Beginner

Tensors Basics

Creating tensors and reading their shape, dtype, and device.

A PyTorch is NumPy's ndarray with two extra superpowers: it can live on a GPU, and it can remember the operations that built it so can flow backward through them via (Lesson 3). If you already know NumPy, most of this will feel familiar: the API is deliberately close.

⚠️ Watch out: torch.rand and torch.randn are NOT interchangeable, despite the near-identical names. rand(shape) is UNIFORM over [0, 1): every value equally likely, always non-negative. randn(shape) is standard NORMAL (Gaussian, mean 0, std 1): can be negative, most values cluster near 0. Mixing them up is a very easy, very common mistake.

Call	Makes	Example
`torch.tensor(data)`	a tensor from existing data (list, NumPy array, …)	`torch.tensor([1, 2, 3])` → `tensor([1, 2, 3])`
`torch.zeros(shape)` / `torch.ones(shape)`	filled with 0s / 1s	`torch.zeros(3)` → `tensor([0., 0., 0.])`
`torch.arange(start, stop, step)`	like NumPy's `arange`	`torch.arange(0, 6, 2)` → `tensor([0, 2, 4])`
`torch.eye(n)`	n×n identity matrix	`torch.eye(2)` → `tensor([[1., 0.], [0., 1.]])`
`torch.rand(shape)`	uniform random values in [0, 1)	`torch.rand(2)` → e.g. `tensor([0.42, 0.91])`, always ≥ 0
`torch.randn(shape)`	normal/Gaussian random values (mean 0, std 1)	`torch.randn(2)` → e.g. `tensor([-0.3, 1.2])`, can be negative
`x.float()`, `x.long()`, `x.to(torch.float32)`	dtype casts	`torch.tensor([1, 2]).float()` → `tensor([1., 2.])`

💡 Tip: torch.manual_seed(0) makes random tensors reproducible, which is essential when you need the same 'random' result on every run (like in these lessons' examples).

Now practice it

Exercises that use what this lesson just covered.

Null VectorEasy Range Vector 10–49Easy Identity MatrixEasy Seeded Random TensorEasy Mean ValueEasy

← All PyTorch lessons Shape Ops & Indexing →