Cuda out of memory error. Why do my tensors use too much memory?

Hello.
I think I’ve done wrong. It uses too much memory.

Before training I made sure any memory isn’t allocated,
but it dies saying
RuntimeError: CUDA out of memory. Tried to allocate 14.00 MiB (GPU 0; 7.93 GiB total capacity; 6.30 GiB already allocated; 25.75 MiB free; 6.78 GiB reserved in total by PyTorch)

I do not know why 6.3GB is alreay allocated.

Is there something I am doing wrong?
This is my training function.

‘’’
def train(input, target):
# initializing grad and hidden state
hidden = model.init_hidden()
optimizer.zero_grad()
loss = 0

for i in range(input.size()[0]):
    output, hidden = model(input.data[i].unsqueeze(0), hidden)
    l = criterion(output, target.data[i])
    loss += l
    output.detach()

# backpropagation
loss.backward()

# optimizing
optimizer.step()

# learning rate scheduling
scheduler.step()

return loss

‘’’

this is my gru unit

‘’’
import torch.nn as nn
import torch

class GRU(nn.Module):
def init(self, input_size, hidden_size, output_size, batch_size, device, num_layers):
super(GRU, self).init()
self.hidden_size = hidden_size
self.num_layers = num_layers
self.batch_size = batch_size
self.gru = nn.GRU(input_size, hidden_size, num_layers, batch_first=False, dropout=0.2)
self.elu = nn.ELU()
self.fc = nn.Linear(hidden_size, output_size)
self.device = device
self.to(device)

def forward(self, x, h):
    out, h = self.gru(x, h)
    out = self.elu(h[0])
    out = self.fc(out)
    return out, h


def init_hidden(self):
    hidden = torch.zeros(self.num_layers, self.batch_size, self.hidden_size).to(self.device)
    return hidden

‘’’

The losses will be accumulated to loss in the first loop and with them the complete computation graphs.
It’s thus expected that your memory usage would grow and your GPU might run out of memory.
You could have a look at this post and see if another approach would be suitable for your use case.

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