LogoAIAny
Icon for item

PyTorch Image Models (timm)

PyTorch Image Models (timm) is an open-source PyTorch library that aggregates a large collection of image model architectures, pretrained weights, training/validation/inference scripts, optimizers, schedulers and augmentations. It is widely used for research, reproducible ImageNet-style training, and as a convenient model/backbone provider for vision projects.

Visit Website
Visit Website

Introduction

PyTorch Image Models (timm)

PyTorch Image Models (timm) is a comprehensive PyTorch library for computer vision. Initially authored by Ross Wightman and now maintained under the Hugging Face organization, timm consolidates a very wide range of image model architectures, utilities and reference training code, making it a one-stop toolkit for researchers and engineers working on image classification, feature extraction and backbone integration.

Key components
  • Model zoo: implementations of many families (ResNet / ResNeXt, EfficientNet, MobileNet variants, ConvNeXt, NFNet, ViT and many transformer-based and hybrid architectures). Most models include convenient factory functions and support for pretrained weights.
  • Pretrained weights: many models ship with ImageNet (and other) pretrained checkpoints, with loaders that adapt input channels or final layers when needed.
  • Training & evaluation scripts: reference scripts for distributed training, validation and inference that support mixed precision, DDP, and many modern training conveniences.
  • Optimizers & schedulers: a rich set of optimizer implementations and a factory to create them (AdamW, Adan, Lion, Muon, Kron, and many more), plus LR schedulers often used in vision training.
  • Data pipeline & augmentations: common augmentations (Mixup, CutMix, RandAugment, AutoAugment, Random Erasing, AugMix) and data loaders geared for ImageNet-style experiments.
  • Utilities: features-only mode, multi-scale feature extraction, dynamic pooling, export helpers (ONNX), checkpointing helpers, and various normalization/attention modules.
Why use timm
  • Breadth: timm aggregates a very large set of modern and legacy vision architectures in a single consistent API.
  • Reproducibility: reference training recipes and results tables help reproduce ImageNet-like experiments.
  • Flexibility: consistent model creation API (create_model) with options for feature extraction, output indices and pretrained weight handling.
  • Engineering-ready: includes production-friendly features like export, checkpointing fixes, and compatibility with Hugging Face Hub for pretrained weight hosting.
Example uses
  • Rapidly try different backbones for a classification or detection pipeline by switching model names in create_model.
  • Use pretrained encoders as feature extractors for transfer learning or as backbones in detection/segmentation.
  • Run large-scale training experiments using the included distributed training scripts and optimizer implementations.
Licensing & provenance
  • The code in the repository is Apache 2.0 licensed. Many pretrained weights originate from ImageNet and other sources; users should check status and dataset licenses before commercial use.
Maintenance & community
  • Active project with frequent releases, contributions from the community, and integration with the Hugging Face Hub (docs and hosted weights). The repository is commonly used in both research and applied ML workflows.
Back

Information

  • Websitegithub.com
  • AuthorsRoss Wightman, Hugging Face
  • Published date2019/02/02

Categories

Tags

More Items

Icon for item

Isaac Lab

2022
NVIDIA (Isaac Sim / Omniverse team)

Isaac Lab is an open-source, GPU-accelerated robotics learning framework built on NVIDIA Isaac Sim. It provides high-fidelity physics and sensor simulation, ready-to-train environments and robot models, and integrations for reinforcement and imitation learning workflows to accelerate sim-to-real research and large-scale robot training.

nvidiaRLphysicsai-frameworkai-train+3
Icon for item

Kornia

2018
Kornia contributors, E. Riba +4

Kornia is a differentiable computer-vision library built on top of PyTorch. It provides a comprehensive collection of differentiable image processing operators, geometric vision modules, advanced augmentation pipelines, and integrations with pre-trained vision models — all designed to run in batch on GPU and integrate with deep learning workflows.

visionpytorchai-librarygithubai-image+2
Icon for item

Dolphin: Document Image Parsing via Heterogeneous Anchor Prompting

2025
ByteDance

Dolphin is an open-source document image parsing project from ByteDance that uses heterogeneous anchor prompting and a document-type-aware two-stage architecture. It handles both digital-born and photographed documents, offering page-level and element-level parsing (text, tables, formulas, code). Dolphin-v2 (3B) improves accuracy and adds multi-page PDF support, deployment recipes (vLLM, TensorRT-LLM), and Hugging Face model hosting. The repository includes code, demos, pretrained models, and a BibTeX citation; license: MIT.

githubbytedanceocrpapervision+3