Cuda and pytorch memory usage

Question

I am using Cuda and Pytorch:1.4.0.

When I try to increase batch_size, I've got the following error:

CUDA out of memory. Tried to allocate 20.00 MiB (GPU 0; 4.00 GiB total capacity; 2.74 GiB already allocated; 7.80 MiB free; 2.96 GiB reserved in total by PyTorch)

I haven't found anything about Pytorch memory usage.

Also, I don't understand why I have only 7.80 mib available?

Should I just use a videocard with better perfomance, or can I free some memory? FYI, I have a GTX 1050 TI, python 3,7 and torch==1.4.0 and my os is Windows 10.

Please don't post error messages as images. They can't be searched for by future visitors — talonmies
– talonmies, Commented Feb 18, 2020 at 8:36

Jeril · Accepted Answer · 2020-02-18 09:06:35Z

5

I had the same problem, the following worked for me:

torch.cuda.empty_cache()
# start training from here

Even after this if you get the error, then you should decrease the batch_size

answered Feb 18, 2020 at 9:06

Jeril

8,6316 gold badges60 silver badges74 bronze badges

Sign up to request clarification or add additional context in comments.

1 Comment

André Pacheco Over a year ago

To complement, one can check the GPU memory using nvidia-smi command on terminal. Also, if you're storing tensors on GPU you can move them to cpu using tensor.cpu(). I solve most of my problems with memory using these commands.

Collectives™ on Stack Overflow

Cuda and pytorch memory usage

1 Answer 1

1 Comment

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

1 Comment

Your Answer

Sign up or log in

Post as a guest

Linked

Related