How to destroy computational graph after setting create_graph to True?

autograd
#1

Hi,

Can I ask how to destroy the computational graph after setting create_graph to True?

Say that I have:

x = torch.randn(10,5)
y = torch.randn(10,5)

def f(x,y):
    return torch.sum(x*y*y)

def jacobian_matrix_with_graph(x, y):
    jac = torch.autograd.functional.jacobian(f,(x,y),create_graph=True)
    return jac

def partial_hes(y): # only compute hessian matrix 1. f /partial_x /partial_y 2. f /partial_x /partial_y
    hes = jacobian_matrix_with_graph(x, y) # x being global variable
    return hes
    
hessian_matrix_pxpy = torch.autograd.functional.jacobian(partial_hes,(y))[1]
print(hessian_matrix_pxpy)

After I have used this computation graph, how do I destroy it manually?

Thanks in advance.

#2

There is no way to manually “destroy” the graph. It will be freed automatically once all references to it are gone. If you wish to obtain a tensor with the value of the output, but no longer attached to this graph, you can use detach