This repository provides PyTorch implementations of various loss functions used in deep learning training, including label smoothing, AMSoftmax, focal loss, triplet loss, dice loss, and IoU-based losses (GIOU/DIOU/CIOU). It also includes activation functions (Swish, Mish), one-hot encoding utilities with CUDA support, exponential moving average operators, and convolution variants like CoordConv and DynamicConvolution. Some operators are implemented as CUDA extensions requiring compilation.