Enable autograd on existing data
An existing tensor was loaded or initialized without gradient tracking and must participate in later differentiation.
PyTorch · Module
tensor.requires_grad_
Changes a tensor’s requires_grad setting in place. By default, it enables autograd recording for subsequent operations involving that tensor.
How it works
Call this method on a tensor to change its requires_grad setting in place. With the default True value, subsequent operations involving the tensor can be recorded by autograd; passing False disables recording for that tensor.
Syntax
Tensor.requires_grad_(requires_grad=True)Parameters
| Name | Type | Description |
|---|---|---|
requires_grad | bool | Whether autograd should record operations on this tensor; defaults to True. |
Returns
The same tensor, after its requires_grad setting is changed.
Key points
- Use the method to control whether autograd records later operations involving the tensor.
- The configured tensor itself is returned, allowing the call to participate in an expression or assignment.
- A common use is enabling autograd on an already-created tensor before subsequent computation.
- It is especially useful when the tensor currently has gradient recording disabled.
Common mistakes
- The Boolean argument controls whether autograd records operations involving the tensor.
Example: Enable gradients on an existing tensor
import torch
x = torch.tensor([1.0, 2.0])
x.requires_grad_()
y = (x ** 2).sum()
y.backward()
print(x.grad)Enables recording on the tensor in place, then computes and prints gradients for the resulting scalar.
When to use it
Start tracking after preprocessing
A preprocessing pipeline produces a tensor that should become the starting point for differentiable computation.
Disable recording
A tensor should stop participating in gradient recording for later operations.
Interview questions
QWhat does Tensor.requires_grad_() do?
It changes the tensor's requires_grad attribute in place and returns that same tensor. Its primary purpose is to start autograd recording for an existing tensor that currently does not require gradients.
Follow-up: Why might an existing tensor have gradients disabled? It may come from a DataLoader or from preprocessing or initialization, after which this method can enable recording.
QWhat parameter does the method accept, and what is its default?
The requires_grad parameter is a Boolean that determines whether autograd should record operations on the tensor. It defaults to True, so an argument-less call enables recording.
Follow-up: What call disables recording? Use tensor.requires_grad_(False), which changes the same setting in place.
QWhat does the method return, and why does in-place behavior matter?
The method returns the tensor after changing its setting. Because the change is in place, the original tensor is configured for subsequent autograd-tracked operations.
Follow-up: How can the enabled setting be used? Perform later operations and then apply backward computation to obtain gradients where autograd supports them.
Related
torch.Tensor.requires_gradtorch.Tensor.is_leaftorch.Tensor.gradtorch.Tensor.backwardtorch.nn.Module.requires_grad_