timm (PyTorch Image Models)

timm — PyTorch Image Models

timm is Ross Wightman's ever-expanding collection of state-of-the-art image models, pretrained weights, and training utilities for PyTorch. With 900+ model architectures and thousands of pretrained checkpoints, it is the go-to library for transfer learning, feature extraction, and vision benchmarking.

Key Features

• 900+ model architectures: ViT, DeiT, Swin, ConvNeXt, EfficientNet, RegNet, ResNet, and more
• Thousands of pretrained weights from ImageNet, ImageNet-21k, LAION, and domain-specific datasets
• Universal create_model() API with consistent feature extraction interface
• Optimised data augmentation (RandAugment, MixUp, CutMix) and training recipes
• ONNX and TorchScript export; timm.onnx helper for direct export

Quick Start

pip install timm

import timm
import torch

model = timm.create_model("convnext_base.fb_in22k", pretrained=True)
model.eval()

data_cfg = timm.data.resolve_model_data_config(model)
transforms = timm.data.create_transform(**data_cfg, is_training=False)

from PIL import Image
img = transforms(Image.open("image.jpg")).unsqueeze(0)
with torch.no_grad():
    output = model(img)
print(output.topk(5))

npx ai-supply add timm-image-models

Curated mirror of the open-source pytorch-image-models (Apache-2.0). Get it from the source.