To calculate custom gradients in PyTorch, you can define a function with the `@torch.autograd.Function` decorator. This function should have forward and backward methods, where forward computes the output and backward computes the gradients of the inputs. You can then use this custom function as a building block to create complex differentiable operations.

Another way to calculate custom gradients is by defining your own autograd function using `torch.autograd.grad`. This allows you to compute gradients using any mathematical operations or external libraries. This flexibility is particularly useful when dealing with non-standard or complex gradients.

PyTorch also provides a higher-level interface called `torch.autograd.Function` that allows you to define custom layers and operations with autograd support. Using this interface, you can easily create reusable custom layers with custom backward passes.