qubvel-org/segmentation_models.pytorch ? reverse-engineered prompt

I need a clean PyTorch library for semantic segmentation on images. Make it dead simple so anyone can create a model in two lines, like picking a Unet or Segformer architecture with a pretrained resnet34 or efficientnet backbone from imagenet weights.

Support tons of encoder options, both convolutional and transformer based, plus different input channels for rgb or grayscale and any number of output classes.

Include popular losses and metrics like dice or jaccard to make training straightforward. Keep everything compatible with ONNX export and common training loops.

Add a few example scripts showing how to train on datasets like pet photos or street scenes from camvid. Focus on a high level api that just works without deep pytorch expertise.