Global reduction
Compute a single aggregate statistic from a tensor.
PyTorch · Function
torch.mean
Computes the mean of all tensor elements or reduces selected dimensions. It supports optional dtype conversion and dimension retention.
How it works
torch.mean computes the arithmetic mean of every element or reduces selected dimensions. The optional dtype casts the input before reduction, while keepdim preserves reduced dimensions as size 1.
Syntax
torch.mean(*input*, ***, *dtype=None*) → Tensor
torch.mean(*input*, *dim*, *keepdim=False*, ***, *dtype=None*, *out=None*) → TensorParameters
| Name | Type | Description |
|---|---|---|
input | Tensor | The input tensor, with floating-point or complex dtype. |
dtype | torch.dtype, optional | Optional desired data type; the input is cast to it before reduction. |
dim | int or tuple of ints, optional | Dimension or dimensions over which to reduce. |
keepdim | bool, optional | Whether to retain reduced dimensions as size 1. |
out | Tensor, optional | Optional output tensor. |
Returns
A Tensor containing the mean of all input elements or the means after reduction over the selected dimension(s).
Key points
- Without dim, the function reduces the complete tensor to its overall mean.
- Supplying dim computes means along the selected axis or axes.
- A dimension list enables one reduction over multiple axes.
- keepdim=True preserves reduced axes as singleton dimensions.
- With keepdim=False, reduced axes are removed from the result.
- dtype controls a pre-reduction cast and the resulting tensor's data type.
Common mistakes
- An empty input produces nan because the mean of no elements is undefined.
- The input tensor must use a floating-point or complex dtype.
Example: Overall and dimension-wise means
import torch
x = torch.tensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
print(torch.mean(x))
print(torch.mean(x, dim=0, keepdim=True))Computes the overall mean, then the column means while retaining the reduced dimension.
When to use it
Axis-wise reduction
Average features, batches, or spatial axes independently.
Shape-preserving reductions
Maintain broadcast-compatible shapes after reducing dimensions.
Interview questions
QWhat does torch.mean return when no dimension is supplied?
With only input, torch.mean returns one scalar containing the mean of all elements. With dim, it reduces the specified dimension or dimensions instead.
Follow-up: What happens when dim is a list? It reduces across every dimension in that list.
QHow does keepdim affect the output shape?
keepdim=True leaves each reduced axis in the result with length one. With the default False, those axes are squeezed from the output shape.
Follow-up: Why use keepdim=True? It keeps shapes suitable for broadcasting against the unreduced tensor.
QWhat is the purpose of dtype?
dtype specifies the desired output data type. When provided, the input is cast to that dtype before the mean is calculated.
Follow-up: What input types does the documented function accept? Floating-point and complex tensors.
Related
torch.nanmeantorch.squeezetorch.Tensor.mean