rusty1s/himp-gnn

Getting RuntimeError

lancioni opened this issue · 3 comments

lancioni commented

When I try to reproduce experiments as stated, I invariably get an error:

RuntimeError: received 0 items of ancdata

The error is somewhat erratic, e.g. in train_zinc_subset.py I get it after 15 to 50 epoch in the first run. It seems to have to do with some kind of memory of other saturation I cannot detect.
Other examples in Python Geometric work just fine, so I think it is something specific to the code in himp-gnn.
I am using Python 3.6.9 under WSL2 Ubuntu 18.04.

rusty1s commented

Adding torch.multiprocessing.set_sharing_strategy('file_system') at the top of the file is likely to fix this issue, see here. Not exactly sure what's causing it though.

lancioni commented

Thank you, it's working now!

lancioni commented

Another runtime error, again erratical and coming out when running train_zinc_full.py (after adding torch.multiprocessing.set_sharing_strategy('file_system')):

RuntimeError: unable to open shared memory object </torch_5535_1638182315> in read-write mode

Some discussion in other pytorch apps suggests it is a memory issue and advises to lower the number of workers. I have no idea if this is the issue, though.