JiaRenChang
why does the testing process need more GPU memory than the training process
with the same size of pictures. I could train my own dataset with the PSMNet code, the train process needs about 8 G GPU memory, but failed to test on it, the error says ' out of memory', why? @JiaRenChang @mileyan
Hi, I think your title and comment contradict each other. In the title you're saying that training needs more GPU memory than testing, and in the description you're sayig that testing needs more GPU memory than training. Which is right?
In my experience the training process needs the most memory. I'm able to test the network using a GTX1060 6GB, but when I try to train a model, I get an RuntimeError: CUDA error: out of memory.
@fela98 thanks, my fault, the description is not clear, I have have edited the title. The fact is when I use submission.py to test the images, and it gives 'out of memory', but the training is OK with the same size of images. Could you give me some advise?
Hi, @passion3394 One pair of KITTI images in size 384x1248 takes around 4GB GPU memory in testing. You would check that the memory usage of your devices.
@JiaRenChang Thanks for your reply. Because of the size of my picture is about 3000*3000, it's too big to test on the PSMNet, could you give me some advise?
hello, @JiaRenChang my computure will be reboot when I run your program,I was so confused,could you give me some advise? thank you
@JiaRenChang Thanks for your reply. Because of the size of my picture is about 3000*3000, it's too big to test on the PSMNet, could you give me some advise?
@passion3394 I'm facing the same problem. Did you resolve it?
@Huyhust13 You could resize the pic to the size which could be loaded by the GPUs.