RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor []], which is output 0 of SelectBackward, is at version 499; expected version 498 instead

Hi everyone,

I am getting this error but I do not see why. I wanted to add a custom parameter to my RNN. The code is as follows :

class RNN_pre_d(nn.Module):
def init(self, hidden_dim=64, num_layers=1):
super(RNN_pre_d, self).init()
self.num_layers = num_layers
self.hidden_dim = hidden_dim
self.lstm = nn.LSTM(1, self.hidden_dim, self.num_layers)
self.fcF_c = nn.Linear(self.hidden_dim, 1)
w_final = torch.randn(1)
self.w_final = nn.Parameter(w_final)

def forward(self, x):
    x1, (h_1, c_1) = self.lstm(x)    ## x.shape = [sequence_len, batch_size, hidden_dim]
    c_train = torch.log(1+torch.exp(self.fcF_c(x1.squeeze(1))))
    c_train = c_train.squeeze(-1)
    for i in range(1,len(c_train)):
        c_train[i] = c_train[i] + ( 1/(1+torch.exp(-self.w_final)) ) * c_train[i-1] 

    return c_train, h_1, c_1  

the problem only appears when trying to backprop

NumbEpochs = 100

learning_rate = 10e-04
NumbHiddLayer = 1
sizeHidd = 128

nn_pre_d = RNN_pre_d(hidden_dim = sizeHidd,
num_layers = NumbHiddLayer)

optimizer = optim.Adam(nn_pre_d.parameters(),

MSE loss function

loss = nn.MSELoss() ; listLossTrain = []

Training data standardized

y = np.random.normal(0, 0.1, 500)
mean_train = np.mean(y) ; std_train = np.std(y) ; Z_stand_train = (y-mean_train) / std_train
data_nn_stand_train = torch.FloatTensor([Z_stand_train])
data_nn_stand_train = data_nn_stand_train.permute(1,0).unsqueeze(-1)
target_train = torch.tensor(y)


for epoch in range(NumbEpochs):

# Make the prediction in training set.
c_train, h_1, c_1  = nn_pre_d.forward(data_nn_stand_train)
# Compute a scalar loss in training set.
loss_train = loss(target_train, c_train)

# Compute the gradient for the model parameters and update model parameters.


Thank you!

I guess the assignment to c_train[i] might raise this error. Could you append the results to a list and use torch.stack on it instead?

PS: you can post code snippets by wrapping your code into three backticks ```, which would make debugging easier. :wink:

Thank you very much! It’s working!

I will take notice for next time!