One of the variables needed for gradient computation has been modified by an inplace operation_

autograd
1

I have made a custom loss function which is taking parameters calculated from a intermediate function ( a function which is taking as input the output of my CNN network). but from loss.backward() I am getting the error
one of the variables needed for gradient computation has been modified by an inplace operation.
I couldn’t find out the source of the error so I am embedding both functions.
Loss function is


class my_Loss(torch.nn.Module):
    
    def __init__(self):
       super(my_Loss,self).__init__()
       #self.X = X
       #self.Y = Y
       
    def forward(self,Y,X):
        
       #Z = ..........
       .
       .
       .
       .
       pdb.set_trace()
       return F

and the intermediate function is

class fb(torch.nn.Module):
    
    def __init__(self):
       super(fb,self).__init__()
       #self.X = X
       #self.Y = Y

    def forward(self,Y,T,R,C,p0,Rinv,zz):
        ll = calc_ll(Y,T,R,C,p0,Rinv,zz)
        return ll
        
def calc_ll(Y,T,R,C,p0,Rinv,zz):
    #pdb.set_trace()
    cuda = torch.device("cuda:0")
    tensor = torch.tensor((), dtype=torch.float64,device=cuda)  
    [pp,tau] = Y.size(); [pp2,kk] = C.size()
    # initialize space
    alpha = tensor.new_zeros((kk,tau),requires_grad=False) ; 
    bb = tensor.new_zeros((kk,tau),requires_grad=False); rho = tensor.new_zeros((1,tau),requires_grad=False)
    
    # compute bb
    for ii in range(0,kk):
        dd = Y-torch.mm(torch.reshape(C[:,ii],(3,1)),tensor.new_ones((1,tau)))
        #dd.requires_grad=True
        R1 = (torch.mm(Rinv,dd))*dd
        #R1.requires_grad=True
        bb[ii,:]= torch.mul(torch.exp(-0.5*torch.sum(R1,0)),zz)
    #pdb.set_trace()
    alpha[:,[0]]= p0+bb[:,[0]]
    rho[0] = torch.sum(alpha[:,0])
    alpha[:,0] = alpha[:,0].clone()/rho[0]
    #pdb.set_trace()    
    for tt in range(1,tau):
        #pdb.set_trace()
        alpha[:,tt] = torch.transpose((torch.mv(torch.transpose(T,0,1),alpha[:,tt-1].clone())*bb[:,tt]),-1,0)
        rho[0,tt] = torch.sum(alpha[:,tt].clone())
        alpha[:,tt] = alpha[:,tt].clone()/rho[0,tt]
    pdb.set_trace()
    ll = torch.sum(torch.log(rho))/tau
    return ll

please help me to locate the error source line ?
Thanks in advance

2

I’d like to help, can you please provide a minimal script that will reproduce that error?

3

I don’t know exact location of the source of the error. I am getting the error like

  File "/home/ashishverma/anaconda3/lib/python3.6/site-packages/spyder/utils/site/sitecustomize.py", line 102, in execfile
    exec(compile(f.read(), filename, 'exec'), namespace)

  File "/home/ashishverma/Documents/Codes/Eye_traj_pred/mytrain.py", line 182, in <module>
    main(args)

  File "/home/ashishverma/Documents/Codes/Eye_traj_pred/mytrain.py", line 76, in main
    train(train_loader, net, criterion,ll_calculation, optimizer, epoch, train_args)

  File "/home/ashishverma/Documents/Codes/Eye_traj_pred/mytrain.py", line 168, in train
    loss.backward()

  File "/home/ashishverma/anaconda3/lib/python3.6/site-packages/torch/tensor.py", line 93, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph)

  File "/home/ashishverma/anaconda3/lib/python3.6/site-packages/torch/autograd/__init__.py", line 89, in backward
    allow_unreachable=True)  # allow_unreachable flag

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation
4

Hmm… I can’t seem to find an implace operator in your code, maybe I’m missing something though. @richard can you take a look?

5

Thank you for reply. For now I have resolved the issue by opening the below loop in terms of array

for tt in range(1,tau):
        #pdb.set_trace()
        alpha[:,tt] = torch.transpose((torch.mv(torch.transpose(T,0,1),alpha[:,tt-1].clone())*bb[:,tt]),-1,0)
        rho[0,tt] = torch.sum(alpha[:,tt].clone())
        alpha[:,tt] = alpha[:,tt].clone()/rho[0,tt]

now the error is gone. There must be another way to keep the loop but i couldn’t figure it out.
Still waiting for suggestions. Thanks in advance.