Where can I find the backward function of torch.nn.functional.interpolate function?

autograd
1

Where can I find the backward function of torch.nn.functional.interpolate function?
https://pytorch.org/docs/stable/generated/torch.nn.functional.interpolate.html#torch.nn.functional.interpolate

2

[quote=“hao_xi, post:1, topic:213447, full:true”] Inscripción
Where can I find the backward function of torch.nn.functional.interpolate function?
https://pytorch.org/docs/stable/generated/torch.nn.functional.interpolate.html#torch.nn.functional.interpolate
[/quote]

Hello,

The backward function for torch.nn.functional.interpolate doesn’t exist directly in PyTorch as the operation is not inherently differentiable in a simple inverse sense. For operations like interpolation, where you upscale or downscale images, there’s no exact reverse operation, but you can achieve similar results by using the torch.nn.functional.grid_sample function for specific tasks, like reversing an upsampling by creating the appropriate grid for sampling.

3

Depending on the used algorithm, you should be able to find the used method by checking the grad_fn:

x = torch.randn(1, 10, 10, requires_grad=True)
torch.nn.functional.interpolate(x, size=(8)).grad_fn
<UpsampleNearest1DBackward0 object at 0x7f9f6f847e80>
5

Thank you very much for your reply. I also want to know where I can find the calculation source code for this upsampleNearest1DBackward0 implementation, or how can I manually call its gradient calculation function?

6

import inspect
data = torch.randn((1,3,224,224), dtype=torch.float32, requires_grad=True)

out = F.interpolate(data, scale_factor=2)
print(inspect.getmro(type(out.grad_fn)))

(<class ‘UpsampleNearest2DBackward1’>, <class ‘object’>)

The grad_fn of this tensor seems to be an object of a class. Then I have an idea. Can I manually call the function of this grad_fn to implement the gradient calculation I want to implement? For example, this scenario: I have a two-layer CNN, and then I want to implement its forward and reverse processes myself instead of calling the existing interface.

7

To implement a custom backward method you could write custom autograd.Functions as described in this tutorial.