conformer

What it is and what it does

Conformer is a PyTorch library implementing the Conformer architecture from the 2020 paper, which augments transformer models with depthwise convolutions to improve local feature extraction. The package provides three main components: ConformerConvModule (the core convolutional layer), ConformerBlock (a complete transformer block with attention, feed-forward, and convolution), and Conformer (a full stack of multiple blocks). It is primarily designed for speech recognition and audio tasks where capturing local patterns alongside global attention is beneficial.

The library depends only on torch and einops, making installation straightforward. However, the project is no longer maintained—the last commit was in May 2023 and the repository shows no active development. This means bug fixes, compatibility updates with newer PyTorch versions, or feature improvements are unlikely. For production use or long-term projects, you should evaluate whether the frozen implementation meets your needs or if you need an actively maintained alternative.

Use it for:

• Building speech recognition models that combine transformer attention with local convolution for improved acoustic modeling.
• Prototyping audio processing pipelines where you need the Conformer architecture as a backbone component.
• Research or educational projects exploring the Conformer paper's approach to augmenting transformers with convolution.
• Adapting existing Conformer implementations into larger custom architectures by using ConformerBlock or ConformerConvModule as building blocks.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Provides PyTorch implementations of the Conformer architecture—a convolutional module and full transformer blocks that add local inductive bias to transformer models, commonly used in speech recognition.

Yes, if you need the Conformer architecture for a specific project and can accept a frozen, unmaintained codebase. The low install friction and permissive license make it easy to adopt, and the implementation is straightforward enough to fork or patch if needed. However, if you require ongoing maintenance, compatibility updates, or active support, look for an actively maintained alternative or be prepared to maintain a fork yourself.

Install

conformer on PyPI

pip install conformer

uv add conformer

poetry add conformer

Installing conformer

Before you install

Low friction: pure Python wheel with only torch and einops as runtime dependencies. However, the package is abandoned—last release was 2023-05-17 and no commits since then, so expect no maintenance or bug fixes.

License in practice

MIT license is permissive; you can use, modify, and distribute this code freely with minimal restrictions, making it suitable for both open and closed projects.

Quickstart

pip install conformer

import torch
from conformer import ConformerBlock

block = ConformerBlock(dim=512, dim_head=64, heads=8, ff_mult=4, conv_expansion_factor=2, conv_kernel_size=31)
x = torch.randn(1, 1024, 512)
output = block(x)

Requires PyTorch and einops to be installed; no other system-level dependencies.

Verify before relying

• Whether the implementation matches the exact Conformer paper specification or if there are known deviations.
• Current compatibility with recent PyTorch versions, given the last release was in 2023.
• Whether the TODO items (relative positional encoding, flash attention) have been addressed elsewhere or remain unimplemented.

Package facts

Evidence: conformer-0.3.2-py3-none-any.whl

Keywords: artificial intelligence, deep learning, transformers, audio

Tags

More Artificial Intelligence packages