Add more Events for MONAI workflow engines

[Nic-Ma]

Description

This PR registered more Events to ignite engine for MONAI workflow trainer and evaluator.
Users can write more powerful handlers based on it.
If user has very special iteration computation, they can still use iteration_update() callback.

Status

Ready

Types of changes

• Non-breaking change (fix or new feature that would not break existing functionality).
• Breaking change (fix or new feature that would cause existing functionality to change).
• New tests added to cover the changes.
• Integration tests passed locally by running ./runtests.sh --codeformat --coverage.
• Quick tests passed locally by running ./runtests.sh --quick.
• In-line docstrings updated.
• Documentation updated, tested make html command in the docs/ folder.

[Nic-Ma]

/black

[Nic-Ma]

/black

[wyli]

thanks this needs more discussions and tests

[Nic-Ma]

/black

[Nic-Ma]

/black

[Nic-Ma]

Hi @wyli and @yanchengnv ,

Thanks for your reminder, I added test cases in the integration tests.
I think it's ready for review now.

Thanks.

[Nic-Ma]

And BTW, this PR refers to ignite example: https://github.com/pytorch/ignite/blob/master/ignite/engine/events.py#L146
Thanks for @vfdev-5 's help.

[vfdev-5]

@Nic-Ma thanks for mentioning ! Just FYI, currently we have undocumented requirements for custom events if using filters with them: pytorch/ignite#1555
Hopefully, docs issue will be fixed in the next release v0.4.3 and in v0.5.0 user wont need to increment state attribute counter:

def train_step(engine, batch):
    engine.fire_event(TestEvents.TEST_EVENT)
    engine.state.test_event += 1

[Nic-Ma]

Hi @vfdev-5 ,

Thanks for your tips, I didn't add event_to_attr dict to the register_events() call because I think we don't have count attributes to keep in State for these 3 Events. So we don't need engine.state.test_event += 1, right?

Thanks.

[vfdev-5]

Yes, correct, in this case you do not need engine.state.test_event += 1.

[wyli]

thanks please see some comments inline, we'll discuss this next monday

[Nic-Ma]

/black

[Nic-Ma]

Hi @wyli ,

Thanks for your review.
I updated the PR according to your comments.
Could you please help review it again?

Thanks.

[wyli]

thanks, this PR adds a few events to each iteration. just wanted to note that state.output is used to cache information across those events. not sure if this is the best way of using state.output but seems like a viable workaround according to @Nic-Ma

[Nic-Ma]

Thanks @wyli ,

And @vfdev-5 , as we discussed in the above thread, I think saving data to engine.state.output is acceptable solution so far, referring to ignite contrib: https://github.com/pytorch/ignite/blob/master/ignite/contrib/engines/tbptt.py#L111.
But please let me know if this design is against your ignite design or any future plan, then we can update it.

Thanks.