Elementwise masking
Replace invalid, masked, or disallowed tensor elements with a fallback value without writing an explicit Python loop.
PyTorch · Function
torch.where
torch.where selects elements from input or other according to a condition, supporting broadcasting across the operands. With only a condition, it returns the indices of true elements.
How it works
torch.where performs elementwise selection: a true condition chooses input, while a false condition chooses other. The three operands are broadcast together, and the result has that broadcasted shape. With only condition, it instead returns the indices of true elements as a tuple.
Syntax
torch.where(condition, input, other, *, out=None) → Tensor
torch.where(condition) → tuple of LongTensorParameters
| Name | Type | Description |
|---|---|---|
condition | BoolTensor | Boolean tensor; true or nonzero entries choose input, while false entries choose other. |
input | Tensor or Scalar | Tensor or scalar chosen where condition is true. |
other | Tensor or Scalar | Tensor or scalar chosen where condition is false. |
out | Tensor, optional | Optional destination tensor for the operation. |
Returns
For the three-argument form, a Tensor whose shape equals the broadcasted shape of condition, input, and other. The one-argument form returns a tuple of LongTensors.
Key points
- Use the three-argument form for elementwise conditional selection between two values.
- A true or nonzero condition selects input; every other condition selects other.
- input may be one scalar or a tensor supplying values at true-condition positions.
- other may be one scalar or a tensor supplying values at false-condition positions.
- The result's dimensions are the broadcasted dimensions of all three operands.
- The one-argument overload provides true-element coordinates as a tuple of LongTensors.
Common mistakes
- condition, input, and other must have shapes that can be broadcast together; incompatible shapes raise an error.
- The one-argument form returns index tensors rather than selected values, matching nonzero with tuple output.
Example: Select values conditionally
import torch
condition = torch.tensor([True, False, True])
values = torch.tensor([1.0, 2.0, 3.0])
result = torch.where(condition, values, torch.tensor(0.0))
print(result) # tensor([1., 0., 3.])Selects each element from values when the corresponding condition is true; otherwise it selects zero.
When to use it
Conditional tensor transformation
Choose between two tensors or between a tensor and a scalar while preserving a broadcasted tensor layout.
Find true-element indices
Obtain coordinates of entries satisfying a Boolean mask using the one-argument overload.
Interview questions
QWhat does torch.where do, and what shape does it return?
torch.where(condition, input, other) returns an elementwise tensor. At each position it takes input when condition is true and other otherwise; the operands must be broadcastable.
Follow-up: How can torch.where return indices? Use the one-argument form torch.where(condition), which is equivalent to torch.nonzero(condition, as_tuple=True).
QWhat types can the operands have?
condition is a BoolTensor: true or nonzero entries select input, and false entries select other. input and other may each be tensors or scalars.
Follow-up: What determines the output shape? Broadcasting condition, input, and other determines the common output shape.
QWhat are the optional output argument and return value?
The optional out keyword supplies the output tensor. The returned tensor has the shape obtained by broadcasting condition, input, and other.
Follow-up: What alternative tensor method is documented? self.where(condition, y) is equivalent to torch.where(condition, self, y).
Related
torch.nonzerotorch.Tensor.where