OtherFreeOpen Source

TIMM

700+ pretrained PyTorch vision models

Apache-2.0

ABOUT

Computer vision practitioners need access to hundreds of pretrained model architectures for transfer learning, benchmarking, and production deployment — but each model family (ResNet, ViT, ConvNeXt, EfficientNet) has its own codebase, loading conventions, and preprocessing requirements, making it tedious to compare architectures or swap backbones. timm provides a unified Model.create() and create_model() interface to 700+ pretrained models with consistent weight loading, preprocessing pipelines, and training recipes — enabling researchers and engineers to rapidly benchmark architectures, extract features for downstream tasks, and deploy optimized Vision Transformers or ConvNets without managing separate model repositories.

INSTALL

pip install timm

INTEGRATION GUIDE

1. Fine-tune a Vision Transformer (ViT, DeiT, Swin) or ConvNet (ConvNeXt, ResNet, EfficientNet) for custom image classification using timm's built-in training scripts and data augmentations 2. Extract feature embeddings from pretrained vision models for image retrieval, similarity search, and zero-shot classification pipelines 3. Benchmark and compare dozens of model architectures on a target dataset using a consistent API, learning rate schedule, and augmentation pipeline

TIMM

ABOUT

INTEGRATION GUIDE

TAGS