---
id: axial-positional-embedding
version: "0.3.12"
license: MIT License Copyright (c) 2020 Phil Wang Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in… (full text in the JSON record)
license_treatment: permissive
maintenance: aging
---

# axial-positional-embedding — Axial Positional Embedding

License: permissive · Maintenance: aging · Downloads: 84.7K/mo

## What it is and what it does

Axial Positional Embedding implements a specialized form of positional encoding designed for transformer networks working with multi-dimensional data. Instead of encoding position as a single sequence, it decomposes position information across multiple axes—useful for images, videos, or any structured tensor input. The package provides two main variants: a standard AxialPositionalEmbedding that combines embeddings from separate axes, and a ContinuousAxialPositionalEmbedding that uses MLPs for better extrapolation across unseen dimensions.

The package depends on torch for tensor operations and einops for flexible tensor reshaping. It targets developers building attention-based models on non-sequential data, where traditional sequence-based positional encodings are inefficient or inappropriate. The implementation is lightweight and integrates directly into PyTorch models by adding the embedding to token representations.

Use it for:

- Encoding position information in vision transformers processing image patches arranged in 2D grids
- Adding positional context to video transformers with temporal and spatial dimensions
- Improving transformer efficiency on multi-dimensional data by factorizing position encoding across axes
- Building language models that benefit from axial decomposition of sequence positions
- Extending transformers to handle variable-sized multi-dimensional inputs with better generalization

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

Provides axial positional embeddings for transformer networks, enabling position encoding across multi-dimensional data like images, videos, or sequences using decomposed attention axes.

Yes, if you are building transformer models on multi-dimensional data (images, videos, or structured tensors) and want a lightweight, permissively licensed positional encoding layer. The aging maintenance status is a minor concern for a stable utility, but verify compatibility with your specific PyTorch version. No known vulnerabilities.

## Install

pip install axial-positional-embedding
uv add axial-positional-embedding
poetry add axial-positional-embedding

## Installing axial-positional-embedding

Before you install: Low install friction with a pure Python wheel. Maintenance status is aging—last release was 535 days ago—but the package remains functional with torch and einops.

License in practice: MIT License permits free use, modification, and distribution with minimal restrictions, making it suitable for both open-source and commercial projects.

Quickstart:

pip install axial-positional-embedding

import torch
from axial_positional_embedding import AxialPositionalEmbedding

pos_emb = AxialPositionalEmbedding(dim=512, axial_shape=(64, 64), axial_dims=(256, 256))
tokens = torch.randn(1, 1024, 512)
tokens = pos_emb(tokens) + tokens

Requires torch and einops as runtime dependencies.

Verify before relying:

- Whether the package works with recent PyTorch versions beyond current support
- Performance characteristics or benchmarks compared to standard positional encoding
- Active community support or responsiveness to issues given aging maintenance status

## Package facts

- License: MIT License Copyright (c) 2020 Phil Wang Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in… (full text in the JSON record) (permissive)
- Python support: supports_current
- Install friction: low
- Maintenance: aging
- Downloads: 84.7K/month (top 15,000 on PyPI)
- Known vulnerabilities: none known

## Tags

positional embedding transformer, axial attention multi-dimensional, position encoding deep learning, transformer positional encoding, multi-dimensional attention, axial positional embedding, sequence position encoding, transformer-architecture, positional-encoding, vision-transformer

[View on SkillFed](https://skillfed.io/packages/axial-positional-embedding) · [View on PyPI](https://pypi.org/project/axial-positional-embedding/)