pytorch-image-models

PyTorch Image Models (timm) is a collection of image models, layers, utilities, optimizers, schedulers, data loaders and augmentations, plus reference training, validation and inference scripts. Its stated goal is to gather a wide variety of image classification and backbone architectures with the ability to reproduce ImageNet training results. The project aggregates many architecture families, including ResNet, ResNeXT, EfficientNet, NFNet, Vision Transformer (ViT), MobileNetV4, MobileNet-V3/V2, RegNet, DPN, CSPNet, Swin Transformer, MaxViT, CoAtNet and ConvNeXt. It also covers newer additions such as EVA, DINOv3, SigLIP-2, MobileCLIP-2, MetaCLIP-2, Perception Encoder, NaFlexViT, and various efficient mobile-oriented models. Model definitions are paired with pretrained weights hosted on the Hugging Face Hub, and the repository notes that some variants intentionally ship without weights. Beyond model definitions, timm provides training and validation scripts, inference and export utilities, and support for variable-resolution and variable-aspect training through the NaFlex pipeline. It includes optimizers such as AdamW variants, Muon, AdaMuon, NAdaMuon, Kron and others, along with schedulers and augmentation utilities. Recent releases emphasize meta-device initialization, device/dtype factory support, gradient checkpointing, ONNX export improvements, and safer checkpoint loading defaults. The README documents frequent releases, benchmark result files for inference timing on various GPUs, and a long list of contributed models with links to papers and source repositories. It also acknowledges contributions from many external authors and projects. This is a library and toolchain for developers and researchers working with PyTorch vision models, rather than an end-user application. It is useful for training, fine-tuning, evaluating, exporting and benchmarking image encoders and backbones.