PyTorch

PyTorch’s dynamic computational graph — building the graph as your code runs, rather than defining it upfront — is what made it the dominant framework in ML research: you can debug a model with standard Python tools and change architecture logic on the fly, instead of working around a static graph definition. That research-first design has increasingly carried over into production use as well.

Sits above NumPy conceptually (tensors instead of arrays, with GPU support and automatic differentiation built in) — usually reached for once a project moves from data analysis into actually training neural networks.