I have batch size of 1 and I am trying to run on multiple GPUs because I need the large memory given I want a large input image into the classifier.
I have a Tesla K80, and GTX 1080 on the same device (total 3) but using DataParallel will cause an issue so I have to exclude the 1080 and only use the two K80 processors. The issue is that they each only have 12GB of RAM (11.2 effective). When I run the image size around 128, 128, 96 it consumes 11GB, so I can run a batch of 2 on these two GPUs. However, I want to run a single batch with larger image size.
Is it even possible to achieve or am I stuck with a small input size?
If there is no solution in PyTorch, would it be possible to virtualize the K80 to represent a single 24GB processor and still use CUDA with it?