-
-
Save sbarratt/37356c46ad1350d4c30aefbd488a4faa to your computer and use it in GitHub Desktop.
| def get_jacobian(net, x, noutputs): | |
| x = x.squeeze() | |
| n = x.size()[0] | |
| x = x.repeat(noutputs, 1) | |
| x.requires_grad_(True) | |
| y = net(x) | |
| y.backward(torch.eye(noutputs)) | |
| return x.grad.data |
Here's a differentiable batch version that doesn't loop over the batch dim (there's still one for loop left over the input dimension though... not sure if it's possible to get rid of that):
def jacobian(f, x):
"""Computes the Jacobian of f w.r.t x.
This is according to the reverse mode autodiff rule,
sum_i v^b_i dy^b_i / dx^b_j = sum_i x^b_j R_ji v^b_i,
where:
- b is the batch index from 0 to B - 1
- i, j are the vector indices from 0 to N-1
- v^b_i is a "test vector", which is set to 1 column-wise to obtain the correct
column vectors out ot the above expression.
:param f: function R^N -> R^N
:param x: torch.tensor of shape [B, N]
:return: Jacobian matrix (torch.tensor) of shape [B, N, N]
"""
B, N = x.shape
y = f(x)
jacobian = list()
for i in range(N):
v = torch.zeros_like(y)
v[:, i] = 1.
dy_i_dx = grad(y,
x,
grad_outputs=v,
retain_graph=True,
create_graph=True,
allow_unused=True)[0] # shape [B, N]
jacobian.append(dy_i_dx)
jacobian = torch.stack(jacobian, dim=2).requires_grad_()
return jacobian
EDIT: x needs to have requires_grad=True.
get_jacobian input x has batch case without loop.
import torch
def get_batch_jacobian(net, x, noutputs):
x = x.unsqueeze(1) # b, 1 ,in_dim
n = x.size()[0]
x = x.repeat(1, noutputs, 1) # b, out_dim, in_dim
x.requires_grad_(True)
y = net(x)
input_val = torch.eye(noutputs).reshape(1,noutputs, noutputs).repeat(n, 1, 1)
y.backward(input_val)
return x.grad.data
usage
class Net(torch.nn.Module):
def __init__(self, in_dim, out_dim):
super(Net, self).__init__()
self.fc1 = torch.nn.Linear(in_dim, out_dim)
def forward(self, x):
return torch.nn.functional.relu(self.fc1(x))
batch = 2
num_features = 3
num_outputs = 5
x = torch.randn(batch, num_features)
print(x.shape)
net = Net(num_features, num_outputs)
print(net(x).shape)
result = get_batch_jacobian(net, x, num_outputs)
print(result.shape)
Here my code tensor shape
input x: torch.Size([2, 3])
output y: torch.Size([2, 5])
jacobian dy/dx: torch.Size([2, 5, 3])
I checked consistency https://gist.github.com/sbarratt/37356c46ad1350d4c30aefbd488a4faa#file-torch_jacobian-py
ret0 = get_batch_jacobian(net, x, num_outputs) # my code
ret1 = get_jacobian(net, x[0], num_outputs) # sbarratt code
ret2 = get_jacobian(net, x[1], num_outputs) # sbarratt code
ret0[0] == ret1
ret0[1] == ret2
@MasanoriYamada , can you explain how this works? What does torch.eye(noutputs).reshape(1,noutputs, noutputs).repeat(n, 1, 1) do?
@MasanoriYamada , how can I make this work when my inputs are 4d Tensors? When I try calling the model's forward method, I receive the following error:
RuntimeError: Expected 4-dimensional input for 4-dimensional weight 6 3 5 5, but got 5-dimensional input of size [batch_size, output_dimension, num_channels, image_width, image_height] instead
@RylanSchaeffer, My code only supports flatten tensor with batch. Please, show me that input-output tensor and network in your case (simple network is the best)
@RylanSchaeffer, how about this https://gist.github.com/MasanoriYamada/d1d8ca884d200e73cca66a4387c7470a
Disclaimer: Value not tested for correctness
Question for anyone: why do we need to tile the input before passing it through the graph (net, in sbarratt's original code)? Why can't we tile the input and the output after the forward pass?
I'm trying to do this currently, but I'm receiving the error: One of the differentiated Tensors appears to not have been used in the graph. Set allow_unused=True if this is the desired behavior.
Here's what I'm doing. Let x be the input to the graph with shape (batch size, input dimension) and let y be the output of the graph with shape (batch size, output dimension). I then select a subset of N random unit vectors. I stack x with itself and y with itself as follows:
x = torch.cat([x for _ in range(N)], dim=0)
and
y = torch.cat([y for _ in range(N)], dim=0)
x then has shape (N * batch size, input dim) and y has shape (N * batch size, output dim). But then, when I try to use autograd, I receive the aforementioned error .
jacobian = torch.autograd.grad(
outputs=y,
inputs=y,
grad_outputs=subset_unit_vectors,
retain_graph=True,
only_inputs=True)[0]
RuntimeError: One of the differentiated Tensors appears to not have been used in the graph. Set allow_unused=True if this is the desired behavior.
Does anyone know why this is, and is there a way to make this post-forward pass tiling work?
@RylanSchaeffer
I was trying the same thing with unsqueeze().expand(), but it leads to the same autograd error. I suppose it's because the newly created x and y nodes are just hanging in the computation graph, and do not really have a dependency, so autograd would no longer work.
I met this page about 1 year ago. This is really a nice trick, but it's a pity that it needs to forward pass a large batch and becomes a huge challenge to my GPU room. Recently I found an interesting way to bypass this problem. It's really interesting to solve a problem I encountered 1 year ago. https://github.com/ChenAo-Phys/pytorch-Jacobian
If I'm understanding this correctly, this code will forward pass noutputs times just to compute the jacobian once (but do it in a vectorized way)... The 1.5.0 autograd jacobian computation seems to compute the output once but then forloops over it and call backward one by one (@rjeli first comment) which will for sure be slow... Both tradeoffs seem sub optimal.
Anyone know if there's an update on this? Or is pytorch really not meant to compute jacobians?
@justinblaber , autodiff either computes matrix-vector products or vector-matrix products (depending on forward mode / reverse mode). The Jacobian is a matrix - there's no easy way to recover this by itself. Either you perform multiple backwards passes, using different elementary basis vector on each pass, or you blow the batch size up and do one massive backwards pass. There's no way around this.
how about this experimental api for jacobian: https://pytorch.org/docs/stable/_modules/torch/autograd/functional.html#jacobian
is it good?
how about this experimental api for jacobian: https://pytorch.org/docs/stable/_modules/torch/autograd/functional.html#jacobian
is it good?
I took a look and:
for j in range(out.nelement()):
vj = _autograd_grad((out.reshape(-1)[j],), inputs, retain_graph=True, create_graph=create_graph)
It's just for-looping over the output and computing the gradient one by one (i.e. each row of the jacobian one by one). This will for sure be slow as hell if you have a lot of outputs. I actually think it's a tad bit deceiving that they advertise this functionality, because really the functionality just isn't there.
And actually, to be honest I wanted the jacobian earlier to do some gauss newton type optimization, but I've actually since discovered that the optim.LBFGS optimizer (now built into pytorch) might work well for my problem. I think it even has some backtracking type stuff built into it. So for now I don't think I even need the jacobian anymore.
where is
nused?