵张量，以获得（batch size，15,98）我如何在tensorflow中做到这一点或keras后端，我可以使用批处理点

Answer 1

You can use

A = K.placeholder((None, 15, 500))
B = K.placeholder((None, 500, 98))

C = tf.einsum("ijk,ljn->ijn", A, B)

In [84]: C
Out[84]: <tf.Tensor 'einsum/transpose_2:0' shape=(?, 15, 98) dtype=float32>

This says contract on index j, which is what a single matrix multiplication would look like, and repeat that for index i