qubvel-org/segmentation_models.pytorch

Provides 12 encoder-decoder architectures (Unet, Unet++, Segformer, DPT, UPerNet) with integrated task-specific losses (Dice, Jaccard, Tversky) and preprocessing functions matched to each encoder's training regime. Supports timm backbone integration for seamless access to 800+ pretrained encoders, with production-ready export to ONNX and TorchScript formats. Models are standard PyTorch modules instantiable in two lines of code with configurable input channels and auxiliary classification outputs.

11,398 stars. Used by 1 other package. Actively maintained with 84 commits in the last 30 days. Available on PyPI.