Timm ❤️ Transformers: Use any timm model with transformers

Get lightning-fast inference, quick quantization, torch.compile boosts, and effortless fine-tuning for any timm model—all within the friendly 🤗 transformers ecosystem.

Enter TimmWrapper—a simple, yet powerful tool that unlocks this potential.

In this post, we’ll cover:

• How the timm integration works and why it’s a game-changer.
• How to integrate timm models with 🤗 transformers.
• Practical examples: pipelines, quantization, fine-tuning, and more.

To follow along with this blog post, install the latest version of transformers and timm by running:

pip install -Uq transformers timm

Check out the full repository for all code examples and notebooks: 🔗 TimmWrapper Examples

What is timm?

The PyTorch Image Models (timm) library offers a rich collection of state-of-the-art computer vision models, along with useful layers, utilities, optimizers, and data augmentations. With more than 32K GitHub stars and more than 200K daily downloads at the time of writing, it's a go-to resource for image classification and feature extraction for object detection, segmentation, image search, and other downstream tasks.

With pre-trained models covering a wide range of architectures, timm simplifies the workflow for computer vision practitioners.

Why Use the timm integration?

While 🤗 transformers supports several vision models, timm offers an even broader collection, including many mobile-friendly and efficient models not available in transformers.

The timm integration bridges this gap, bringing the best of both worlds:

Pipeline API: Using timm Models for Image Classification

One of the standout features of the timm integration is that it allows you to leverage the 🤗 pipeline API. The pipeline API abstracts away a lot of complexity, making it easy to load a pre-trained model, perform inference, and view results with a few lines of code.

Let's see how to use a transformers pipeline with the MobileNetV4. This architecture does not have a native transformers implementation, but can be easily used from timm: