Genomics and RNA

Windowed max pooling and coordinate routing

EasyRaw PyTorchAbout 10 min1 public example plus private tests

Free account with email, Google or GitHub. Your code runs in an isolated PyTorch sandbox.

Your task

Take channelwise maxima within valid sequence windows. Return pooled values, validity, and the winning original sequence index for every output channel.

Function signature

windowed_max_pool(
    x: torch.Tensor,
    valid_mask: torch.Tensor,
    options: dict,
) -> dict[str, torch.Tensor]

Worked example

Different channels can choose different source positions

Use one batch with x=[(1,-2),(3,-5),(3,-1)] and all positions valid. Set window=2, stride=1, and padding=[1,0]. The three source-index windows are [-1,0], [0,1], and [1,2]; index -1 is padding and never competes. The pooled vectors are [(1,-2),(3,-2),(3,-1)], and winning indices are [(0,0),(1,0),(1,2)]. In the last window, channel 0 ties at value 3 and keeps earlier index 1. All output windows are valid, and the negative channel remains negative.

Parameters

Parameter	Shape and meaning
`x`torch.Tensor	$[B, L, C]$ Sequence features; only valid positions need finite values.
`valid_mask`torch.Tensor	$[B, L]$ True includes all channels at this sequence position.
`options`dict	Scalar settings. Omitted keys use the defaults below.

Return

dict[str, torch.Tensor]

Return exactly the named fields applicable to this operation.

Required behavior

For $x [B,L,C]$ and valid_mask $[B,L]$ , compute independent channel maxima over indices $j\cdot \mathrm{stride}-\mathrm{padding\_left} + k$ , $k=0$ .. $\mathrm{window}-1$ . window defaults to $4$ , stride to window, and padding to $[0,0]$ ; window/stride are positive and padding nonnegative.
$O=\operatorname{max}(0,\operatorname{floor}((L+\mathrm{left}+\mathrm{right}-\mathrm{window})/\mathrm{stride})+1)$ . Exclude out-of-range and masked positions, including nonfinite missing payloads, before taking maxima.
Return $y [B,O,C]$ , valid_mask $[B,O]$ (any eligible position), and indices $[B,O,C]$ containing the original input index of the first maximum. All-invalid windows return $y=0$ and $\mathrm{index}=-1$ ; $O=0$ returns correctly shaped empty outputs. Negative data must not compete with artificial zero padding.
A tied maximum chooses the earliest eligible index, matching the specified subgradient convention.

Why this matters

Pooling changes values and coordinates together; overlapping windows may route several output gradients back to the same input position.

Starter code

You write the body of this function in the BioTorch workspace. Your code runs in an isolated PyTorch sandbox; Run checks the public examples and Submit grades the full test suite, including gradients.

import torch


def windowed_max_pool(
    x: torch.Tensor,  # (B, L, C)
    valid_mask: torch.Tensor,  # (B, L)
    options: dict,
) -> dict[str, torch.Tensor]:
    """Return dict[str, torch.Tensor]:
    
    y: (B, O, C).
    valid_mask: (B, O).
    indices: (B, O, C)."""
    # Implement the operation. Inputs and test setup are supplied by BioTorch.
    raise NotImplementedError("Implement windowed_max_pool")

Ready to implement it?

3 progressive hints, a solution walkthrough, and grading that checks values, shapes and gradients are waiting in the workspace.

Write and run your solution Sign in

Free account with email, Google or GitHub. Your code runs in an isolated PyTorch sandbox.

Sources and further reading

A primitive exercise, not an end-to-end reproduction of the named models. No affiliation or endorsement is implied.