pocket.utils.DetectionAPMeter [Errno24] Too many open files

[fredzzhang]

Following pocket.utils.AveragePrecisionMeter, DetectionAPMeter uses multiprocessing to evaluate APs of different classes. multiprocessing.Pool.imap is used specifically. There have been issues previously regarding the choice of chunk size. It was resolved by applying the heuristics adopted in multiprocessing.Pool.map. The current issue is a bit bizarre because Errno24 appears when there are a large number of classes (e.g. >300) . Yet AveragePrecisionMeter, with similar mechanism, works perfectly regardless of the number of classes or examples.

[fredzzhang]

The distinction of DetectionAPMeter from AveragePrecisionMeter is the number of examples being different across classes. But even when the numbers of samples are forced to be the same, the error still occurs.

[fredzzhang]

A simple fix for this issue is to increase the number of files allowed to be opened, using the following command (replace $NUM with a number)
ulimit -n $NUM
For a linux machine, the current limits can be accessed using the following commands
ulimit -Sn for currently applied soft limit (1024 by default)
ulimit -Hn for hard limit set by system (4096)
Changing the soft limit to 4096 works for the previous scenario that generated the error. The downside is that this command has to be run every time when starting a new terminal. A workaround this is to write it in .basrhc.

[fredzzhang]

The problem appears to have been fixed with 94e22cf, which specified the number of subprocesses to spawn