# Creating tensors: shape, dtype and device — PyTorch

Source: https://www.skillbyai.com/en/pytorch/pt-tensors

> Create tensors from data and factory functions, read their shape, dtype and device, and avoid float64 and int-label mix-ups.

## Every tensor has a shape, a dtype and a device

A **tensor** is an n-dimensional array whose elements all share one data type. Three attributes describe it: `shape` (size along each dimension), `dtype` (for example `torch.float32` or `torch.int64`) and `device` (`cpu`, `cuda`, `mps`).

`torch.tensor(data)` **copies** Python data and infers the dtype: Python floats become `float32`, Python ints become `int64`. Factory functions build tensors of a given shape: `torch.zeros`, `torch.ones`, `torch.rand` (uniform in [0, 1)), `torch.randn` (standard normal), `torch.arange` and `torch.linspace`.

`torch.from_numpy(arr)` does not copy: the tensor **shares memory** with the NumPy array, so changing one changes the other. Note that NumPy's default float type is `float64`, while PyTorch models use `float32` because it is faster on GPUs and takes half the memory. Mixing them causes errors like *expected scalar type Float but found Double*. Classification labels, on the other hand, are usually `int64` (also called `long`).

## From raw numbers to batched tensors

Data starts as Python lists or NumPy arrays, becomes a tensor with a fixed dtype, gets reshaped or indexed into the layout a model expects, and is finally combined with other tensors through broadcasting and matrix operations.

![Diagram of 4 connected stages from left to right: Python or NumPy data, tensor with dtype, views and reshapes, broadcast operations.](assets/figures/pytorch/section-2-map.svg) — Figure 2.1: Python or NumPy data -> tensor with dtype -> views and reshapes -> broadcast operations

## Creating tensors

The comments state what each line produces. Watch the last line: the NumPy-backed tensor is `float64` and changes when the array changes.

```python
import numpy as np
import torch

a = torch.tensor([1, 2, 3])          # dtype torch.int64
b = torch.tensor([1.0, 2.0, 3.0])    # dtype torch.float32
z = torch.zeros(2, 3)                # shape (2, 3), all zeros
r = torch.randn(4, 5)                # samples from a standard normal
s = torch.arange(0, 10, 2)           # tensor([0, 2, 4, 6, 8])

arr = np.array([1.0, 2.0])
t = torch.from_numpy(arr)            # float64, shares memory with arr
arr[0] = 99.0
print(t[0].item())                   # 99.0: same memory
print(a.dtype, b.dtype, t.dtype, z.shape)
```

## Converting dtypes on purpose

Use `.float()`, `.long()` or `.to(torch.float32)` to convert. A common pattern after loading NumPy data is `torch.from_numpy(arr).float()`. Each conversion returns a new tensor; it does not change the original in place.

**Quiz:** What is the dtype of `torch.tensor([1, 2, 3])`?

- [ ] torch.float32
- [x] torch.int64
- [ ] torch.int32
- [ ] torch.float64

*Answer:* torch.int64. Python integers are inferred as `torch.int64`. Python floats would give `torch.float32`.