huggingface/pytorch-image-models

The largest collection of PyTorch image encoders / backbones. Including train, eval, inference, export scripts, and pretrained weights -- ResNet, ResNeXT, EfficientNet, NFNet, Vision Transformer (ViT), MobileNetV4, MobileNet-V3 & V2, RegNet, DPN, CSPNet, Swin Transformer, MaxViT, CoAtNet, ConvNeXt, and more

What it solves

PyTorch Image Models (timm) provides a centralized, standardized library of state-of-the-art (SOTA) image models and the necessary infrastructure to use them. It eliminates the need for developers to manually implement or hunt for disparate model architectures and pretrained weights across different repositories, enabling reproducible ImageNet training results.

How it works

timm) acts as a comprehensive model zoo and toolkit. It integrates a vast array of vision architectures (such as ViT, ConvNeXt, EfficientNet, and MobileNet) and provides unified interfaces for model creation, pretrained weight loading, and inference. The library also includes essential utilities for training and validation, including specialized optimizers (like Muon, Kron, and AdamW), data loaders, and augmentation strategies.

Who it’s for

It is designed for computer vision researchers and developers who need access to a wide variety of high-performance image encoders and backbones for tasks like image classification, feature extraction, and fine-tuning.

Highlights

  • Extensive Model Collection: Supports a massive range of SOTA models including Vision Transformers (ViT), ConvNeXt, and various efficient mobile backbones.
  • NaFlexViT Support: Implements a flexible Vision Transformer capable of handling variable aspect ratios, resolutions, and patch sizes.
  • Advanced Optimizers: Includes specialized optimization algorithms such as Muon, AdaMuon, and Kron to improve training efficiency.
  • Ready-to-use Infrastructure: Provides reference scripts for training, validation, and inference, along with support for bfloat16 and float16 precision.
  • Broad Integration: Seamlessly integrates with the Hugging Face Hub for weight distribution and supports various PyTorch versions.

Related

  • Project
    lightly-ai/lightly-train

    A computer vision framework for pretraining, fine-tuning, and distilling SOTA models like DINOv3 and LTDETRv2 for tasks such as object detection and segmentation.

  • Project
    jamjamjon/usls

    A cross-platform Rust library powered by ONNX Runtime for efficient inference of vision and vision-language models under 1B parameters.

  • Project
    timeseriesAI/tsai

    A state-of-the-art deep learning library built on PyTorch and fastai for time series tasks including classification, regression, and forecasting.

  • Project
    pytorch/vision

    A PyTorch library providing popular datasets, model architectures, and image transformations for computer vision.

  • Project
    facebookresearch/multimodal

    A PyTorch library for training state-of-the-art multimodal models at scale, providing modular building blocks and pretrained weights for models like CLIP and BLIP-2.