How to work with pin_memory?

[tea1528]

Really appreciate a great blog and clear demo.

I implemented my own version of COCO loader based on your TorchSerializedList and the default RAM usage decreases significantly. However, when I set pin_memory=True in PyTorch dataloader, I see that USS gets increased quite a lot while the shared memory stays the same for the default COCODetection and TorchSerializedList

	RSS	USS	Shared
default / pin_memory=True	7.31GB	3.45GB	3.87GB
default / pin_memory=False	4.31GB	455.76MB	3.87GB
shm / pin_memory=True	4.72GB	3.65GB	1.22GB
shm / pin_memory=False	1.72GB	508.69MB	1.22GB

From a quick search, the shared memory doesn't seem to work with pin_memory well according to this thread
. Detectron2 repo doesn't seem to use any pin_memory in the dataloader. Can you provide your thoughts on relationship of using the serialized dataset with pin_memory=True?

Thanks!