Running on GPU with less memory - CUDA out of memeory

[syed-abdul-baqi]

I was trying this model on GTX 1080 Ti (12GB memory). All trainings have worked fine for me but with smaller batch size.
However, testing for I-frame model is giving CUDA out of memory error. Which is strange because testing should not require that much or memory especially when training has worked fine.
What is the best way to resolve this issue?

[SethuGitHub]

I was trying this model on GTX 1080 Ti (12GB memory). All trainings have worked fine for me but with smaller batch size.
However, testing for I-frame model is giving CUDA out of memory error. Which is strange because testing should not require that much or memory especially when training has worked fine.
What is the best way to resolve this issue?

Hi Syed,
Please try to change the test.py as below.
In forward_video(),
def forward_video(data):
#input_var = torch.autograd.Variable(data, volatile=True)
input_var = torch.autograd.Variable(data)
with torch.no_grad(): #sethu
scores = net(input_var)
scores = scores.view((-1, args.test_segments * args.test_crops) + scores.size()[1:])
scores = torch.mean(scores, dim=1)
return scores.data.cpu().numpy().copy()
Try this and check.
-- [email protected]

[Maruidear]

Hi，I met this problem in training. How can I solve it?

[Maruidear]

I was trying this model on GTX 1080 Ti (12GB memory). All trainings have worked fine for me but with smaller batch size.
However, testing for I-frame model is giving CUDA out of memory error. Which is strange because testing should not require that much or memory especially when training has worked fine.
What is the best way to resolve this issue?

Hi，I met this problem in training.Can you tell me about your batch-size in training?

[syed-abdul-baqi]

Reduce the batch size for training.
It was reduced to 4 for I-Frame. You can go further down to 2 or even 1.
Batch-size 40 worked fine for mv and residual frame.

In testing, passing the following parameter in the command made it run successfully.

--test-crops=1

Note that the default value is 10 for test crops.

Best of luck

[JimLee1996]

Half precision is also helpful to reduce memory consumptiuon. Try use tensor.half() and model.half().