PyTorch · Function

torch.flatten

Flattens a tensor by combining a selected inclusive range of dimensions into one. By default, all dimensions are flattened while element order is preserved.

How it works

torch.flatten reshapes a contiguous range of dimensions into one dimension, preserving the input element order. With the defaults, every dimension is flattened; specifying start_dim and end_dim limits the range. Depending on the input layout, the result can be the original tensor, a view, or a copied tensor.

Syntax

torch.flatten(input, start_dim=0, end_dim=-1) → Tensor

Parameters

Name	Type	Description
`input`	`Tensor`	Input tensor to reshape by flattening the selected dimensions.
`start_dim`	`int`	First dimension included in the flattened range; default is 0.
`end_dim`	`int`	Last dimension included in the flattened range; default is -1.

Returns

A Tensor with the selected dimensions combined into one dimension; the result may be the original object, a view, or a copy.

Key points

With default arguments, the function reshapes the entire input into one dimension.
Passing dimension bounds restricts flattening to that inclusive range.
Flattening preserves the original traversal order of tensor elements.
When the selected range requires no flattening, the exact input object is returned.
A view is preferred when the input layout supports the requested reshaping.
The input parameter is a Tensor whose selected dimensions are flattened.

Common mistakes

Do not assume the result always owns copied data; it may alias the input through a view.
A copy is made only when the requested flattened shape cannot be represented as a view.
A scalar becomes a one-dimensional tensor rather than remaining zero-dimensional.

Example: Flatten tensor dimensions

import torch

x = torch.tensor([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])
print(torch.flatten(x).shape)          # torch.Size([8])
print(torch.flatten(x, start_dim=1).shape)  # torch.Size([2, 4])

Flattens all dimensions, or only the selected dimension range while preserving element order.

When to use it

Connect convolutional output to linear layers

Prepare a multidimensional activation tensor for a fully connected layer while retaining the batch dimension.

Create a one-dimensional representation

Convert an entire tensor into a single vector for serialization, inspection, or downstream processing.

Flatten a selected dimension range

Collapse only feature or spatial dimensions while preserving surrounding dimensions.

Interview questions

QWhat does torch.flatten do by default?

The default range is start_dim=0 through end_dim=-1, so all dimensions are combined into one dimension. The output is one-dimensional and retains the input element order.

Follow-up: What happens for a non-contiguous tensor? If the requested shape cannot be represented as a view, torch.flatten copies the data.

QHow do start_dim and end_dim affect the output shape?

It combines dimensions from start_dim through end_dim into a single dimension, leaving dimensions outside that range unchanged. For example, flattening a shape (2, 3, 4) from dimension 1 produces (2, 12).

Follow-up: What does end_dim=-1 mean? It selects the last dimension, using Python-style negative indexing.

QDoes torch.flatten always copy the input?

The result can be the original object when no dimensions are flattened, a view when the flattened shape is view-compatible, or a copy otherwise. Thus, code should not depend on a copy being created.

Follow-up: How can view compatibility be investigated? The documentation points to torch.Tensor.view() for the conditions governing when a view is returned.

torch.Tensor.view
torch.Tensor.flatten
torch.unflatten

Source: torch.flatten — PyTorch 2.14 documentation. Checked 2026-10-16.