Tangent
Source-to-source debuggable derivatives in pure Python
Existing libraries implement automatic differentiation by tracing a program's execution (at runtime, like PyTorch) or by staging out a dynamic data-flow graph and then differentiating the graph (ahead-of-time, like TensorFlow). In contrast, Tangent performs ahead-of-time autodiff on the Python source code itself, and produces Python source code as its output. Tangent fills a unique location in the space of machine learning tools. As a result, you can finally read your automatic derivative code...