Trouble avoiding "Trying to backward through the graph a second time" error on simple model

autograd
1

Hi,

I’m trying to teach myself the nuts and bolts of pytorch by implementing a relatively low level model (a kalman filter).

I keep running into the error: Trying to backward through the graph a second time, but the buffers have already been freed. Specify retain_graph=True when calling backward the first time.

I know this is a common error and there are a lot of posts with responses asking about it, but I haven’t gotten any solution to work. This is may be because I can’t see how the situation applies to my model, which makes me concerned I don’t understand the nuts and bolts very well.

I’ve tried to simplify the example as much as possible:

import torch
from torch.autograd import Variable
from torch.nn import Parameter
from torch import Tensor

N = 50
vel_true = torch.cumsum(torch.randn(N)/10., 0) # true velocity
pos_true = torch.cumsum(vel_true, 0) # true position
pos_obs = pos_true + torch.randn(N) # observed position

class KalmanFilter1(torch.nn.Module):
    def __init__(self):
        super(KalmanFilter1, self).__init__()
        self.log_obs_std = Parameter(Tensor([ 0.]))
        
        # if this parameter is set like so, the error doesn't occur:
        # self.log_process_std = Parameter(Tensor([-1.])) 

        # if the parameter is defined as its ratio to the other parameter, the error occurs.
        # no real reason to do this, but it was the easiest way to make the error occur in
        # this simple example (I originally encountered the error in a multivariate kalman-filter
        # but I think that code would just make the problem more obscure)
        self.log_ratio = Parameter(Tensor([ 0.]))
        self.log_process_std = torch.log( torch.exp(self.log_obs_std) * torch.exp(self.log_ratio) )
        
    def predict(self, x, k_mean, k_var):
        k_mean = k_mean
        k_var = k_var + torch.pow(torch.exp(self.log_process_std),2)
        return k_mean, k_var
    
    def update(self, x, k_mean, k_var):
        resid = x - k_mean
        K = k_var / ( k_var + torch.pow(torch.exp(self.log_obs_std),2) )
        k_mean = k_mean + K*resid
        k_var = (1.-K) * k_var
        return k_mean, k_var
        
    def forward(self, x):
        output = []
        
        k_mean = Variable(torch.zeros(1))
        k_var = Variable(torch.ones(1))
        output.append(k_mean)
        for i in range(len(x)-1):
            k_mean, k_var = self.predict(x[i], k_mean, k_var)
            k_mean, k_var = self.update(x[i], k_mean, k_var)
            output.append(k_mean)

        output = torch.cat(output, 0)
        return output

model = KalmanFilter1()
optimizer = torch.optim.RMSprop(model.parameters(),lr=0.01)

for t in range(100):
    obs_x = Variable(pos_obs)
    pred = model(obs_x)
    loss = torch.sum(torch.pow(obs_x - pred, 2))
    optimizer.zero_grad()
    loss.backward(retain_graph=False)
    optimizer.step()

As far as I can understand what retaining the graph means, I don’t think it should be needed here, as each training iteration is independent – I think I must be specifying something incorrectly about my model. Any help conceptually understanding why this error is occurring would be appreciated.

2

PyTorch uses dynamic graph. It’s not symbolic expression as tf is. You are backproping via this graph multiple times. Manually compute at each iteration please.

3 Likes
3

Hello,
I faced this error once. Here is the solution: RuntimeError: Trying to backward through the graph a second time, but the buffers have already been freed. Specify retain_graph=True when calling backward the first time

In my case, since I called .backward() twice, It happend to me. I fixed this by adding retain_graph = True as an input of the first backward call

4

This was the trick! I had exactly the conceptual confusion that you’ve identified. It was also the source of my problems in my original code.

Thanks!

1 Like
5

Hey @jwdink, what was the change you made in your code to get it to work finally? I have a similar optimization process I am trying to solve and not sure how to work around this error.