PyTorch Lab
Working through the framework, layer by layer
Read the sourceMostly Jupyter Notebook
Overview
Tensors, autograd, custom nn.Modules and training loops written out rather than abstracted away — where the from-scratch maths meets the framework that does it faster.