Compute the Jacobian of a Given Function in PyTorch

The jacobian() function computes the Jacobian of a given function. The jacobian() function can be accessed from the torch.autograd.functional module. The function whose Jacobian is being computed takes a tensor as the input and returns a tuple of tensors or a tensor. The jacobian() function returns a tensor with Jacobian values computed for a function with the given input.

Syntax

torch.autograd.functional.jacobian(func, input)

Parameters

• func − It's a Python function for which the Jacobian is computed.

• input − It’s the input to the function, func.

Steps

We could use the following steps to compute the Jacobian of a given function −

• Import the required library. In all the following examples, the required Python libraries are torch. Make sure you have already installed it.

import torch
from torch.autograd.functional import jacobian

• Define a function func for which the Jacobian is to be calculated. The input to this function is input.

def func(x):
   return x**3 + 4*x -10

• Define the tensor input to the function, func.

input = torch.tensor([2.,3.,4.])

• Compute the Jacobian of the function defined above for the given input input.

output = jacobian(func, input)

• Print the tensor containing the computed Jacobians.

print("Jacobians Tensor:
", output)

Example 1

# Import the required libraries
import torch
from torch.autograd.functional import jacobian

# define a function
def func(x):
   return x**3 + 4*x -10

# define the inputs
input1 = torch.tensor([2.])
input2 = torch.tensor([2.,3.])
input3 = torch.tensor([2.,3.,4.])

# compute the jacobians
output1 = jacobian(func, input1)
output2 = jacobian(func, input2)
output3 = jacobian(func, input3)

# print the Jacobians calculated above
print("Jacobian Tensor:
", output1)
print("Jacobian Tensor:
", output2)
print("Jacobian Tensor:
", output3)

Output

Jacobian Tensor:
   tensor([[16.]])
Jacobian Tensor:
   tensor([[16., 0.],
      [ 0., 31.]])
Jacobian Tensor:
   tensor([[16., 0., 0.],
      [ 0., 31., 0.],
      [ 0., 0., 52.]])

In the above example, we computed the Jacobians for a function for different inputs.

Example 2

import torch
from torch.autograd.functional import jacobian

# define a function
def func(x,y):
   return x.pow(3) + y

# here input is tuple of two tensors, one for x and other for y
input1 = (torch.tensor([2.]), torch.tensor([5.]))
input2 = (torch.tensor([2., 3., 4.]), torch.tensor([5., 6., 7.]))

output1 = jacobian(func, input1)
output2 = jacobian(func, input2)

print(output1)
print(output2)

Output

(tensor([[12.]]), tensor([[1.]]))
(tensor([[12., 0., 0.],
   [ 0., 27., 0.],
   [ 0., 0., 48.]]), tensor([[1., 0., 0.],
   [0., 1., 0.],
   [0., 0., 1.]]))