Automatic Gradients with PyTorch
Let’s see how we can compute automatic gradients with PyTorch and how PyTorch is different from plain Python and NumPy.
Creating a tensor with requires_grad=True argument
The following code creates a tensor called x just like before, but this time we give PyTorch an additional option requires_grad=True. We’ll see very soon what this option does.
Run the code to see what printing x does.
Press + to interact
import torch# pytorch tensorx = torch.tensor(3.5, requires_grad=True)print("x:", x)
We can see that x has a value of and is of type tensor. The output also shows that the tensor x has the requires_grad option set to True.
Functional relationship between variables
Let’s create a new variable y from x, just like before but this time using a different expression.
Press + to interact
import torch# pytorch tensorx = torch.tensor(3.5, requires_grad=True)print("x:", x)# y is defined as a function of xy = (x-1) * (x-2) * (x-3)print("y=(x-1) * (x-2) * (x-3):", y)
We can see that y is calculated from the expression (x-1) * (x-2) * (x-3).
As expected, the value of y is . This is because x is , so (3.5-1) * (3.5-2) * (3.5-3) is ...