yassouali/pytorch-segmentation

Implements multiple encoder-decoder architectures (DeepLab V3+, PSPNet, UperNet, U-Net, etc.) with atrous convolution and multi-scale parsing strategies for dense pixel-level prediction. Provides specialized loss functions including Lovász-Softmax for direct mIoU optimization and focal loss for handling class imbalance, alongside poly and one-cycle learning rate schedulers commonly used in segmentation workflows. Supports Pascal VOC, Cityscapes, ADE20K, and COCO Stuff datasets with JSON-based configuration for reproducible training pipelines.

1,814 stars. No commits in the last 6 months.