Tensor by Tensor

02.02 · UNIT 01 · PyTorch in One Hour: foundations · Lesson

Tensors, shapes, and data types

A tensor is a grid of numbers. Rank counts the axes, and shape gives each axis length.

PLAIN-LANGUAGE INTRODUCTION

The idea

A tensor is a grid of numbers. Rank counts the axes, and shape gives each axis length.

A worked example

torch.tensor([[1,2,3],[4,5,6]]) has shape (2,3). Its transpose has shape (3,2).

The input

Python numbers or nested Python lists.

The result

A tensor with a rank, a shape, and a data type.

Why you use it

Shape and data type decide which operations are valid.

A limit to remember

A scalar of shape () is not a one-item vector of shape (1,).

Check your understanding

  1. Rank 0 is a scalar, rank 1 is a vector, and rank 2 is a matrix. Higher ranks have no special name.
  2. torch.tensor reads a Python number or nested list. Each extra list level adds one axis.
  3. With the usual defaults, Python integers become torch.int64. Python floats become torch.float32. Set dtype explicitly when a particular type matters.
  4. Reshape and transpose keep the number of values. Matrix multiply needs matching inner sizes: (2,3) times (3,2) gives (2,2).
  5. Common mistake: a reshape changes the view, not the meaning of each axis. Track what every axis represents.
Open the detailed notes ↗