CODEWITHSUNDEEP
pythonnumpytensorflowkerastensor
Khurram Sultan
Khurram Sultan

Reputation: 1

Row-wise processing of tensors in a batch

I need to get the outer product of each row of a tensor separately. The code goes like:

input1=K.placeholder(shape=(None, 12)) prod=K.map_fn(someMethod,input1)

the someMethod needs to do the following:

*def someMethod(x):*
    ## get the outerproduct row-wise of input1 #
    outer=x*x.T
    ## subtract the identity matrix from the outer product #
    diff=outer-np.eye(12) 
    ## and return the trace of the difference matrix #
    trace=np.trace(diff)
    return trace

I expect the trace value to be a scalar but the prod to be a list of input of the batch size? I am using plaidml as the backend, therefore, would like something working with numpy or keras backend to work, or perhaps tensorflow.

Upvotes: 0

Views: 1271

Answers (2)

wwilliamcook
wwilliamcook

Reputation: 311

The following example performs the requested operation on a tensor x of shape [None, None] (variable batch size and variable vector dimensions). It has been tested in Tensorflow 1.13.1 and 2.0RC (tf.placeholder must be removed for 2.0). Comments assume input shape is [None, 12] for explanatory purposes.

import tensorflow as tf

x = tf.placeholder(tf.float32, shape=[None, None])  # Remove for TF 2.0
# Explicitly perform batch-wise outer product using Einstein summation notation
outer = tf.einsum('bi, bj -> bij', x, x)
outer.shape
# TensorShape([Dimension(None), Dimension(12), Dimension(12)])
diff = outer - tf.eye(tf.shape(x)[1])
trace = tf.linalg.trace(diff)
trace.shape
# TensorShape([Dimension(None)])

As you can see, mapping a helper function over the batch dimension of the input is not required with Tensorflow. You can learn more about tf.einsum here.

Upvotes: 0

Sıddık Açıl
Sıddık Açıl

Reputation: 967

Hello and welcome to Stack Overflow.

For row-wise outer-product of matrix A use the following:

outer_product = np.matmul(A[:,:,np.newaxis], A[:,np.newaxis,:])

Upvotes: 1

Related Questions