jacobian_hessian.py

import torch                                                                                          
                                                                                                      
def jacobian(y, x, create_graph=False):                                                               
    jac = []                                                                                          
    flat_y = y.reshape(-1)                                                                            
    grad_y = torch.zeros_like(flat_y)                                                                 
    for i in range(len(flat_y)):                                                                      
        grad_y[i] = 1.                                                                                
        grad_x, = torch.autograd.grad(flat_y, x, grad_y, retain_graph=True, create_graph=create_graph)
        jac.append(grad_x.reshape(x.shape))                                                           
        grad_y[i] = 0.                                                                                
    return torch.stack(jac).reshape(y.shape + x.shape)                                                
                                                                                                      
def hessian(y, x):                                                                                    
    return jacobian(jacobian(y, x, create_graph=True), x)                                             
                                                                                                      
def f(x):                                                                                             
    return x * x * torch.arange(4, dtype=torch.float)                                                 
                                                                                                      
x = torch.ones(4, requires_grad=True)                                                                 
print(jacobian(f(x), x))                                                                              
print(hessian(f(x), x))                                                                               

I don't think there's any other way with the current AD methods. You don't have to keep the whole Hessian in memory of course (you can throw away a row of the Hessian once you've picked out the element you're interested in), but you'll still need to compute each row, just like the hessian function does.

Thanks Adam!

Is there a reason why you use grad_y instead of just indexing flat_y[i] in the autograd?

Hello, I am relatively new to PyTorch and came across your Hessian function. It is much more elegant than some Hessian code from an academic paper that I am trying to reproduce. I've put together a toy example, but keep getting the error

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor [2]] is at version 4; expected version 3 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!

I've been scouring the docs and googling, but for the life of me I can't figure out what I'm doing wrong. Any help you could offer would be greatly appreciated!

Here is my code:

import torch
import torch.autograd as autograd
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import numpy as np

torch.set_printoptions(precision=20, linewidth=180)

def jacobian(y, x, create_graph=False):
    jac = []                             
    flat_y = y.reshape(-1)     
    grad_y = torch.zeros_like(flat_y)
    for i in range(len(flat_y)):         
        grad_y[i] = 1.
        grad_x, = torch.autograd.grad(flat_y, x, grad_y, retain_graph=True, create_graph=create_graph)
        jac.append(grad_x.reshape(x.shape))
        grad_y[i] = 0.
    return torch.stack(jac).reshape(y.shape + x.shape)           
                                                                                                      
def hessian(y, x):  
    return jacobian(jacobian(y, x, create_graph=True), x)                                             
                                                                                                      
def f(x):                                                                                             
    return x * x                                            

np.random.seed(435537698)

num_dims = 2
num_samples = 3

X = [np.random.uniform(size=num_dims) for i in range(num_samples)]

mean = torch.Tensor(np.mean(X, axis=0))
mean.requires_grad = True

cov = torch.Tensor(np.cov(X, rowvar=False))

with autograd.detect_anomaly():
    hessian_matrices = hessian(f(mean), mean)
    print('hessian: \n{}\n\n'.format(hessian_matrices))

The output with anomaly detection turned on is here:

---

RuntimeError Traceback (most recent call last)
in ()
67
68 with autograd.detect_anomaly():
---> 69 hessian_matrices = hessian(f(mean), mean)
70 print('hessian: \n{}\n\n'.format(hessian_matrices))

2 frames
in hessian(y, x)
45 print('--> hessian()')
46 j = jacobian(y, x, create_graph=True)
---> 47 return jacobian(j, x)
48
49 def f(x):

in jacobian(y, x, create_graph)
28 print('\tgrad_y: \n\t{}\n'.format(grad_y))
29
---> 30 grad_x, = torch.autograd.grad(flat_y, x, grad_y, retain_graph=True, create_graph=create_graph)
31 print('\tgrad_x: \n\t{}\n')
32

/usr/local/lib/python3.6/dist-packages/torch/autograd/init.py in grad(outputs, inputs, grad_outputs, retain_graph, create_graph, only_inputs, allow_unused)
155 return Variable._execution_engine.run_backward(
156 outputs, grad_outputs, retain_graph, create_graph,
--> 157 inputs, allow_unused)
158
159

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor [2]] is at version 4; expected version 3 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!

