What's the usage of `event_to_attr`

[Nic-Ma]

❓ Questions/Help/Support

Hi @vfdev-5 ,

I see there is a dict property event_to_attr in State:
https://github.com/pytorch/ignite/blob/master/ignite/engine/events.py#L374
And users can register new event-to-attr "mapping" in engines.
But I didn't find the usage of this property, may I know why you defined this dict and what the usage case it is for?

In my simple mind, maybe "register events" and "register attributes" can be 2 independent features of State?
I didn't see the relationship between "event names" and "attributes" of State.
I am enabling it in MONAI and Clara workflows, users use custom events and attributes to share data between handlers.

Thanks in advance.

[vfdev-5]

Hi @Nic-Ma ,

it is, for example, used internally to resolve events filtering:

@trainer.on(Events.EPOCH_COMPLETED(every=5))
def do_something():
    # do something

it adds a if-else clause: trainer.state.epoch % 5 == 0 and we have to connect Events.EPOCH_COMPLETED <-> State.epoch. Please, keep in mind that some parts of this design is subject to change in future release as we'd like to have more splitting between Engine and State. We'll ensure BC as much as possible but filtering will be done in a bit different way. State will be more sync with internal Engine's event counters...

As for custom events, we discussed that previously and we updated our docs here:

• https://pytorch.org/ignite/master/concepts.html#custom-events

If you want to use filtering with custom events (e.g. CustomEvents.CUSTOM_STARTED(every=5)),

• engine.state should have corresponding attributes for the events, e.g. engine.state.custom_started
• you need to pass a dict event_to_attr to register_events(), which maps between events and state attributes,...
• you should increase the counter for the event whenever you fire the event, e.g. engine.state.custom_started += 1

[Nic-Ma]

Hi @vfdev-5 ,

Thanks for your example, makes sense.
As Clara users usually use JSON config to develop DL pipeline, we highlight "event-handlers" as the main approach to bring their own python code.
So "register events", "register attrs", "end-to-end context dict data" would 3 most important features to construct a flexible program based on ignite handlers.

Thanks.

[vfdev-5]

As Clara users usually use JSON config to develop DL pipeline, we highlight "event-handlers" as the main approach to bring their own python code.

@Nic-Ma, I had a side thought about DL pipeline development. What do you think about providing an app similar to this one for your Clara users : http://pytorch-ignite-code-generator-dev.eu-west-3.elasticbeanstalk.com/ where the main goal is to "generate your training scripts" ?

[Nic-Ma]

Hi @vfdev-5 ,

Thanks for your sharing, a very cool online tool.
NVIDIA Clara is based on MMAR standard archive and JSON configs, many users are not programmers.
https://docs.nvidia.com/clara/tlt-mi/nvmidl/model.html#
The new version that is based on MONAI & ignite is not released yet.
Maybe we can leverage your application in the future.

Thanks.

[vfdev-5]

Let me close this issue as answered. We can reopen it if needed.