causal-mask — Razorpay case

Razorpay-inspired transaction-sequence model must not allow a prediction at position t to inspect events that occur later in the sequence. You need to construct the causal attention mask that blocks those future positions.

Build the lower-triangular boolean mask of size $n\times nn\backslash times\; n$: entry $(i,j)(i,j)$ is True when position $ii$ may look at position $jj$, i.e. $j\le ij\backslash le\; i$.

Implement solve(n).

Returns. Return an $n\times nn\backslash times\; n$ boolean NumPy array.

Examples

Example 1

Input:

solve(3)

Output:

[[True, False, False], [True, True, False], [True, True, True]]

Example 2

Input:

solve(1)

Output:

[[True]]