Enhancement of class Metric by adding attachment events to __init__() arguments
#644
Conversation
…_()` arguments: The original `Metric.__init__(self, output_transform=lambda x: x)` is enhanced to `Metric.__init__(output_transform=lambda x: x, started_event=Events.EPOCH_STARTED, iteration_completed_event=Events.ITERATION_COMPLETED, completed_event=Events.EPOCH_COMPLETED) to provide metric updating and computing at any individual event. The default values guarantee backward compatability.
|
Great, I'm happy to see that merged 👍 ... by the way despite some subclasses of |
|
@DrStoop sorry, yes, you are right about other classes. Maybe we would need to get them all in this PR. |
|
no problem, wait with the merge. I will open separate issues for each topic I was working on. After deciding on the features to integrate, I'll tie the winners up and pull-request them at once... |
|
@DrStoop I started to write some tests for this new feature and seems like with current events system it is rather difficult to setup something useful, e.g. # Would like to compute accuracy on the half of the loader
# the following should compute only on the first half
e1 = Events.ITERATION_STARTED(once=1),
e2 = Events.ITERATION_COMPLETED,
e3 = Events.ITERATION_COMPLETED(once=50)
m = Accuracy(..., started_event=e1, iteration_completed_event=e2, completed_event=e3)The problem is with reset event. I added another issue #710 such that if we could pass a composite event we could |
|
Hi @vfdev-5, yeah that's maybe because it was implemented before the event filter... but i'll come up with something possibly in a week or two... unfortunately won't make it earlier |
2 participants
The original
Metric.__init__(self, output_transform=lambda x: x)is enhanced to`Metric.init(output_transform=lambda x: x, started_event=Events.EPOCH_STARTED,
iteration_completed_event=Events.ITERATION_COMPLETED, completed_event=Events.EPOCH_COMPLETED)
to provide metric updating and computing at any individual event. The default values
guarantee backward compatibility.
Modified metrics inheriting from
Metrics will follow up...Fixes #643
Description:
Check list:
pytests were passed with constant number of failed tests, before/after = 14 of 341 / 14 of 341