This repository provides a complete PyTorch implementation of FlowNet 2.0, a convolutional neural network architecture for estimating optical flow between consecutive video frames. It includes multiple network variants (FlowNet2S, FlowNet2C, FlowNet2CS, FlowNet2CSS, FlowNet2SD), custom CUDA layers for correlation and resampling operations, multi-GPU training support, and fp16 inference capabilities. The codebase provides dataloaders for standard optical flow datasets (FlyingChairs, FlyingThings, MPI-Sintel) and multi-scale loss functions for training.