PyTorch Fundamentals

PyTorch is a framework for numerical computing and machine learning. Its central data type is the tensor. A tensor holds numbers in one or more dimensions and can run on supported accelerators as well as CPUs.

The framework makes a training system concrete. You represent data as tensors. You express a model as a composition of operations. You define a loss that measures error. PyTorch records the operations needed for gradients, and an optimizer uses those gradients to update learnable parameters.

What PyTorch gives you

PyTorch has a small core and a broad ecosystem. The core provides tensor operations, automatic differentiation through torch.autograd, neural-network building blocks in torch.nn, and optimization algorithms in torch.optim.

You write ordinary Python around that core. This matters when you need to inspect intermediate values, branch on data, or debug a training step. The cost is that you still own the training design. PyTorch does not choose the dataset, objective, model architecture, evaluation metric, or deployment boundary for you.