Accessing and Multiplying Individual Elements of a Layer's Output in Keras

Question

Problem Description:

I am trying to access the individual elements (i.e., scalar) from a softmax layer's output (with dimension (,2)) and multiply this with a tensor from another model, which has a dimension of (,10). A mock-up digram describing my problem is shown in the attached Figure.

I am using Keras with Tensorflow as the back-end. So far, my approach has been the following: Lets say the output dimension of the softmax layer is (,2) (i.e., a vector of size 2). First, my plan is to access the individual elements of this vector (tensor according to the Keras/tensorflow) using the keras.backend.gather(a_k_p,0). Where, the variable a_k_p references the SoftmaxLayer. However, gather simply gives the entire row and does not give the individual element. So, my first question is how to access individual elements of a layer's output?

Assuming I get an answer to the above question, I am describing my idea to multiply a scalar with a tensor as follows. First, replicate this scalar element to create a vector (i.e., a tensor) of size that matches with the DenseLayerA and DenseLayerB (in this case (,10)) and perform the multiplication. Now, I am not sure if this is the right approach since I am unsuccessful in retrieving the individual elements of the Sofmax Layer's output. So, is my approach correct? if not, what is the right way to solve my problem.

Answer 1

This line of code will select the highest value in your softmax function. If you'd like more than one value just change the 1 to whichever value you'd like. I hope this helps.

top_value = tf.nn.in_top_k(logits, y, 1)