When I run the pytorch memory profiler it returns the following:
Name Self CPU total % Self CPU total CPU total % CPU total CPU time avg CPU Mem Self CPU Mem CUDA Mem Self CUDA Mem Number of Calls
------------------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- ---------------
empty 1.11% 443.155ms 100.19% 39.822s 1.851ms 25.95 Gb 7.98 Gb 0 b 0 b
This ´empty´ is taking up a lot of memory (more than anything else in the network). I wonder whether this is expected behaviour, or a sign that there are a lot of empty tensors taking up memory, without being useful?
If so, where should I start looking for the culprits?