Catgrad is a unique deep learning framework that utilizes category theory to statically compile models for both forward and backward passes. This means you can run your training loop without requiring any deep learning framework at all! You can easily install catgrad using pip. The framework is built upon various papers and research, including those on data-parallel algorithms, differentiable polynomial circuits, and gradient-based learning. One interesting aspect is the use of category theory in reverse derivative ascent for learning Boolean circuits. Overall, catgrad offers a new approach to deep learning compilation that sets it apart in the field.
https://catgrad.com/