GPU utility usage is zero and CPU utility is very high even with Pytorch Cuda version of code

vision
1

I am training a network and I put my network parameters and data loader on Cuda but when I run training, nvidia-smi shows 0% GPU usage with memory occupied. All the computation is running on CPU as CPU usage is very high. I checked every variable and networks parameters with “variable_name.is_cuda” and “next(net.parameters()).is_cuda” respectively in my training loop. They are all true. Can anyone help me regarding this issue?

1 Like
2

You will have to share the code for this one, or an abstract of it.

3

I am sending an abstract of my code. I am sending data and network to cuda but they are running on cpu. Can you help?

def to_var(self, x):
        if torch.cuda.is_available():
            x = x.cuda()
        return Variable(x)

net = network.generator()
params = net.parameters()
optimizer = optim.Adam(params, lrG, [beta1, beta2], amsgrad=True)

if torch.cuda.is_available():
     net.cuda()

for i, data in enumerate(t_loader):

                source_img, target_img  = data[0], data[1]

                source_img = to_var(source_img)
                target_img = to_var(target_img)

                fake_generated_target = net(source_img)

                g_loss = torch.mean((fake_generated_target- 1) **2)

                optimizer.zero_grad()
                g_loss.backward()
                self.optimizer.step()```
4

You have to call “net = network.generator().cuda()” and give its paramters to the optimizer.