Delete Custom Instances Pretty Slow

[jmdelahanty]

Hey SLEAPy heads!

I thought my videos were solid with predictions but found a couple where my fish ended up in a spot I didn't expect/label very well, but still valid. I'm trying to use Labels -> Custom Instance Delete -> Delete Predicted Instances in Current Video

To delete about 57k predictions, it took around 5 minutes. I read a little about this kind of thing here but wanted to see if that's expected. It's not a big deal that it takes a little bit to do, but just wanted to ask about it.

[roomrys]

Hi @jmdelahanty,

I started a little investigation and have not finished yet, but need to switch tasks now - will update when I get a chance.

Profiling

If you run sleap-label --profiling then redo the action, we should be able to narrow it down to the function that is obviously taking the most time

(sleap_1.4) me:~$sleap-label $dsdmc --profiling

Happy SLEAPing! :)
Saving config: /Users/me/.sleap/1.4.1a2/preferences.yaml
         223035 function calls (222552 primitive calls) in 32.377 seconds

   Ordered by: standard name

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
      153    0.000    0.000    0.000    0.000 <attrs generated init sleap.gui.widgets.slider.SliderMark>:1(__init__)
        2    0.000    0.000    0.000    0.000 <attrs generated init sleap.instance.LabeledFrame>:1(__init__)
        1    0.000    0.000   32.377   32.377 <string>:1(<module>)
        1    0.000    0.000    0.000    0.000 QtGui.py:35(__getattr__)
        1    0.000    0.000    0.002    0.002 __init__.py:215(dump_all)
        1    0.000    0.000    0.002    0.002 __init__.py:248(dump)
        1    0.000    0.000    0.000    0.000 _bootlocale.py:33(getpreferredencoding)

My readily-available test dataset isn't so big so all the times are pretty small, but it would be more prevelant in your large dataset.

Code Investigation

Immediately after click "Delete Predicted Instances in Current Video", we loop over all LabeledFrames and all Instances in each labeled frame:

sleap/sleap/gui/commands.py

Lines 2324 to 2334 in d3ad226

	class DeleteAllPredictions(InstanceDeleteCommand):
	@staticmethod
	def get_frame_instance_list(
	context: CommandContext, params: dict
	) -> List[Tuple[LabeledFrame, Instance]]:
	return [
	(lf, inst)
	for lf in context.labels
	for inst in lf
	if type(inst) == PredictedInstance
	]

. Then, we confirm the deletion with the user and start the actual deletion where we loop over only the LabeledFrames and PredictedInstances to delete:

sleap/sleap/gui/commands.py

Lines 2294 to 2298 in d3ad226

	def _do_deletion(context: CommandContext, lf_inst_list: List[int]):
	# Delete the instances
	lfs_to_remove = []
	for lf, inst in lf_inst_list:
	context.labels.remove_instance(lf, inst, in_transaction=True)

during which we call Labels.remove_instance

sleap/sleap/io/dataset.py

Lines 1318 to 1322 in d3ad226

	def remove_instance(
	self, frame: LabeledFrame, instance: Instance, in_transaction: bool = False
	):
	"""Remove instance from frame, updating track occupancy."""
	frame.instances.remove(instance)

which calls the List.remove method (another iteration through only Instances in that particular LabeledFrame until it finds the Instance to remove - at worst iterating through all the Instances in the LabeledFrame.instances list).

At this point, the PredictedInstances have been deleted. While removing PredictedInstances, we were also keeping track of now-empty LabeledFrames to remove which is what we will do now

sleap/sleap/gui/commands.py

Lines 2301 to 2304 in d3ad226

	if len(lf.instances) == 0:
	lfs_to_remove.append(lf)
	context.labels.remove_frames(lfs_to_remove)

.

Incomplete investigation, but will return with an update.

Thanks,
Liezl