Tensorflow: Iterate row-wise, elementwise multiplication, reduce-sum. Custom convolution

Question

I've got two tensors,

x = shape(batchsize, 29, 64), 
y = shape(batchsize, 29, 29, 64)

I want to iterate row-wise over y, perform an elementwise multiplication with x, sum_reduce the result, and stack those results to a new tensor. The result should be of a shape (batchsize, 29, 64). It's quite similar to a convolution.

How I would program it sequentially:

for batchnr in range(x.shape[0]): 
    for n in range(y.shape[1]):
        temp = tf.multiply(x[batchnr][n], y[batchnr]) #shape(29,64)
        prod = tf.reduce_sum(temp) # shape(1,64)
        res[batchnr][n] = prod

I've created that explanation figure: Since reduce_sum is done for each row, the result is a tensor of shape (batchsize, 29, 64) again.

I can't figure out how to do it right and efficient. Thank you.

Answer 1

I think i've found the solution. Instead of iterating y:

• amplify x to shape of y
• multiply
• reduce

In code it looks like:

m = K.tf.constant([1, 29, 1], dtype=K.tf.int32)
x = K.tf.tile(Z_RBF[0], m) #vermehrfache Z und stacke es zu RBF shape
x = K.tf.reshape(x, shape=(-1, *Z_RBF[1].shape[1:]))
x = K.tf.multiply(x, Z_RBF[1])
x = K.tf.reduce_sum(x, axis=2)  # shape (batchsize, 29, 64)

.