jittor

Why we included this project

Jittor is worth a look for anyone who wants a deep learning framework built around just-in-time compilation rather than the usual eager PyTorch-style loop. The framework and its meta-operators compile on the fly, so the op compiler and tuner generate code specialized to your model instead of handing you generic kernels. The front end is plain Python with dynamic graph execution and a Module API, so it feels familiar if you are coming from PyTorch, while the backend runs on CUDA and C++ for performance. Beyond training, the project ships model libraries for detection, segmentation, generation, differentiable rendering, and reinforcement learning, which makes it handy as a reference for reproducing published methods as well as a tool for your own research.