timm

Includes hundreds of pretrained models across diverse architectures like Vision Transformers, ConvNeXt, and EfficientNet, all with reproducible ImageNet training results as documented in the Models section.

Provides reference scripts for training, validation, and inference that support multiple GPU modes and advanced augmentations like MixUp and RandAugment, enabling efficient experimentation.

Supports multi-scale feature map extraction via `features_only` and `forward_features` methods, making it easy to adapt models for downstream tasks like detection or segmentation.

Offers a wide range of optimizers, schedulers, and attention modules through common APIs, such as the Muon optimizer and SplitBachNorm regularization, as listed in the Features section.