Index sequences
Generate integer indices for tensor operations or iteration.
PyTorch · Function
torch.arange
Creates a 1-D tensor containing values from start up to, but excluding, end, separated by step. start defaults to 0 and step to 1; dtype is inferred unless specified.
How it works
torch.arange returns a one-dimensional tensor whose values begin at start and advance by step while remaining below end. The result size is determined by the interval and step, and omitted start or step values default to 0 and 1 respectively.
Syntax
torch.arange(start=0, end, step=1, *, out=None, dtype=None, layout=torch.strided, device=None, requires_grad=False) → TensorParameters
| Name | Type | Description |
|---|---|---|
start | Number, optional | Starting value of the sequence; defaults to 0. |
end | Number | Exclusive upper endpoint of the sequence. |
step | Number, optional | Difference between adjacent values; defaults to 1. |
out | Tensor, optional | Optional destination tensor. |
dtype | torch.dtype, optional | Desired dtype; if omitted, PyTorch infers it from the inputs. |
layout | torch.layout, optional | Desired tensor layout; defaults to torch.strided. |
device | torch.device, optional | Device on which to create the tensor. |
requires_grad | bool, optional | Whether autograd records operations on the returned tensor; defaults to false. |
Returns
A one-dimensional Tensor containing the generated range.
Key points
- The result is always one-dimensional, with its length determined by the interval and step.
- The end boundary is exclusive, so a value equal to end is not included.
- Adjacent values differ by step, starting with start.
- When the relevant inputs are not floating-point and dtype is omitted, the result uses int64.
Common mistakes
- Floating-point bounds or steps cause dtype inference to use the current default floating-point dtype; specify dtype when exact type control matters.
Example: Create an evenly spaced range
import torch
steps = torch.arange(0, 10, 2, dtype=torch.float32)
print(steps) # tensor([0., 2., 4., 6., 8.])Creates floating-point values starting at 0, advancing by 2, and stopping before 10.
When to use it
Stepped numeric ranges
Build regularly spaced numeric coordinates with a chosen increment.
Device- and dtype-specific initialization
Create a range directly on a selected device and with a chosen dtype.
Interview questions
QWhat does torch.arange generate?
It creates a one-dimensional tensor containing values in the half-open interval [start, end), separated by step. The number of elements is the ceiling of (end - start) divided by step.
Follow-up: What happens when start is omitted? start defaults to 0, so torch.arange(end) generates values from 0 up to, but excluding, end.
QWhich arguments are optional?
The required argument is end; start defaults to 0 and step defaults to 1. Thus torch.arange(5) produces integer values 0 through 4.
Follow-up: How can the output type be controlled? Pass dtype explicitly, such as dtype=torch.float32.
QHow is the result dtype selected?
If dtype is omitted, PyTorch infers it from the inputs: floating-point inputs use the default floating-point dtype, while otherwise the result defaults to torch.int64. The function also accepts device, layout, out, and requires_grad keyword arguments.
Follow-up: Why should dtype often be explicit for fractional ranges? Floating-point inference depends on the configured default dtype, so explicit dtype makes the result predictable.
Related
torch.linspacetorch.zerostorch.onestorch.tensor