huggingface

pytorch-image-models: PyTorch Vision Models and Tools

PyTorch Image Models (timm) is a PyTorch collection of image encoders and backbones with pretrained weights, plus tools for training, evaluation, inference, and export. It suits teams building or comparing computer-vision models without assembling these components from scratch.

Stars
37.2k
Forks
5.2k
Stars, last 30 days
+63
Last commit
4 days ago
Contributors
100+

Our take

Activeregular commits and releases

Health score

73/100

How it is scored
Activity
75
Community
61
Issues
82
Pull requests
74

+63 stars in the last 30 days

  • Closes issues faster than 81% of the projects we track
  • More contributors than 78% of the projects we track
  • Merges more pull requests than 77% of the projects we track

timm is a broad, actively maintained PyTorch vision library, though teams should account for its heavily concentrated contributor activity and check the terms attached to individual pretrained weights.

Good fit if

  • You need a broad selection of PyTorch image encoders, backbones, or pretrained weights behind a consistent interface.
  • You want reusable training and evaluation scripts alongside model definitions, rather than models alone.
  • Your team can work in Python and review the Apache-2.0 code license as well as any separate dataset or weight conditions.

Look elsewhere if

  • You need a turnkey application or a dedicated end-to-end detection or segmentation framework. The repository is framed as a model library and reference tooling.
  • Your project does not use PyTorch or image models.
  • Your team requires project activity to be distributed among many contributors. Recent activity is strong, but most measured contributions come from one person.
All health signals
Last commit2026-10-20 (4 days ago)
Commits, last 90 days100+
Releases, last 12 months10 (latest v1.0.30, 2026-09-22)
Contributors100+ (top contributor: 85% of commits)
Issues closed, last 90 days10+ (typically closed in 0 days)
Pull requests merged, last 90 days58 (typically merged in 0 days)
Project age7 years

Checked on 2026-10-25 with the GitHub API.

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

37,289

Forks

5,230

Change since Nov 17, 2026: +2 stars·-1 forks

Overview

PyTorch Image Models, also known as timm, is a collection built on PyTorch for image encoders and backbones, including convolutional networks and Vision Transformers. It addresses the practical work around using these models too: consistent model interfaces, pretrained-weight loading, feature extraction, and reference scripts for training, validation, and inference.

It is aimed at researchers and developers who need a broad set of image models and supporting components in one library. The repository also includes augmentation and regularization tools, optimizers, schedulers, and export-related utilities. Official documentation is available at https://huggingface.co/docs/timm.

Key Features

  • A wide range of image model families, including ResNet, EfficientNet, ConvNeXt, MobileNet, Swin, and ViT variants.
  • Pretrained weights and a common model-creation and classifier interface.
  • Feature-only and multi-scale feature-map extraction for downstream vision tasks.
  • Reference training, validation, and inference scripts for image classification and other supported tasks, including multilabel classification.
  • Image augmentations and regularization methods such as Mixup, CutMix, AugMix, DropPath, and DropBlock.
  • A selection of optimizers and learning-rate schedulers, with training utilities for distributed and GPU workflows.
  • Model resizing and export support for integrating models into different inference workflows.

Use Cases

  • A computer-vision researcher can compare architectures or pretrained checkpoints using a shared model API.
  • An ML engineer can use pretrained backbones as feature extractors for a downstream vision system.
  • A team training image classifiers can adapt the included scripts and data-loading tools to its dataset and hardware.
  • A deployment engineer can evaluate model variants and use the repository's export-related tooling as part of an inference pipeline.

What you need

Detected in the repository

  • Python >=3.8 (from pyproject.toml)
  • A test suite and automated checks on GitHub Actions

License in plain words

Apache-2.0permissive

  • Commercial use: yes
  • Modify and redistribute: yes
  • You must keep: the license, the NOTICE file and a note of your changes
  • Share your changes: no
  • Includes an explicit patent grant from the contributors:

A summary, not legal advice: the LICENSE file is what applies.

Getting Started

Install the library with pip install timm, then see the README and official documentation for model creation, pretrained weights, and training scripts.

Considerations

  • The project is actively maintained, with regular commits and releases, and recent issue and pull-request handling is fast. However, measured contribution activity is concentrated heavily in one contributor, which is a continuity risk for teams relying on sustained upstream maintenance.
  • Python is required. The repository's reference training workflows include GPU and distributed options, but hardware needs depend on the chosen model and workload.
  • The code is licensed Apache-2.0, but pretrained weights may have additional conditions. The README specifically cautions that ImageNet-derived weights may carry dataset-related restrictions and flags some weights with non-commercial terms. Check the applicable weight and dataset terms before commercial use.
  • The library covers many model families, but the README describes its scripts as adaptable reference scripts, not turnkey solutions for every dataset or task.

Found this useful?

Share it with someone who would like pytorch-image-models.