Picodl: numpy-only deep learning framework from scratch

Picodl represents a pedagogical approach to deep learning frameworks—prioritizing transparency and readability over performance optimization. By building on numpy alone and hand-tracing every gradient into a computation graph that is topologically sorted on backward passes, the framework eliminates abstraction layers that obscure how neural networks actually function. This design choice makes it well-suited for learners and researchers who want to see the machinery beneath popular frameworks like PyTorch or TensorFlow.

The modular structure—seven focused modules (tensor, layers, loss, optimizers, NeuralNet, and data/training utilities)—allows each component to be understood in isolation while remaining composable. The inclusion of modern features like decoupled weight decay in AdamW and attention primitives suggests the framework is not merely a toy, but capable of supporting contemporary architectures. The ability to save and load weights to any file format and the support for both raw numpy arrays and Tensor inputs indicates a pragmatic design that does not sacrifice usability for simplicity.