dfdx

Why we included this project

dfdx is one of the few deep learning libraries that actually exploits Rust's type system rather than just mimicking PyTorch. Its main draw is compile-time shape checking: a mismatched tensor dimension in your model won't even compile, so you catch those mistakes before you run anything. Beyond that, it has a solid tensor library, an autodiff engine, common neural network layers (linear, conv, transformer), and optimizers such as SGD, Adam, and AdamW. You can also enable CUDA for GPU acceleration. This is a good fit for Rust developers who want to build models without leaving the language, or who need a compiled alternative for porting research code. Just note that the project is still pre-alpha and plans breaking releases, so treat it as a tool for learning and prototyping rather than production.