Finally, I'm running my code in a Google Colab notebook with PyTorch 1.4 if that makes a difference.

Thanks!

I did manage to get the code to run now. I made a "simplification" that broke it.

Your function f is:

def f(x):                                                                                             
    return x * x * torch.arange(4, dtype=torch.float)  

While mine was:

def f(x):                                                                                             
    return x * x  

I've since fixed it to:

def f(x):                                                                                             
    return x * x  * torch.ones_like(x)

and it works like a charm. @apaszke any idea why that is the case?

I did manage to get the code to run now. I made a "simplification" that broke it.

Your function f is:

def f(x):                                                                                             
    return x * x * torch.arange(4, dtype=torch.float)  

While mine was:

def f(x):                                                                                             
    return x * x  

I've since fixed it to:

def f(x):                                                                                             
    return x * x  * torch.ones_like(x)

and it works like a charm. @apaszke any idea why that is the case?

you can switch torch.ones_like(x) to 1 and it still works...

Hello Adam! How could I give credit to you if I use this code? Can it be a doc-string in documentation, paper citation or something?

Now the function torch.autograd.functional.jacobian can do the same thing, I think.

def jacobian(y, x, create_graph=False):
    # xx, yy = x.detach().numpy(), y.detach().numpy()
    jac = []
    flat_y = y.reshape(-1)
    grad_y = torch.zeros_like(flat_y)
    for i in range(len(flat_y)):
        grad_y[i] = 1.
        grad_x, = torch.autograd.grad(flat_y, x, grad_y, retain_graph=True, create_graph=True)
        jac.append(grad_x.reshape(x.shape))
        grad_y[i] = 0.
    return torch.stack(jac).reshape(y.shape + x.shape)


def hessian(y, x):
    return jacobian(jacobian(y, x, create_graph=True), x)


def f(xx):
    # y = x * x * torch.arange(4, dtype=torch.float)
    matrix = torch.tensor([[0.2618, 0.2033, 0.7280, 0.8618],
        [0.1299, 0.6498, 0.6675, 0.0527],
        [0.3006, 0.9691, 0.0824, 0.8513],
        [0.7914, 0.2796, 0.3717, 0.9483]], requires_grad=True)
    y = torch.einsum('ji, i -> j', (matrix, xx))
    return y


if __name__ == "__main__":
    # matrix = torch.rand(4, 4, requires_grad=True)
    # print(matrix)
    x = torch.arange(4,  dtype=torch.float, requires_grad=True)
    print(jacobian(f(x), x))
    grad = torch.autograd.functional.jacobian(f, x).numpy()
    # grad = grad.flatten()
    print(grad)
    # print(hessian(f(x, matrix), x))

output

        [0.1299, 0.6498, 0.6675, 0.0527],
        [0.3006, 0.9691, 0.0824, 0.8513],
        [0.7914, 0.2796, 0.3717, 0.9483]], grad_fn=<ViewBackward>)
[[0.2618 0.2033 0.728  0.8618]
 [0.1299 0.6498 0.6675 0.0527]
 [0.3006 0.9691 0.0824 0.8513]
 [0.7914 0.2796 0.3717 0.9483]]```

Hi,
I want to find a Hessian matrix for the loss function of the pre-trained neural network with respect to the parameters of the network. How can I use this method? Can someone please share an example? Thanks.

Hi,
I want to find a Hessian matrix for the loss function of the pre-trained neural network with respect to the parameters of the network. How can I use this method? Can someone please share an example? Thanks.

Hi,
I am looking for the same thing. Could you figure out how we can do it?

I think this has now been added to recent versions of torch's autograd module. Maybe look at the examples here

I think this has now been added to recent versions of torch's autograd module. Maybe look at the examples here

Right. I checked it. When I use this method I am getting multiple errors. I am looking for an example or similar code to see how the implementation is done.