attention-weighted-sum — Swiggy case

Swiggy-inspired recommendation model combines several encoded signals using learned attention weights. You need to turn the raw attention scores into weights with a softmax and compute the weighted sum of the value vectors, so the model produces the intended context representation.

scores are the raw (unnormalised) attention scores for one query over $nn$ items and V is the $n\times dn\backslash times\; d$ matrix of their value vectors. Turn the scores into weights with a stable softmax and return the weighted sum of the rows of V.

Implement solve(scores,V).

Returns. Return a float NumPy vector of length $dd$.

Examples

Example 1

Input:

solve([0.0, 0.0], [[1.0, 2.0], [3.0, 4.0]])

Output:

[2.0, 3.0]

Example 2

Input:

solve([10.0, 0.0], [[1.0, 0.0], [0.0, 1.0]])

Output:

[0.999955, 4.5e-05]