'CUDA error: out of memory' after several epochs


#1

Hi, I’m having some trouble which results in the following error:

  pred_log_probs = estimator.forward(x_train[:, :])
  File "D:\Google_Drive\Projects\FeedForward_detection\NN_detector_module_seperated_output.py", line 84, in forward
    x = F.relu((self.layer_1(x)))
  File "C:\Users\sholev\AppData\Local\Continuum\anaconda3\lib\site-packages\torch\nn\functional.py", line 643, in relu
    return torch.relu(input)
RuntimeError: CUDA error: out of memory

The strange thing is that this error arises after 7 epochs, so it seems like some GPU memory allocation is not being released.

The NN architecture is the following:

class symbols_detector(nn.Module):
    def __init__(self, num_of_symbols, N1, N2):
        super(symbols_detector, self).__init__()
        self.num_of_symbols = num_of_symbols
        self.N1 = N1
        self.N2 = N2
        self.Lin = (N1*2 + 2)*N2
        self.layer_1 = nn.Linear(self.Lin , L1) 
        self.layer_2 = nn.Linear(L1,  L2)
        self.layer_3 = nn.Linear(L2,  L3)
        self.layer_4 = nn.Linear(L3,  L4)
        self.layer_5 = nn.Linear(L4, self.num_of_symbols * N2)
        self.log_p = nn.LogSoftmax(dim=2)

    def forward(self, x):
        x = F.relu((self.layer_1(x)))
        x = F.relu((self.layer_2(x)))
        x = F.relu((self.layer_3(x)))
        x = F.relu((self.layer_4(x)))
        x = self.layer_5(x)
        x = x.view(-1, self.2, self.num_of_symbols) 
        x = self.log_p(x)
        return x

The training code is the following:

rand_idx = torch.randperm(x_train.shape[0])
    x_train.pin_memory()
    x_train = x_train[rand_idx, :].cuda()

    y_train = y_train[:, rand_idx]
    y_train = y_train.cuda()

    pred_log_probs = estimator.forward(x_train[:, :])
    train_loss = torch.zeros([ep_num+1])
    SER_train = np.zeros([ep_num+1])
    BER_train = np.zeros([ep_num + 1])
    train_loss[0] = cost_func(pred_log_probs.permute([0, 2, 1]), y_train[0, :])
train_loss[0].cpu().data.numpy())

    for i in range(0, ep_num):
        estimator.train()
        for j in range(0, (int(N_train_samples/batch_Size)-2)):
            #print(j)
            pred_log_probs = estimator.forward(x_train[j * batch_Size:(j + 1) * batch_Size, :])
            model_optimizer.zero_grad()
            loss1 = cost_func(pred_log_probs.permute([0, 2, 1]), y_train[0, j * batch_Size:(j + 1) * batch_Size])
            loss1.backward()
            model_optimizer.step()

        estimator.eval()
        pred_log_probs = estimator.forward(x_train[:, :])
        train_loss[i+1] = cost_func(pred_log_probs.permute([0, 2, 1]), y_train[0, :])

        rand_idx = torch.randperm(x_train.shape[0])
        x_train = x_train[rand_idx, :].cuda()
        y_train = y_train[:, rand_idx].cuda()

print('model trained data set')
 return train_loss

Is there something I’m missing that causes the memory the overflow?
Thanks


(Lelouedec) #2

Maybe don’t use

pin_memory()


#3

It looks like you are directly appending the training loss to train_loss[i+1], which might hold a reference to the computation graph. If that’s the case, you are storing the computation graph in each epoch, which will grow your memory. You need to detach the loss from the computation, so that the graph can be cleared.
Change this line of code to:

train_loss[i+1] = cost_func(pred_log_probs.permute([0, 2, 1]), y_train[0, :]).detach().item()

and try it again.


#4

I t really solved it!
Thanks for the catch!