additional events in the deepgrow interaction engine

[wyli]

Is your feature request related to a problem? Please describe.
This is a feature request for adding extra engine events within the click simulation loops during the deepgrow model training:

MONAI/monai/apps/deepgrow/interaction.py

Lines 61 to 77 in abad841

	for j in range(self.max_interactions):
	inputs, _ = engine.prepare_batch(batchdata)
	inputs = inputs.to(engine.state.device)
	engine.network.eval()
	with torch.no_grad():
	if engine.amp:
	with torch.cuda.amp.autocast():
	predictions = engine.inferer(inputs, engine.network)
	else:
	predictions = engine.inferer(inputs, engine.network)
	batchdata.update({CommonKeys.PRED: predictions})
	batchdata[self.key_probability] = torch.as_tensor(
	([1.0 - ((1.0 / self.max_interactions) * j)] if self.train else [1.0]) * len(inputs)
	)
	batchdata = self.transforms(batchdata)

the main benefit is to have flexible simulation handlers attached to the inner loops

cc @danieltudosiu @diazandr3s @SachidanandAlle

[Nic-Ma]

Please refer to the existing SupervisedTrainer:
https://github.com/Project-MONAI/MONAI/blob/master/monai/engines/trainer.py#L164

Thanks.

[diazandr3s]

Thanks @wyli .
@Nic-Ma in the current SupervisedTrainer implementation can I access the simulated positive and negative points for a single image?

[Nic-Ma]

Hi @diazandr3s ,

I mean you can refer to the current SupervisedTrainer to add additional events.

Thanks.

[diazandr3s]

Superb! @danieltudosiu and I implemented a couple of engine events called INNER_ITERATION_STARTED and INNER_ITERATION_COMPLETED for a custom Interaction implementation.

        engine.fire_event(DeepEditEvents.INNER_ITERATION_STARTED)

        engine.network.eval()
        with torch.no_grad():
            if engine.amp:
                with torch.cuda.amp.autocast():
                    predictions = engine.inferer(inputs, engine.network)
            else:
                predictions = engine.inferer(inputs, engine.network)

        engine.state.batch_data_inner = batchdata
        engine.state.preds_inner = predictions
        engine.state.step_inner = j

        engine.fire_event(DeepEditEvents.INNER_ITERATION_COMPLETED)

It works quite well for my purpose (DeepEdit implementation). Not sure if that's reasonable of other MONAI users

[SachidanandAlle]

@diazandr3s can we check if this is good enough solve your requirement for deepedit usecase..

In that case.. we can add a tutorial in monai on how to extend deepgrow app and train a deepedit model..
May be you want to add one specific transform to drop the signals based on percentage into monai