Deeplab error

[dinisov]

Hi I am trying to train a Deeplab semantic segmentation model in detectron2. However, I keep getting CUDA out of memory errors. My images are 1024x1024 which, for object detection, is fairly standard.

My machine has two RTX 2080 Ti GPUs, each with 11GB memory.

I don't think 1024x1024 is large, so before I start reducing image size I wanted to ask if there is anything else that could be generating these CUDA out of memory errors.

The configuration is the same as in detectron2/projects/DeepLab/ except for

cfg.MODEL.ROI_HEADS.NUM_CLASSES = 1
cfg.MODEL.SEM_SEG_HEAD.NUM_CLASSES = 1

As there is only one class (I am trying to segment foreground from background).

EDIT: actually with cfg.SOLVER.IMS_PER_BATCH = 2 I get a different error message:

ValueError: pool_kernel_size must be divisible by the shape of inputs. Input size: torch.Size([32, 32]) pool_kernel_size: (32, 64)