Many-to-one — take the last timestep

Background

For many-to-one prediction, the RNN reads the whole window and you use only the last timestep's hidden vector — out[:, -1, :] in PyTorch — as the summary to feed the output head:

$$\text{out}:(B, T, H) \;\longrightarrow\; \text{out}[:, -1, :] : (B, H)$$

Problem statement

Implement last_timestep(out) returning the last timestep of each sequence in the batch.

Input

• out — array-like of shape (B, T, H).

Output

Returns an np.ndarray of shape (B, H): the hidden vector at the final timestep for each batch item.

Examples

Example 1

Input:  out = [[[1], [2], [3]]]     # B=1, T=3, H=1
Output: [[3]]

Explanation: the last (T=3) timestep.

Example 2 — batch of 2

Input:  out = [[[1, 1], [2, 2]], [[3, 3], [4, 4]]]   # B=2, T=2, H=2
Output: [[2, 2], [4, 4]]

Constraints

• Index the time axis at -1, keeping the batch and hidden axes.
• Return shape (B, H).

Notes

• For many-to-many tasks you'd keep all timesteps instead; the choice of which outputs to read defines the task shape.